lavu: add Vulkan hwcontext code
authorLynne <dev@lynne.ee>
Wed, 28 Aug 2019 20:58:10 +0000 (21:58 +0100)
committerLynne <dev@lynne.ee>
Tue, 4 Feb 2020 23:19:48 +0000 (23:19 +0000)
This commit adds the necessary code to initialize and use a Vulkan device
within the hwcontext libavutil framework.
Currently direct mapping to VAAPI and DRM frames is functional, and
transfers to CUDA and native frames are supported.

Lets hope the future Vulkan video decode extension fits well within this
framework.

configure
doc/APIchanges
libavutil/Makefile
libavutil/hwcontext.c
libavutil/hwcontext.h
libavutil/hwcontext_cuda.c
libavutil/hwcontext_internal.h
libavutil/hwcontext_vulkan.c [new file with mode: 0644]
libavutil/hwcontext_vulkan.h [new file with mode: 0644]
libavutil/pixdesc.c
libavutil/pixfmt.h

index fbb1a86..ad3894f 100755 (executable)
--- a/configure
+++ b/configure
@@ -316,6 +316,7 @@ External library support:
   --disable-securetransport disable Secure Transport, needed for TLS support
                            on OSX if openssl and gnutls are not used [autodetect]
   --enable-vapoursynth     enable VapourSynth demuxer [no]
+  --enable-vulkan          enable Vulkan code [no]
   --disable-xlib           disable xlib [autodetect]
   --disable-zlib           disable zlib [autodetect]
 
@@ -1854,6 +1855,7 @@ HWACCEL_LIBRARY_LIST="
     mmal
     omx
     opencl
+    vulkan
 "
 
 DOCUMENT_LIST="
@@ -3641,7 +3643,7 @@ avformat_deps="avcodec avutil"
 avformat_suggest="libm network zlib"
 avresample_deps="avutil"
 avresample_suggest="libm"
-avutil_suggest="clock_gettime ffnvcodec libm libdrm libmfx opencl user32 vaapi videotoolbox corefoundation corevideo coremedia bcrypt"
+avutil_suggest="clock_gettime ffnvcodec libm libdrm libmfx opencl user32 vaapi vulkan videotoolbox corefoundation corevideo coremedia bcrypt"
 postproc_deps="avutil gpl"
 postproc_suggest="libm"
 swresample_deps="avutil"
@@ -6629,6 +6631,9 @@ enabled vdpau &&
 
 enabled crystalhd && check_lib crystalhd "stdint.h libcrystalhd/libcrystalhd_if.h" DtsCrystalHDVersion -lcrystalhd
 
+enabled vulkan &&
+    require_pkg_config vulkan "vulkan >= 1.1.97" "vulkan/vulkan.h" vkCreateInstance
+
 if enabled x86; then
     case $target_os in
         mingw32*|mingw64*|win32|win64|linux|cygwin*)
index 2494a39..31ddc0a 100644 (file)
@@ -15,6 +15,10 @@ libavutil:     2017-10-21
 
 API changes, most recent first:
 
+2020-ww-xx - xxxxxxxxxx - lavu yy.yy.yyy - hwcontext.h
+  Add AV_PIX_FMT_VULKAN
+  Add AV_HWDEVICE_TYPE_VULKAN and implementation.
+
 2020-01-30 - xxxxxxxxxx - lavf 58.37.100 - avio.h
   Add avio_protocol_get_class().
 
index 57e6e3d..b189f9a 100644 (file)
@@ -43,6 +43,7 @@ HEADERS = adler32.h                                                     \
           hwcontext_vaapi.h                                             \
           hwcontext_videotoolbox.h                                      \
           hwcontext_vdpau.h                                             \
+          hwcontext_vulkan.h                                            \
           imgutils.h                                                    \
           intfloat.h                                                    \
           intreadwrite.h                                                \
@@ -175,6 +176,7 @@ OBJS-$(CONFIG_QSV)                      += hwcontext_qsv.o
 OBJS-$(CONFIG_VAAPI)                    += hwcontext_vaapi.o
 OBJS-$(CONFIG_VIDEOTOOLBOX)             += hwcontext_videotoolbox.o
 OBJS-$(CONFIG_VDPAU)                    += hwcontext_vdpau.o
+OBJS-$(CONFIG_VULKAN)                   += hwcontext_vulkan.o
 
 OBJS += $(COMPAT_OBJS:%=../compat/%)
 
@@ -191,6 +193,7 @@ SKIPHEADERS-$(CONFIG_OPENCL)           += hwcontext_opencl.h
 SKIPHEADERS-$(CONFIG_VAAPI)            += hwcontext_vaapi.h
 SKIPHEADERS-$(CONFIG_VIDEOTOOLBOX)     += hwcontext_videotoolbox.h
 SKIPHEADERS-$(CONFIG_VDPAU)            += hwcontext_vdpau.h
+SKIPHEADERS-$(CONFIG_VULKAN)           += hwcontext_vulkan.h
 
 TESTPROGS = adler32                                                     \
             aes                                                         \
index 3189391..d09a15a 100644 (file)
@@ -59,6 +59,9 @@ static const HWContextType * const hw_table[] = {
 #if CONFIG_MEDIACODEC
     &ff_hwcontext_type_mediacodec,
 #endif
+#if CONFIG_VULKAN
+    &ff_hwcontext_type_vulkan,
+#endif
     NULL,
 };
 
@@ -73,6 +76,7 @@ static const char *const hw_type_names[] = {
     [AV_HWDEVICE_TYPE_VDPAU]  = "vdpau",
     [AV_HWDEVICE_TYPE_VIDEOTOOLBOX] = "videotoolbox",
     [AV_HWDEVICE_TYPE_MEDIACODEC] = "mediacodec",
+    [AV_HWDEVICE_TYPE_VULKAN] = "vulkan",
 };
 
 enum AVHWDeviceType av_hwdevice_find_type_by_name(const char *name)
index f5a4b62..f874af9 100644 (file)
@@ -36,6 +36,7 @@ enum AVHWDeviceType {
     AV_HWDEVICE_TYPE_DRM,
     AV_HWDEVICE_TYPE_OPENCL,
     AV_HWDEVICE_TYPE_MEDIACODEC,
+    AV_HWDEVICE_TYPE_VULKAN,
 };
 
 typedef struct AVHWDeviceInternal AVHWDeviceInternal;
index 30611b1..ba8ca58 100644 (file)
@@ -21,6 +21,9 @@
 #include "hwcontext.h"
 #include "hwcontext_internal.h"
 #include "hwcontext_cuda_internal.h"
+#if CONFIG_VULKAN
+#include "hwcontext_vulkan.h"
+#endif
 #include "cuda_check.h"
 #include "mem.h"
 #include "pixdesc.h"
@@ -42,6 +45,9 @@ static const enum AVPixelFormat supported_formats[] = {
     AV_PIX_FMT_YUV444P16,
     AV_PIX_FMT_0RGB32,
     AV_PIX_FMT_0BGR32,
+#if CONFIG_VULKAN
+    AV_PIX_FMT_VULKAN,
+#endif
 };
 
 #define CHECK_CU(x) FF_CUDA_CHECK_DL(device_ctx, cu, x)
@@ -205,6 +211,10 @@ static int cuda_transfer_data_from(AVHWFramesContext *ctx, AVFrame *dst,
     CUcontext dummy;
     int i, ret;
 
+    /* We don't support transfers to HW devices. */
+    if (dst->hw_frames_ctx)
+        return AVERROR(ENOSYS);
+
     ret = CHECK_CU(cu->cuCtxPushCurrent(hwctx->cuda_ctx));
     if (ret < 0)
         return ret;
@@ -247,6 +257,10 @@ static int cuda_transfer_data_to(AVHWFramesContext *ctx, AVFrame *dst,
     CUcontext dummy;
     int i, ret;
 
+    /* We don't support transfers from HW devices. */
+    if (src->hw_frames_ctx)
+        return AVERROR(ENOSYS);
+
     ret = CHECK_CU(cu->cuCtxPushCurrent(hwctx->cuda_ctx));
     if (ret < 0)
         return ret;
@@ -389,6 +403,123 @@ error:
     return AVERROR_UNKNOWN;
 }
 
+static int cuda_device_derive(AVHWDeviceContext *device_ctx,
+                              AVHWDeviceContext *src_ctx,
+                              int flags) {
+    AVCUDADeviceContext *hwctx = device_ctx->hwctx;
+    CudaFunctions *cu;
+    const char *src_uuid = NULL;
+    CUcontext dummy;
+    int ret, i, device_count, dev_active = 0;
+    unsigned int dev_flags = 0;
+
+    const unsigned int desired_flags = CU_CTX_SCHED_BLOCKING_SYNC;
+
+#if CONFIG_VULKAN
+    VkPhysicalDeviceIDProperties vk_idp = {
+        .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES,
+    };
+#endif
+
+    switch (src_ctx->type) {
+#if CONFIG_VULKAN
+    case AV_HWDEVICE_TYPE_VULKAN: {
+        AVVulkanDeviceContext *vkctx = src_ctx->hwctx;
+        VkPhysicalDeviceProperties2 vk_dev_props = {
+            .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2,
+            .pNext = &vk_idp,
+        };
+        vkGetPhysicalDeviceProperties2(vkctx->phys_dev, &vk_dev_props);
+        src_uuid = vk_idp.deviceUUID;
+        break;
+    }
+#endif
+    default:
+        return AVERROR(ENOSYS);
+    }
+
+    if (!src_uuid) {
+        av_log(device_ctx, AV_LOG_ERROR,
+               "Failed to get UUID of source device.\n");
+        goto error;
+    }
+
+    if (cuda_device_init(device_ctx) < 0)
+        goto error;
+
+    cu = hwctx->internal->cuda_dl;
+
+    ret = CHECK_CU(cu->cuInit(0));
+    if (ret < 0)
+        goto error;
+
+    ret = CHECK_CU(cu->cuDeviceGetCount(&device_count));
+    if (ret < 0)
+        goto error;
+
+    hwctx->internal->cuda_device = -1;
+    for (i = 0; i < device_count; i++) {
+        CUdevice dev;
+        CUuuid uuid;
+
+        ret = CHECK_CU(cu->cuDeviceGet(&dev, i));
+        if (ret < 0)
+            goto error;
+
+        ret = CHECK_CU(cu->cuDeviceGetUuid(&uuid, dev));
+        if (ret < 0)
+            goto error;
+
+        if (memcmp(src_uuid, uuid.bytes, sizeof (uuid.bytes)) == 0) {
+            hwctx->internal->cuda_device = dev;
+            break;
+        }
+    }
+
+    if (hwctx->internal->cuda_device == -1) {
+        av_log(device_ctx, AV_LOG_ERROR, "Could not derive CUDA device.\n");
+        goto error;
+    }
+
+    hwctx->internal->flags = flags;
+
+    if (flags & AV_CUDA_USE_PRIMARY_CONTEXT) {
+        ret = CHECK_CU(cu->cuDevicePrimaryCtxGetState(hwctx->internal->cuda_device, &dev_flags, &dev_active));
+        if (ret < 0)
+            goto error;
+
+        if (dev_active && dev_flags != desired_flags) {
+            av_log(device_ctx, AV_LOG_ERROR, "Primary context already active with incompatible flags.\n");
+            goto error;
+        } else if (dev_flags != desired_flags) {
+            ret = CHECK_CU(cu->cuDevicePrimaryCtxSetFlags(hwctx->internal->cuda_device, desired_flags));
+            if (ret < 0)
+                goto error;
+        }
+
+        ret = CHECK_CU(cu->cuDevicePrimaryCtxRetain(&hwctx->cuda_ctx, hwctx->internal->cuda_device));
+        if (ret < 0)
+            goto error;
+    } else {
+        ret = CHECK_CU(cu->cuCtxCreate(&hwctx->cuda_ctx, desired_flags, hwctx->internal->cuda_device));
+        if (ret < 0)
+            goto error;
+
+        CHECK_CU(cu->cuCtxPopCurrent(&dummy));
+    }
+
+    hwctx->internal->is_allocated = 1;
+
+    // Setting stream to NULL will make functions automatically use the default CUstream
+    hwctx->stream = NULL;
+
+    return 0;
+
+error:
+    cuda_device_uninit(device_ctx);
+    return AVERROR_UNKNOWN;
+}
+
 const HWContextType ff_hwcontext_type_cuda = {
     .type                 = AV_HWDEVICE_TYPE_CUDA,
     .name                 = "CUDA",
@@ -397,6 +528,7 @@ const HWContextType ff_hwcontext_type_cuda = {
     .frames_priv_size     = sizeof(CUDAFramesContext),
 
     .device_create        = cuda_device_create,
+    .device_derive        = cuda_device_derive,
     .device_init          = cuda_device_init,
     .device_uninit        = cuda_device_uninit,
     .frames_get_constraints = cuda_frames_get_constraints,
index 77dc47d..dba0f39 100644 (file)
@@ -172,5 +172,6 @@ extern const HWContextType ff_hwcontext_type_vaapi;
 extern const HWContextType ff_hwcontext_type_vdpau;
 extern const HWContextType ff_hwcontext_type_videotoolbox;
 extern const HWContextType ff_hwcontext_type_mediacodec;
+extern const HWContextType ff_hwcontext_type_vulkan;
 
 #endif /* AVUTIL_HWCONTEXT_INTERNAL_H */
diff --git a/libavutil/hwcontext_vulkan.c b/libavutil/hwcontext_vulkan.c
new file mode 100644 (file)
index 0000000..0fe88e6
--- /dev/null
@@ -0,0 +1,2825 @@
+/*
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "config.h"
+#include "pixdesc.h"
+#include "avstring.h"
+#include "imgutils.h"
+#include "hwcontext.h"
+#include "hwcontext_internal.h"
+#include "hwcontext_vulkan.h"
+
+#if CONFIG_LIBDRM
+#include <unistd.h>
+#include <xf86drm.h>
+#include <drm_fourcc.h>
+#include "hwcontext_drm.h"
+#if CONFIG_VAAPI
+#include <va/va_drmcommon.h>
+#include "hwcontext_vaapi.h"
+#endif
+#endif
+
+#if CONFIG_CUDA
+#include "hwcontext_cuda_internal.h"
+#include "cuda_check.h"
+#define CHECK_CU(x) FF_CUDA_CHECK_DL(cuda_cu, cu, x)
+#endif
+
+typedef struct VulkanExecCtx {
+    VkCommandPool pool;
+    VkCommandBuffer buf;
+    VkQueue queue;
+    VkFence fence;
+} VulkanExecCtx;
+
+typedef struct VulkanDevicePriv {
+    /* Properties */
+    VkPhysicalDeviceProperties props;
+    VkPhysicalDeviceMemoryProperties mprops;
+
+    /* Debug callback */
+    VkDebugUtilsMessengerEXT debug_ctx;
+
+    /* Image uploading */
+    VulkanExecCtx cmd;
+
+    /* Extensions */
+    uint64_t extensions;
+
+    /* Settings */
+    int use_linear_images;
+
+    /* Nvidia */
+    int dev_is_nvidia;
+} VulkanDevicePriv;
+
+typedef struct VulkanFramesPriv {
+    VulkanExecCtx cmd;
+} VulkanFramesPriv;
+
+typedef struct AVVkFrameInternal {
+#if CONFIG_CUDA
+    /* Importing external memory into cuda is really expensive so we keep the
+     * memory imported all the time */
+    AVBufferRef *cuda_fc_ref; /* Need to keep it around for uninit */
+    CUexternalMemory ext_mem[AV_NUM_DATA_POINTERS];
+    CUmipmappedArray cu_mma[AV_NUM_DATA_POINTERS];
+    CUarray cu_array[AV_NUM_DATA_POINTERS];
+    CUexternalSemaphore cu_sem[AV_NUM_DATA_POINTERS];
+#endif
+} AVVkFrameInternal;
+
+#define VK_LOAD_PFN(inst, name) PFN_##name pfn_##name = (PFN_##name)           \
+                                              vkGetInstanceProcAddr(inst, #name)
+
+#define DEFAULT_USAGE_FLAGS (VK_IMAGE_USAGE_SAMPLED_BIT      |                 \
+                             VK_IMAGE_USAGE_STORAGE_BIT      |                 \
+                             VK_IMAGE_USAGE_TRANSFER_SRC_BIT |                 \
+                             VK_IMAGE_USAGE_TRANSFER_DST_BIT)
+
+#define ADD_VAL_TO_LIST(list, count, val)                                      \
+    do {                                                                       \
+        list = av_realloc_array(list, sizeof(*list), ++count);                 \
+        if (!list) {                                                           \
+            err = AVERROR(ENOMEM);                                             \
+            goto end;                                                          \
+        }                                                                      \
+        list[count - 1] = val;                                                 \
+    } while(0)
+
+static const struct {
+    enum AVPixelFormat pixfmt;
+    const VkFormat vkfmts[3];
+} vk_pixfmt_map[] = {
+    { AV_PIX_FMT_GRAY8,   { VK_FORMAT_R8_UNORM } },
+    { AV_PIX_FMT_GRAY16,  { VK_FORMAT_R16_UNORM } },
+    { AV_PIX_FMT_GRAYF32, { VK_FORMAT_R32_SFLOAT } },
+
+    { AV_PIX_FMT_NV12, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8G8_UNORM } },
+    { AV_PIX_FMT_P010, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } },
+    { AV_PIX_FMT_P016, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16G16_UNORM } },
+
+    { AV_PIX_FMT_YUV420P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
+    { AV_PIX_FMT_YUV422P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
+    { AV_PIX_FMT_YUV444P, { VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM, VK_FORMAT_R8_UNORM } },
+
+    { AV_PIX_FMT_YUV420P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
+    { AV_PIX_FMT_YUV422P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
+    { AV_PIX_FMT_YUV444P16, { VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM, VK_FORMAT_R16_UNORM } },
+
+    { AV_PIX_FMT_ABGR,   { VK_FORMAT_A8B8G8R8_UNORM_PACK32 } },
+    { AV_PIX_FMT_BGRA,   { VK_FORMAT_B8G8R8A8_UNORM } },
+    { AV_PIX_FMT_RGBA,   { VK_FORMAT_R8G8B8A8_UNORM } },
+    { AV_PIX_FMT_RGB24,  { VK_FORMAT_R8G8B8_UNORM } },
+    { AV_PIX_FMT_BGR24,  { VK_FORMAT_B8G8R8_UNORM } },
+    { AV_PIX_FMT_RGB48,  { VK_FORMAT_R16G16B16_UNORM } },
+    { AV_PIX_FMT_RGBA64, { VK_FORMAT_R16G16B16A16_UNORM } },
+    { AV_PIX_FMT_RGB565, { VK_FORMAT_R5G6B5_UNORM_PACK16 } },
+    { AV_PIX_FMT_BGR565, { VK_FORMAT_B5G6R5_UNORM_PACK16 } },
+    { AV_PIX_FMT_BGR0,   { VK_FORMAT_B8G8R8A8_UNORM } },
+    { AV_PIX_FMT_0BGR,   { VK_FORMAT_A8B8G8R8_UNORM_PACK32 } },
+    { AV_PIX_FMT_RGB0,   { VK_FORMAT_R8G8B8A8_UNORM } },
+
+    { AV_PIX_FMT_GBRPF32, { VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_SFLOAT } },
+};
+
+const VkFormat *av_vkfmt_from_pixfmt(enum AVPixelFormat p)
+{
+    for (enum AVPixelFormat i = 0; i < FF_ARRAY_ELEMS(vk_pixfmt_map); i++)
+        if (vk_pixfmt_map[i].pixfmt == p)
+            return vk_pixfmt_map[i].vkfmts;
+    return NULL;
+}
+
+static int pixfmt_is_supported(AVVulkanDeviceContext *hwctx, enum AVPixelFormat p,
+                               int linear)
+{
+    const VkFormat *fmt = av_vkfmt_from_pixfmt(p);
+    int planes = av_pix_fmt_count_planes(p);
+
+    if (!fmt)
+        return 0;
+
+    for (int i = 0; i < planes; i++) {
+        VkFormatFeatureFlags flags;
+        VkFormatProperties2 prop = {
+            .sType = VK_STRUCTURE_TYPE_FORMAT_PROPERTIES_2,
+        };
+        vkGetPhysicalDeviceFormatProperties2(hwctx->phys_dev, fmt[i], &prop);
+        flags = linear ? prop.formatProperties.linearTilingFeatures :
+                         prop.formatProperties.optimalTilingFeatures;
+        if (!(flags & DEFAULT_USAGE_FLAGS))
+            return 0;
+    }
+
+    return 1;
+}
+
+enum VulkanExtensions {
+    EXT_EXTERNAL_DMABUF_MEMORY = 1ULL <<  0, /* VK_EXT_external_memory_dma_buf */
+    EXT_DRM_MODIFIER_FLAGS     = 1ULL <<  1, /* VK_EXT_image_drm_format_modifier */
+    EXT_EXTERNAL_FD_MEMORY     = 1ULL <<  2, /* VK_KHR_external_memory_fd */
+    EXT_EXTERNAL_FD_SEM        = 1ULL <<  3, /* VK_KHR_external_semaphore_fd */
+
+    EXT_OPTIONAL               = 1ULL << 62,
+    EXT_REQUIRED               = 1ULL << 63,
+};
+
+typedef struct VulkanOptExtension {
+    const char *name;
+    uint64_t flag;
+} VulkanOptExtension;
+
+static const VulkanOptExtension optional_instance_exts[] = {
+    /* For future use */
+};
+
+static const VulkanOptExtension optional_device_exts[] = {
+    { VK_KHR_EXTERNAL_MEMORY_FD_EXTENSION_NAME,               EXT_EXTERNAL_FD_MEMORY,     },
+    { VK_EXT_EXTERNAL_MEMORY_DMA_BUF_EXTENSION_NAME,          EXT_EXTERNAL_DMABUF_MEMORY, },
+    { VK_EXT_IMAGE_DRM_FORMAT_MODIFIER_EXTENSION_NAME,        EXT_DRM_MODIFIER_FLAGS,     },
+    { VK_KHR_EXTERNAL_SEMAPHORE_FD_EXTENSION_NAME,            EXT_EXTERNAL_FD_SEM,        },
+};
+
+/* Converts return values to strings */
+static const char *vk_ret2str(VkResult res)
+{
+#define CASE(VAL) case VAL: return #VAL
+    switch (res) {
+    CASE(VK_SUCCESS);
+    CASE(VK_NOT_READY);
+    CASE(VK_TIMEOUT);
+    CASE(VK_EVENT_SET);
+    CASE(VK_EVENT_RESET);
+    CASE(VK_INCOMPLETE);
+    CASE(VK_ERROR_OUT_OF_HOST_MEMORY);
+    CASE(VK_ERROR_OUT_OF_DEVICE_MEMORY);
+    CASE(VK_ERROR_INITIALIZATION_FAILED);
+    CASE(VK_ERROR_DEVICE_LOST);
+    CASE(VK_ERROR_MEMORY_MAP_FAILED);
+    CASE(VK_ERROR_LAYER_NOT_PRESENT);
+    CASE(VK_ERROR_EXTENSION_NOT_PRESENT);
+    CASE(VK_ERROR_FEATURE_NOT_PRESENT);
+    CASE(VK_ERROR_INCOMPATIBLE_DRIVER);
+    CASE(VK_ERROR_TOO_MANY_OBJECTS);
+    CASE(VK_ERROR_FORMAT_NOT_SUPPORTED);
+    CASE(VK_ERROR_FRAGMENTED_POOL);
+    CASE(VK_ERROR_SURFACE_LOST_KHR);
+    CASE(VK_ERROR_NATIVE_WINDOW_IN_USE_KHR);
+    CASE(VK_SUBOPTIMAL_KHR);
+    CASE(VK_ERROR_OUT_OF_DATE_KHR);
+    CASE(VK_ERROR_INCOMPATIBLE_DISPLAY_KHR);
+    CASE(VK_ERROR_VALIDATION_FAILED_EXT);
+    CASE(VK_ERROR_INVALID_SHADER_NV);
+    CASE(VK_ERROR_OUT_OF_POOL_MEMORY);
+    CASE(VK_ERROR_INVALID_EXTERNAL_HANDLE);
+    CASE(VK_ERROR_NOT_PERMITTED_EXT);
+    CASE(VK_ERROR_INVALID_DRM_FORMAT_MODIFIER_PLANE_LAYOUT_EXT);
+    CASE(VK_ERROR_INVALID_DEVICE_ADDRESS_EXT);
+    CASE(VK_ERROR_FULL_SCREEN_EXCLUSIVE_MODE_LOST_EXT);
+    default: return "Unknown error";
+    }
+#undef CASE
+}
+
+static VkBool32 vk_dbg_callback(VkDebugUtilsMessageSeverityFlagBitsEXT severity,
+                                VkDebugUtilsMessageTypeFlagsEXT messageType,
+                                const VkDebugUtilsMessengerCallbackDataEXT *data,
+                                void *priv)
+{
+    int l;
+    AVHWDeviceContext *ctx = priv;
+
+    switch (severity) {
+    case VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT: l = AV_LOG_VERBOSE; break;
+    case VK_DEBUG_UTILS_MESSAGE_SEVERITY_INFO_BIT_EXT:    l = AV_LOG_INFO;    break;
+    case VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT: l = AV_LOG_WARNING; break;
+    case VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT:   l = AV_LOG_ERROR;   break;
+    default:                                              l = AV_LOG_DEBUG;   break;
+    }
+
+    av_log(ctx, l, "%s\n", data->pMessage);
+    for (int i = 0; i < data->cmdBufLabelCount; i++)
+        av_log(ctx, l, "\t%i: %s\n", i, data->pCmdBufLabels[i].pLabelName);
+
+    return 0;
+}
+
+static int check_extensions(AVHWDeviceContext *ctx, int dev,
+                            const char * const **dst, uint32_t *num, int debug)
+{
+    const char *tstr;
+    const char **extension_names = NULL;
+    VulkanDevicePriv *p = ctx->internal->priv;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+    int err = 0, found, extensions_found = 0;
+
+    const char *mod;
+    int optional_exts_num;
+    uint32_t sup_ext_count;
+    VkExtensionProperties *sup_ext;
+    const VulkanOptExtension *optional_exts;
+
+    if (!dev) {
+        mod = "instance";
+        optional_exts = optional_instance_exts;
+        optional_exts_num = FF_ARRAY_ELEMS(optional_instance_exts);
+        vkEnumerateInstanceExtensionProperties(NULL, &sup_ext_count, NULL);
+        sup_ext = av_malloc_array(sup_ext_count, sizeof(VkExtensionProperties));
+        if (!sup_ext)
+            return AVERROR(ENOMEM);
+        vkEnumerateInstanceExtensionProperties(NULL, &sup_ext_count, sup_ext);
+    } else {
+        mod = "device";
+        optional_exts = optional_device_exts;
+        optional_exts_num = FF_ARRAY_ELEMS(optional_device_exts);
+        vkEnumerateDeviceExtensionProperties(hwctx->phys_dev, NULL,
+                                             &sup_ext_count, NULL);
+        sup_ext = av_malloc_array(sup_ext_count, sizeof(VkExtensionProperties));
+        if (!sup_ext)
+            return AVERROR(ENOMEM);
+        vkEnumerateDeviceExtensionProperties(hwctx->phys_dev, NULL,
+                                             &sup_ext_count, sup_ext);
+    }
+
+    for (int i = 0; i < optional_exts_num; i++) {
+        int req = optional_exts[i].flag & EXT_REQUIRED;
+        tstr = optional_exts[i].name;
+
+        found = 0;
+        for (int j = 0; j < sup_ext_count; j++) {
+            if (!strcmp(tstr, sup_ext[j].extensionName)) {
+                found = 1;
+                break;
+            }
+        }
+        if (!found) {
+            int lvl = req ? AV_LOG_ERROR : AV_LOG_VERBOSE;
+            av_log(ctx, lvl, "Extension \"%s\" not found!\n", tstr);
+            if (req) {
+                err = AVERROR(EINVAL);
+                goto end;
+            }
+            continue;
+        }
+        if (!req)
+            p->extensions |= optional_exts[i].flag;
+
+        av_log(ctx, AV_LOG_VERBOSE, "Using %s extension \"%s\"\n", mod, tstr);
+
+        ADD_VAL_TO_LIST(extension_names, extensions_found, tstr);
+    }
+
+    if (debug && !dev) {
+        tstr = VK_EXT_DEBUG_UTILS_EXTENSION_NAME;
+        found = 0;
+        for (int j = 0; j < sup_ext_count; j++) {
+            if (!strcmp(tstr, sup_ext[j].extensionName)) {
+                found = 1;
+                break;
+            }
+        }
+        if (found) {
+            ADD_VAL_TO_LIST(extension_names, extensions_found, tstr);
+        } else {
+            av_log(ctx, AV_LOG_ERROR, "Debug extension \"%s\" not found!\n",
+                   tstr);
+            err = AVERROR(EINVAL);
+            goto end;
+        }
+    }
+
+    *dst = extension_names;
+    *num = extensions_found;
+
+end:
+    av_free(sup_ext);
+    return err;
+}
+
+/* Creates a VkInstance */
+static int create_instance(AVHWDeviceContext *ctx, AVDictionary *opts)
+{
+    int err = 0;
+    VkResult ret;
+    VulkanDevicePriv *p = ctx->internal->priv;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+    AVDictionaryEntry *debug_opt = av_dict_get(opts, "debug", NULL, 0);
+    const int debug_mode = debug_opt && strtol(debug_opt->value, NULL, 10);
+    VkApplicationInfo application_info = {
+        .sType              = VK_STRUCTURE_TYPE_APPLICATION_INFO,
+        .pEngineName        = "libavutil",
+        .apiVersion         = VK_API_VERSION_1_1,
+        .engineVersion      = VK_MAKE_VERSION(LIBAVUTIL_VERSION_MAJOR,
+                                              LIBAVUTIL_VERSION_MINOR,
+                                              LIBAVUTIL_VERSION_MICRO),
+    };
+    VkInstanceCreateInfo inst_props = {
+        .sType            = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO,
+        .pApplicationInfo = &application_info,
+    };
+
+    /* Check for present/missing extensions */
+    err = check_extensions(ctx, 0, &inst_props.ppEnabledExtensionNames,
+                           &inst_props.enabledExtensionCount, debug_mode);
+    if (err < 0)
+        return err;
+
+    if (debug_mode) {
+        static const char *layers[] = { "VK_LAYER_LUNARG_standard_validation" };
+        inst_props.ppEnabledLayerNames = layers;
+        inst_props.enabledLayerCount = FF_ARRAY_ELEMS(layers);
+    }
+
+    /* Try to create the instance */
+    ret = vkCreateInstance(&inst_props, hwctx->alloc, &hwctx->inst);
+
+    /* Free used memory */
+    av_free((void *)inst_props.ppEnabledExtensionNames);
+
+    /* Check for errors */
+    if (ret != VK_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Instance creation failure: %s\n",
+               vk_ret2str(ret));
+        return AVERROR_EXTERNAL;
+    }
+
+    if (debug_mode) {
+        VkDebugUtilsMessengerCreateInfoEXT dbg = {
+            .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_MESSENGER_CREATE_INFO_EXT,
+            .messageSeverity = VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT |
+                               VK_DEBUG_UTILS_MESSAGE_SEVERITY_INFO_BIT_EXT    |
+                               VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT |
+                               VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT,
+            .messageType = VK_DEBUG_UTILS_MESSAGE_TYPE_GENERAL_BIT_EXT    |
+                           VK_DEBUG_UTILS_MESSAGE_TYPE_VALIDATION_BIT_EXT |
+                           VK_DEBUG_UTILS_MESSAGE_TYPE_PERFORMANCE_BIT_EXT,
+            .pfnUserCallback = vk_dbg_callback,
+            .pUserData = ctx,
+        };
+        VK_LOAD_PFN(hwctx->inst, vkCreateDebugUtilsMessengerEXT);
+
+        pfn_vkCreateDebugUtilsMessengerEXT(hwctx->inst, &dbg,
+                                           hwctx->alloc, &p->debug_ctx);
+    }
+
+    return 0;
+}
+
+typedef struct VulkanDeviceSelection {
+    uint8_t uuid[VK_UUID_SIZE]; /* Will use this first unless !has_uuid */
+    int has_uuid;
+    const char *name; /* Will use this second unless NULL */
+    uint32_t pci_device; /* Will use this third unless 0x0 */
+    uint32_t vendor_id; /* Last resort to find something deterministic */
+    int index; /* Finally fall back to index */
+} VulkanDeviceSelection;
+
+static const char *vk_dev_type(enum VkPhysicalDeviceType type)
+{
+    switch (type) {
+    case VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU: return "integrated";
+    case VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU:   return "discrete";
+    case VK_PHYSICAL_DEVICE_TYPE_VIRTUAL_GPU:    return "virtual";
+    case VK_PHYSICAL_DEVICE_TYPE_CPU:            return "software";
+    default:                                     return "unknown";
+    }
+}
+
+/* Finds a device */
+static int find_device(AVHWDeviceContext *ctx, VulkanDeviceSelection *select)
+{
+    int err = 0, choice = -1;
+    uint32_t num;
+    VkResult ret;
+    VkPhysicalDevice *devices = NULL;
+    VkPhysicalDeviceIDProperties *idp = NULL;
+    VkPhysicalDeviceProperties2 *prop = NULL;
+    VulkanDevicePriv *p = ctx->internal->priv;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+
+    ret = vkEnumeratePhysicalDevices(hwctx->inst, &num, NULL);
+    if (ret != VK_SUCCESS || !num) {
+        av_log(ctx, AV_LOG_ERROR, "No devices found: %s!\n", vk_ret2str(ret));
+        return AVERROR(ENODEV);
+    }
+
+    devices = av_malloc_array(num, sizeof(VkPhysicalDevice));
+    if (!devices)
+        return AVERROR(ENOMEM);
+
+    ret = vkEnumeratePhysicalDevices(hwctx->inst, &num, devices);
+    if (ret != VK_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Failed enumerating devices: %s\n",
+               vk_ret2str(ret));
+        err = AVERROR(ENODEV);
+        goto end;
+    }
+
+    prop = av_mallocz_array(num, sizeof(*prop));
+    if (!prop) {
+        err = AVERROR(ENOMEM);
+        goto end;
+    }
+
+    idp = av_mallocz_array(num, sizeof(*idp));
+    if (!idp) {
+        err = AVERROR(ENOMEM);
+        goto end;
+    }
+
+    av_log(ctx, AV_LOG_VERBOSE, "GPU listing:\n");
+    for (int i = 0; i < num; i++) {
+        idp[i].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES;
+        prop[i].sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2;
+        prop[i].pNext = &idp[i];
+
+        vkGetPhysicalDeviceProperties2(devices[i], &prop[i]);
+        av_log(ctx, AV_LOG_VERBOSE, "    %d: %s (%s) (0x%x)\n", i,
+               prop[i].properties.deviceName,
+               vk_dev_type(prop[i].properties.deviceType),
+               prop[i].properties.deviceID);
+    }
+
+    if (select->has_uuid) {
+        for (int i = 0; i < num; i++) {
+            if (!strncmp(idp[i].deviceUUID, select->uuid, VK_UUID_SIZE)) {
+                choice = i;
+                goto end;
+             }
+        }
+        av_log(ctx, AV_LOG_ERROR, "Unable to find device by given UUID!\n");
+        err = AVERROR(ENODEV);
+        goto end;
+    } else if (select->name) {
+        av_log(ctx, AV_LOG_VERBOSE, "Requested device: %s\n", select->name);
+        for (int i = 0; i < num; i++) {
+            if (strstr(prop[i].properties.deviceName, select->name)) {
+                choice = i;
+                goto end;
+             }
+        }
+        av_log(ctx, AV_LOG_ERROR, "Unable to find device \"%s\"!\n",
+               select->name);
+        err = AVERROR(ENODEV);
+        goto end;
+    } else if (select->pci_device) {
+        av_log(ctx, AV_LOG_VERBOSE, "Requested device: 0x%x\n", select->pci_device);
+        for (int i = 0; i < num; i++) {
+            if (select->pci_device == prop[i].properties.deviceID) {
+                choice = i;
+                goto end;
+            }
+        }
+        av_log(ctx, AV_LOG_ERROR, "Unable to find device with PCI ID 0x%x!\n",
+               select->pci_device);
+        err = AVERROR(EINVAL);
+        goto end;
+    } else if (select->vendor_id) {
+        av_log(ctx, AV_LOG_VERBOSE, "Requested vendor: 0x%x\n", select->vendor_id);
+        for (int i = 0; i < num; i++) {
+            if (select->vendor_id == prop[i].properties.vendorID) {
+                choice = i;
+                goto end;
+            }
+        }
+        av_log(ctx, AV_LOG_ERROR, "Unable to find device with Vendor ID 0x%x!\n",
+               select->vendor_id);
+        err = AVERROR(ENODEV);
+        goto end;
+    } else {
+        if (select->index < num) {
+            choice = select->index;
+            goto end;
+        }
+        av_log(ctx, AV_LOG_ERROR, "Unable to find device with index %i!\n",
+               select->index);
+        err = AVERROR(ENODEV);
+        goto end;
+    }
+
+end:
+    if (choice > -1) {
+        p->dev_is_nvidia = (prop[choice].properties.vendorID == 0x10de);
+        hwctx->phys_dev = devices[choice];
+    }
+    av_free(devices);
+    av_free(prop);
+    av_free(idp);
+
+    return err;
+}
+
+static int search_queue_families(AVHWDeviceContext *ctx, VkDeviceCreateInfo *cd)
+{
+    uint32_t num;
+    VkQueueFamilyProperties *qs = NULL;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+    int graph_index = -1, comp_index = -1, tx_index = -1;
+    VkDeviceQueueCreateInfo *pc = (VkDeviceQueueCreateInfo *)cd->pQueueCreateInfos;
+
+    /* First get the number of queue families */
+    vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &num, NULL);
+    if (!num) {
+        av_log(ctx, AV_LOG_ERROR, "Failed to get queues!\n");
+        return AVERROR_EXTERNAL;
+    }
+
+    /* Then allocate memory */
+    qs = av_malloc_array(num, sizeof(VkQueueFamilyProperties));
+    if (!qs)
+        return AVERROR(ENOMEM);
+
+    /* Finally retrieve the queue families */
+    vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &num, qs);
+
+#define SEARCH_FLAGS(expr, out)                                                \
+    for (int i = 0; i < num; i++) {                                            \
+        const VkQueueFlagBits flags = qs[i].queueFlags;                        \
+        if (expr) {                                                            \
+            out = i;                                                           \
+            break;                                                             \
+        }                                                                      \
+    }
+
+    SEARCH_FLAGS(flags & VK_QUEUE_GRAPHICS_BIT, graph_index)
+
+    SEARCH_FLAGS((flags &  VK_QUEUE_COMPUTE_BIT) && (i != graph_index),
+                 comp_index)
+
+    SEARCH_FLAGS((flags & VK_QUEUE_TRANSFER_BIT) && (i != graph_index) &&
+                 (i != comp_index), tx_index)
+
+#undef SEARCH_FLAGS
+#define QF_FLAGS(flags)                                                        \
+    ((flags) & VK_QUEUE_GRAPHICS_BIT      ) ? "(graphics) " : "",              \
+    ((flags) & VK_QUEUE_COMPUTE_BIT       ) ? "(compute) "  : "",              \
+    ((flags) & VK_QUEUE_TRANSFER_BIT      ) ? "(transfer) " : "",              \
+    ((flags) & VK_QUEUE_SPARSE_BINDING_BIT) ? "(sparse) "   : ""
+
+    av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i for graphics, "
+           "flags: %s%s%s%s\n", graph_index, QF_FLAGS(qs[graph_index].queueFlags));
+
+    hwctx->queue_family_index      = graph_index;
+    hwctx->queue_family_tx_index   = graph_index;
+    hwctx->queue_family_comp_index = graph_index;
+
+    pc[cd->queueCreateInfoCount++].queueFamilyIndex = graph_index;
+
+    if (comp_index != -1) {
+        av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i for compute, "
+               "flags: %s%s%s%s\n", comp_index, QF_FLAGS(qs[comp_index].queueFlags));
+        hwctx->queue_family_tx_index                    = comp_index;
+        hwctx->queue_family_comp_index                  = comp_index;
+        pc[cd->queueCreateInfoCount++].queueFamilyIndex = comp_index;
+    }
+
+    if (tx_index != -1) {
+        av_log(ctx, AV_LOG_VERBOSE, "Using queue family %i for transfers, "
+               "flags: %s%s%s%s\n", tx_index, QF_FLAGS(qs[tx_index].queueFlags));
+        hwctx->queue_family_tx_index                    = tx_index;
+        pc[cd->queueCreateInfoCount++].queueFamilyIndex = tx_index;
+    }
+
+#undef QF_FLAGS
+
+    av_free(qs);
+
+    return 0;
+}
+
+static int create_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd,
+                           int queue_family_index)
+{
+    VkResult ret;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+
+    VkCommandPoolCreateInfo cqueue_create = {
+        .sType              = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO,
+        .flags              = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT,
+        .queueFamilyIndex   = queue_family_index,
+    };
+    VkCommandBufferAllocateInfo cbuf_create = {
+        .sType              = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
+        .level              = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
+        .commandBufferCount = 1,
+    };
+
+    VkFenceCreateInfo fence_spawn = {
+        .sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO,
+    };
+
+    ret = vkCreateFence(hwctx->act_dev, &fence_spawn,
+                        hwctx->alloc, &cmd->fence);
+    if (ret != VK_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Failed to create frame fence: %s\n",
+               vk_ret2str(ret));
+        return AVERROR_EXTERNAL;
+    }
+
+    ret = vkCreateCommandPool(hwctx->act_dev, &cqueue_create,
+                              hwctx->alloc, &cmd->pool);
+    if (ret != VK_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Command pool creation failure: %s\n",
+               vk_ret2str(ret));
+        return AVERROR_EXTERNAL;
+    }
+
+    cbuf_create.commandPool = cmd->pool;
+
+    ret = vkAllocateCommandBuffers(hwctx->act_dev, &cbuf_create, &cmd->buf);
+    if (ret != VK_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Command buffer alloc failure: %s\n",
+               vk_ret2str(ret));
+        return AVERROR_EXTERNAL;
+    }
+
+    vkGetDeviceQueue(hwctx->act_dev, cqueue_create.queueFamilyIndex, 0,
+                     &cmd->queue);
+
+    return 0;
+}
+
+static void free_exec_ctx(AVHWDeviceContext *ctx, VulkanExecCtx *cmd)
+{
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+
+    if (cmd->fence)
+        vkDestroyFence(hwctx->act_dev, cmd->fence, hwctx->alloc);
+    if (cmd->buf)
+        vkFreeCommandBuffers(hwctx->act_dev, cmd->pool, 1, &cmd->buf);
+    if (cmd->pool)
+        vkDestroyCommandPool(hwctx->act_dev, cmd->pool, hwctx->alloc);
+}
+
+static void vulkan_device_free(AVHWDeviceContext *ctx)
+{
+    VulkanDevicePriv *p = ctx->internal->priv;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+
+    free_exec_ctx(ctx, &p->cmd);
+
+    vkDestroyDevice(hwctx->act_dev, hwctx->alloc);
+
+    if (p->debug_ctx) {
+        VK_LOAD_PFN(hwctx->inst, vkDestroyDebugUtilsMessengerEXT);
+        pfn_vkDestroyDebugUtilsMessengerEXT(hwctx->inst, p->debug_ctx,
+                                            hwctx->alloc);
+    }
+
+    vkDestroyInstance(hwctx->inst, hwctx->alloc);
+}
+
+static int vulkan_device_create_internal(AVHWDeviceContext *ctx,
+                                         VulkanDeviceSelection *dev_select,
+                                         AVDictionary *opts, int flags)
+{
+    int err = 0;
+    VkResult ret;
+    AVDictionaryEntry *opt_d;
+    VulkanDevicePriv *p = ctx->internal->priv;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+    VkDeviceQueueCreateInfo queue_create_info[3] = {
+        {   .sType            = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
+            .pQueuePriorities = (float []){ 1.0f },
+            .queueCount       = 1, },
+        {   .sType            = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
+            .pQueuePriorities = (float []){ 1.0f },
+            .queueCount       = 1, },
+        {   .sType            = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
+            .pQueuePriorities = (float []){ 1.0f },
+            .queueCount       = 1, },
+    };
+
+    VkDeviceCreateInfo dev_info = {
+        .sType                = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO,
+        .pQueueCreateInfos    = queue_create_info,
+        .queueCreateInfoCount = 0,
+    };
+
+    ctx->free = vulkan_device_free;
+
+    /* Create an instance if not given one */
+    if ((err = create_instance(ctx, opts)))
+        goto end;
+
+    /* Find a device (if not given one) */
+    if ((err = find_device(ctx, dev_select)))
+        goto end;
+
+    vkGetPhysicalDeviceProperties(hwctx->phys_dev, &p->props);
+    av_log(ctx, AV_LOG_VERBOSE, "Using device: %s\n", p->props.deviceName);
+    av_log(ctx, AV_LOG_VERBOSE, "Alignments:\n");
+    av_log(ctx, AV_LOG_VERBOSE, "    optimalBufferCopyOffsetAlignment:   %li\n",
+           p->props.limits.optimalBufferCopyOffsetAlignment);
+    av_log(ctx, AV_LOG_VERBOSE, "    optimalBufferCopyRowPitchAlignment: %li\n",
+           p->props.limits.optimalBufferCopyRowPitchAlignment);
+    av_log(ctx, AV_LOG_VERBOSE, "    minMemoryMapAlignment:              %li\n",
+           p->props.limits.minMemoryMapAlignment);
+
+    /* Search queue family */
+    if ((err = search_queue_families(ctx, &dev_info)))
+        goto end;
+
+    if ((err = check_extensions(ctx, 1, &dev_info.ppEnabledExtensionNames,
+                                &dev_info.enabledExtensionCount, 0)))
+        goto end;
+
+    ret = vkCreateDevice(hwctx->phys_dev, &dev_info, hwctx->alloc,
+                         &hwctx->act_dev);
+
+    av_free((void *)dev_info.ppEnabledExtensionNames);
+
+    if (ret != VK_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Device creation failure: %s\n",
+               vk_ret2str(ret));
+        err = AVERROR_EXTERNAL;
+        goto end;
+    }
+
+    /* Tiled images setting, use them by default */
+    opt_d = av_dict_get(opts, "linear_images", NULL, 0);
+    if (opt_d)
+        p->use_linear_images = strtol(opt_d->value, NULL, 10);
+
+end:
+    return err;
+}
+
+static int vulkan_device_init(AVHWDeviceContext *ctx)
+{
+    int err;
+    uint32_t queue_num;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+    VulkanDevicePriv *p = ctx->internal->priv;
+
+    vkGetPhysicalDeviceQueueFamilyProperties(hwctx->phys_dev, &queue_num, NULL);
+    if (!queue_num) {
+        av_log(ctx, AV_LOG_ERROR, "Failed to get queues!\n");
+        return AVERROR_EXTERNAL;
+    }
+
+#define CHECK_QUEUE(type, n)                                                         \
+if (n >= queue_num) {                                                                \
+    av_log(ctx, AV_LOG_ERROR, "Invalid %s queue index %i (device has %i queues)!\n", \
+           type, n, queue_num);                                                      \
+    return AVERROR(EINVAL);                                                          \
+}
+
+    CHECK_QUEUE("graphics", hwctx->queue_family_index)
+    CHECK_QUEUE("upload",   hwctx->queue_family_tx_index)
+    CHECK_QUEUE("compute",  hwctx->queue_family_comp_index)
+
+#undef CHECK_QUEUE
+
+    /* Create exec context - if there's something invalid this will error out */
+    err = create_exec_ctx(ctx, &p->cmd, hwctx->queue_family_tx_index);
+    if (err)
+        return err;
+
+    /* Get device capabilities */
+    vkGetPhysicalDeviceMemoryProperties(hwctx->phys_dev, &p->mprops);
+
+    return 0;
+}
+
+static int vulkan_device_create(AVHWDeviceContext *ctx, const char *device,
+                                AVDictionary *opts, int flags)
+{
+    VulkanDeviceSelection dev_select = { 0 };
+    if (device && device[0]) {
+        char *end = NULL;
+        dev_select.index = strtol(device, &end, 10);
+        if (end == device) {
+            dev_select.index = 0;
+            dev_select.name  = device;
+        }
+    }
+
+    return vulkan_device_create_internal(ctx, &dev_select, opts, flags);
+}
+
+static int vulkan_device_derive(AVHWDeviceContext *ctx,
+                                AVHWDeviceContext *src_ctx, int flags)
+{
+    av_unused VulkanDeviceSelection dev_select = { 0 };
+
+    /* If there's only one device on the system, then even if its not covered
+     * by the following checks (e.g. non-PCIe ARM GPU), having an empty
+     * dev_select will mean it'll get picked. */
+    switch(src_ctx->type) {
+#if CONFIG_LIBDRM
+#if CONFIG_VAAPI
+    case AV_HWDEVICE_TYPE_VAAPI: {
+        AVVAAPIDeviceContext *src_hwctx = src_ctx->hwctx;
+
+        const char *vendor = vaQueryVendorString(src_hwctx->display);
+        if (!vendor) {
+            av_log(ctx, AV_LOG_ERROR, "Unable to get device info from VAAPI!\n");
+            return AVERROR_EXTERNAL;
+        }
+
+        if (strstr(vendor, "Intel"))
+            dev_select.vendor_id = 0x8086;
+        if (strstr(vendor, "AMD"))
+            dev_select.vendor_id = 0x1002;
+
+        return vulkan_device_create_internal(ctx, &dev_select, NULL, flags);
+    }
+#endif
+    case AV_HWDEVICE_TYPE_DRM: {
+        AVDRMDeviceContext *src_hwctx = src_ctx->hwctx;
+
+        drmDevice *drm_dev_info;
+        int err = drmGetDevice(src_hwctx->fd, &drm_dev_info);
+        if (err) {
+            av_log(ctx, AV_LOG_ERROR, "Unable to get device info from DRM fd!\n");
+            return AVERROR_EXTERNAL;
+        }
+
+        if (drm_dev_info->bustype == DRM_BUS_PCI)
+            dev_select.pci_device = drm_dev_info->deviceinfo.pci->device_id;
+
+        drmFreeDevice(&drm_dev_info);
+
+        return vulkan_device_create_internal(ctx, &dev_select, NULL, flags);
+    }
+#endif
+#if CONFIG_CUDA
+    case AV_HWDEVICE_TYPE_CUDA: {
+        AVHWDeviceContext *cuda_cu = src_ctx;
+        AVCUDADeviceContext *src_hwctx = src_ctx->hwctx;
+        AVCUDADeviceContextInternal *cu_internal = src_hwctx->internal;
+        CudaFunctions *cu = cu_internal->cuda_dl;
+
+        int ret = CHECK_CU(cu->cuDeviceGetUuid((CUuuid *)&dev_select.uuid,
+                                               cu_internal->cuda_device));
+        if (ret < 0) {
+            av_log(ctx, AV_LOG_ERROR, "Unable to get UUID from CUDA!\n");
+            return AVERROR_EXTERNAL;
+        }
+
+        dev_select.has_uuid = 1;
+
+        return vulkan_device_create_internal(ctx, &dev_select, NULL, flags);
+    }
+#endif
+    default:
+        return AVERROR(ENOSYS);
+    }
+}
+
+static int vulkan_frames_get_constraints(AVHWDeviceContext *ctx,
+                                         const void *hwconfig,
+                                         AVHWFramesConstraints *constraints)
+{
+    int count = 0;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+    VulkanDevicePriv *p = ctx->internal->priv;
+
+    for (enum AVPixelFormat i = 0; i < AV_PIX_FMT_NB; i++)
+        count += pixfmt_is_supported(hwctx, i, p->use_linear_images);
+
+#if CONFIG_CUDA
+    if (p->dev_is_nvidia)
+        count++;
+#endif
+
+    constraints->valid_sw_formats = av_malloc_array(count + 1,
+                                                    sizeof(enum AVPixelFormat));
+    if (!constraints->valid_sw_formats)
+        return AVERROR(ENOMEM);
+
+    count = 0;
+    for (enum AVPixelFormat i = 0; i < AV_PIX_FMT_NB; i++)
+        if (pixfmt_is_supported(hwctx, i, p->use_linear_images))
+            constraints->valid_sw_formats[count++] = i;
+
+#if CONFIG_CUDA
+    if (p->dev_is_nvidia)
+        constraints->valid_sw_formats[count++] = AV_PIX_FMT_CUDA;
+#endif
+    constraints->valid_sw_formats[count++] = AV_PIX_FMT_NONE;
+
+    constraints->min_width  = 0;
+    constraints->min_height = 0;
+    constraints->max_width  = p->props.limits.maxImageDimension2D;
+    constraints->max_height = p->props.limits.maxImageDimension2D;
+
+    constraints->valid_hw_formats = av_malloc_array(2, sizeof(enum AVPixelFormat));
+    if (!constraints->valid_hw_formats)
+        return AVERROR(ENOMEM);
+
+    constraints->valid_hw_formats[0] = AV_PIX_FMT_VULKAN;
+    constraints->valid_hw_formats[1] = AV_PIX_FMT_NONE;
+
+    return 0;
+}
+
+static int alloc_mem(AVHWDeviceContext *ctx, VkMemoryRequirements *req,
+                     VkMemoryPropertyFlagBits req_flags, void *alloc_extension,
+                     VkMemoryPropertyFlagBits *mem_flags, VkDeviceMemory *mem)
+{
+    VkResult ret;
+    int index = -1;
+    VulkanDevicePriv *p = ctx->internal->priv;
+    AVVulkanDeviceContext *dev_hwctx = ctx->hwctx;
+    VkMemoryAllocateInfo alloc_info = {
+        .sType           = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
+        .pNext           = alloc_extension,
+    };
+
+    /* Align if we need to */
+    if (req_flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)
+        req->size = FFALIGN(req->size, p->props.limits.minMemoryMapAlignment);
+
+    alloc_info.allocationSize = req->size;
+
+    /* The vulkan spec requires memory types to be sorted in the "optimal"
+     * order, so the first matching type we find will be the best/fastest one */
+    for (int i = 0; i < p->mprops.memoryTypeCount; i++) {
+        /* The memory type must be supported by the requirements (bitfield) */
+        if (!(req->memoryTypeBits & (1 << i)))
+            continue;
+
+        /* The memory type flags must include our properties */
+        if ((p->mprops.memoryTypes[i].propertyFlags & req_flags) != req_flags)
+            continue;
+
+        /* Found a suitable memory type */
+        index = i;
+        break;
+    }
+
+    if (index < 0) {
+        av_log(ctx, AV_LOG_ERROR, "No memory type found for flags 0x%x\n",
+               req_flags);
+        return AVERROR(EINVAL);
+    }
+
+    alloc_info.memoryTypeIndex = index;
+
+    ret = vkAllocateMemory(dev_hwctx->act_dev, &alloc_info,
+                           dev_hwctx->alloc, mem);
+    if (ret != VK_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Failed to allocate memory: %s\n",
+               vk_ret2str(ret));
+        return AVERROR(ENOMEM);
+    }
+
+    *mem_flags |= p->mprops.memoryTypes[index].propertyFlags;
+
+    return 0;
+}
+
+static void vulkan_free_internal(AVVkFrameInternal *internal)
+{
+    if (!internal)
+        return;
+
+#if CONFIG_CUDA
+    if (internal->cuda_fc_ref) {
+        AVHWFramesContext *cuda_fc = (AVHWFramesContext *)internal->cuda_fc_ref->data;
+        int planes = av_pix_fmt_count_planes(cuda_fc->sw_format);
+        AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
+        AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
+        AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
+        CudaFunctions *cu = cu_internal->cuda_dl;
+
+        for (int i = 0; i < planes; i++) {
+            if (internal->cu_sem[i])
+                CHECK_CU(cu->cuDestroyExternalSemaphore(internal->cu_sem[i]));
+            if (internal->cu_mma[i])
+                CHECK_CU(cu->cuMipmappedArrayDestroy(internal->cu_mma[i]));
+            if (internal->ext_mem[i])
+                CHECK_CU(cu->cuDestroyExternalMemory(internal->ext_mem[i]));
+        }
+
+        av_buffer_unref(&internal->cuda_fc_ref);
+    }
+#endif
+
+    av_free(internal);
+}
+
+static void vulkan_frame_free(void *opaque, uint8_t *data)
+{
+    AVVkFrame *f = (AVVkFrame *)data;
+    AVHWFramesContext *hwfc = opaque;
+    AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
+    int planes = av_pix_fmt_count_planes(hwfc->sw_format);
+
+    vulkan_free_internal(f->internal);
+
+    for (int i = 0; i < planes; i++) {
+        vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
+        vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
+        vkDestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc);
+    }
+
+    av_free(f);
+}
+
+static int alloc_bind_mem(AVHWFramesContext *hwfc, AVVkFrame *f,
+                          void *alloc_pnext, size_t alloc_pnext_stride)
+{
+    int err;
+    VkResult ret;
+    AVHWDeviceContext *ctx = hwfc->device_ctx;
+    const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
+    VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS] = { { 0 } };
+
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+
+    for (int i = 0; i < planes; i++) {
+        int use_ded_mem;
+        VkImageMemoryRequirementsInfo2 req_desc = {
+            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_REQUIREMENTS_INFO_2,
+            .image = f->img[i],
+        };
+        VkMemoryDedicatedAllocateInfo ded_alloc = {
+            .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO,
+            .pNext = (void *)(((uint8_t *)alloc_pnext) + i*alloc_pnext_stride),
+        };
+        VkMemoryDedicatedRequirements ded_req = {
+            .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS,
+        };
+        VkMemoryRequirements2 req = {
+            .sType = VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2,
+            .pNext = &ded_req,
+        };
+
+        vkGetImageMemoryRequirements2(hwctx->act_dev, &req_desc, &req);
+
+        /* In case the implementation prefers/requires dedicated allocation */
+        use_ded_mem = ded_req.prefersDedicatedAllocation |
+                      ded_req.requiresDedicatedAllocation;
+        if (use_ded_mem)
+            ded_alloc.image = f->img[i];
+
+        /* Allocate memory */
+        if ((err = alloc_mem(ctx, &req.memoryRequirements,
+                             f->tiling == VK_IMAGE_TILING_LINEAR ?
+                             VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT :
+                             VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
+                             use_ded_mem ? &ded_alloc : (void *)ded_alloc.pNext,
+                             &f->flags, &f->mem[i])))
+            return err;
+
+        f->size[i] = req.memoryRequirements.size;
+        bind_info[i].sType  = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
+        bind_info[i].image  = f->img[i];
+        bind_info[i].memory = f->mem[i];
+    }
+
+    /* Bind the allocated memory to the images */
+    ret = vkBindImageMemory2(hwctx->act_dev, planes, bind_info);
+    if (ret != VK_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
+               vk_ret2str(ret));
+        return AVERROR_EXTERNAL;
+    }
+
+    return 0;
+}
+
+static int prepare_frame(AVHWFramesContext *hwfc, AVVkFrame *frame)
+{
+    VkResult ret;
+    AVHWDeviceContext *ctx = hwfc->device_ctx;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+    VulkanFramesPriv *s = hwfc->internal->priv;
+    const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
+
+    VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
+
+    VkCommandBufferBeginInfo cmd_start = {
+        .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
+        .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
+    };
+
+    VkSubmitInfo s_info = {
+        .sType                = VK_STRUCTURE_TYPE_SUBMIT_INFO,
+        .commandBufferCount   = 1,
+        .pCommandBuffers      = &s->cmd.buf,
+
+        .pSignalSemaphores    = frame->sem,
+        .signalSemaphoreCount = planes,
+    };
+
+    ret = vkBeginCommandBuffer(s->cmd.buf, &cmd_start);
+    if (ret != VK_SUCCESS)
+        return AVERROR_EXTERNAL;
+
+    /* Change the image layout to something more optimal for writes.
+     * This also signals the newly created semaphore, making it usable
+     * for synchronization */
+    for (int i = 0; i < planes; i++) {
+        img_bar[i].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
+        img_bar[i].srcAccessMask = 0x0;
+        img_bar[i].dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
+        img_bar[i].oldLayout = frame->layout[i];
+        img_bar[i].newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
+        img_bar[i].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+        img_bar[i].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+        img_bar[i].image = frame->img[i];
+        img_bar[i].subresourceRange.levelCount = 1;
+        img_bar[i].subresourceRange.layerCount = 1;
+        img_bar[i].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
+
+        frame->layout[i] = img_bar[i].newLayout;
+        frame->access[i] = img_bar[i].dstAccessMask;
+    }
+
+    vkCmdPipelineBarrier(s->cmd.buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+                         VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
+                         0, NULL, 0, NULL, planes, img_bar);
+
+    ret = vkEndCommandBuffer(s->cmd.buf);
+    if (ret != VK_SUCCESS)
+        return AVERROR_EXTERNAL;
+
+    ret = vkQueueSubmit(s->cmd.queue, 1, &s_info, s->cmd.fence);
+    if (ret != VK_SUCCESS) {
+        return AVERROR_EXTERNAL;
+    } else {
+        vkWaitForFences(hwctx->act_dev, 1, &s->cmd.fence, VK_TRUE, UINT64_MAX);
+        vkResetFences(hwctx->act_dev, 1, &s->cmd.fence);
+    }
+
+    return 0;
+}
+
+static int create_frame(AVHWFramesContext *hwfc, AVVkFrame **frame,
+                        VkImageTiling tiling, VkImageUsageFlagBits usage,
+                        void *create_pnext)
+{
+    int err;
+    VkResult ret;
+    AVHWDeviceContext *ctx = hwfc->device_ctx;
+    VulkanDevicePriv *p = ctx->internal->priv;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+    enum AVPixelFormat format = hwfc->sw_format;
+    const VkFormat *img_fmts = av_vkfmt_from_pixfmt(format);
+    const int planes = av_pix_fmt_count_planes(format);
+
+    VkExportSemaphoreCreateInfo ext_sem_info = {
+        .sType = VK_STRUCTURE_TYPE_EXPORT_SEMAPHORE_CREATE_INFO,
+        .handleTypes = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
+    };
+
+    VkSemaphoreCreateInfo sem_spawn = {
+        .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
+        .pNext = p->extensions & EXT_EXTERNAL_FD_SEM ? &ext_sem_info : NULL,
+    };
+
+    AVVkFrame *f = av_vk_frame_alloc();
+    if (!f) {
+        av_log(ctx, AV_LOG_ERROR, "Unable to allocate memory for AVVkFrame!\n");
+        return AVERROR(ENOMEM);
+    }
+
+    /* Create the images */
+    for (int i = 0; i < planes; i++) {
+        const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(format);
+        int w = hwfc->width;
+        int h = hwfc->height;
+        const int p_w = i > 0 ? AV_CEIL_RSHIFT(w, desc->log2_chroma_w) : w;
+        const int p_h = i > 0 ? AV_CEIL_RSHIFT(h, desc->log2_chroma_h) : h;
+
+        VkImageCreateInfo image_create_info = {
+            .sType         = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
+            .pNext         = create_pnext,
+            .imageType     = VK_IMAGE_TYPE_2D,
+            .format        = img_fmts[i],
+            .extent.width  = p_w,
+            .extent.height = p_h,
+            .extent.depth  = 1,
+            .mipLevels     = 1,
+            .arrayLayers   = 1,
+            .flags         = VK_IMAGE_CREATE_ALIAS_BIT,
+            .tiling        = tiling,
+            .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
+            .usage         = usage,
+            .sharingMode   = VK_SHARING_MODE_EXCLUSIVE,
+            .samples       = VK_SAMPLE_COUNT_1_BIT,
+        };
+
+        ret = vkCreateImage(hwctx->act_dev, &image_create_info,
+                            hwctx->alloc, &f->img[i]);
+        if (ret != VK_SUCCESS) {
+            av_log(ctx, AV_LOG_ERROR, "Image creation failure: %s\n",
+                   vk_ret2str(ret));
+            err = AVERROR(EINVAL);
+            goto fail;
+        }
+
+        /* Create semaphore */
+        ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
+                                hwctx->alloc, &f->sem[i]);
+        if (ret != VK_SUCCESS) {
+            av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
+                   vk_ret2str(ret));
+            return AVERROR_EXTERNAL;
+        }
+
+        f->layout[i] = image_create_info.initialLayout;
+        f->access[i] = 0x0;
+    }
+
+    f->flags     = 0x0;
+    f->tiling    = tiling;
+
+    *frame = f;
+    return 0;
+
+fail:
+    vulkan_frame_free(hwfc, (uint8_t *)f);
+    return err;
+}
+
+/* Checks if an export flag is enabled, and if it is ORs it with *iexp */
+static void try_export_flags(AVHWFramesContext *hwfc,
+                             VkExternalMemoryHandleTypeFlags *comp_handle_types,
+                             VkExternalMemoryHandleTypeFlagBits *iexp,
+                             VkExternalMemoryHandleTypeFlagBits exp)
+{
+    VkResult ret;
+    AVVulkanFramesContext *hwctx = hwfc->hwctx;
+    AVVulkanDeviceContext *dev_hwctx = hwfc->device_ctx->hwctx;
+    VkExternalImageFormatProperties eprops = {
+        .sType = VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES_KHR,
+    };
+    VkImageFormatProperties2 props = {
+        .sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_PROPERTIES_2,
+        .pNext = &eprops,
+    };
+    VkPhysicalDeviceExternalImageFormatInfo enext = {
+        .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO,
+        .handleType = exp,
+    };
+    VkPhysicalDeviceImageFormatInfo2 pinfo = {
+        .sType  = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_FORMAT_INFO_2,
+        .pNext  = !exp ? NULL : &enext,
+        .format = av_vkfmt_from_pixfmt(hwfc->sw_format)[0],
+        .type   = VK_IMAGE_TYPE_2D,
+        .tiling = hwctx->tiling,
+        .usage  = hwctx->usage,
+        .flags  = VK_IMAGE_CREATE_ALIAS_BIT,
+    };
+
+    ret = vkGetPhysicalDeviceImageFormatProperties2(dev_hwctx->phys_dev,
+                                                    &pinfo, &props);
+    if (ret == VK_SUCCESS) {
+        *iexp |= exp;
+        *comp_handle_types |= eprops.externalMemoryProperties.compatibleHandleTypes;
+    }
+}
+
+static AVBufferRef *vulkan_pool_alloc(void *opaque, int size)
+{
+    int err;
+    AVVkFrame *f;
+    AVBufferRef *avbuf = NULL;
+    AVHWFramesContext *hwfc = opaque;
+    AVVulkanFramesContext *hwctx = hwfc->hwctx;
+    VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
+    VkExportMemoryAllocateInfo eminfo[AV_NUM_DATA_POINTERS];
+    VkExternalMemoryHandleTypeFlags e = 0x0;
+
+    VkExternalMemoryImageCreateInfo eiinfo = {
+        .sType       = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
+        .pNext       = hwctx->create_pnext,
+    };
+
+    if (p->extensions & EXT_EXTERNAL_FD_MEMORY)
+        try_export_flags(hwfc, &eiinfo.handleTypes, &e,
+                         VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT);
+
+    if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
+        try_export_flags(hwfc, &eiinfo.handleTypes, &e,
+                         VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
+
+    for (int i = 0; i < av_pix_fmt_count_planes(hwfc->sw_format); i++) {
+        eminfo[i].sType       = VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO;
+        eminfo[i].pNext       = hwctx->alloc_pnext[i];
+        eminfo[i].handleTypes = e;
+    }
+
+    err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage,
+                       eiinfo.handleTypes ? &eiinfo : NULL);
+    if (err)
+        return NULL;
+
+    err = alloc_bind_mem(hwfc, f, eminfo, sizeof(*eminfo));
+    if (err)
+        goto fail;
+
+    err = prepare_frame(hwfc, f);
+    if (err)
+        goto fail;
+
+    avbuf = av_buffer_create((uint8_t *)f, sizeof(AVVkFrame),
+                             vulkan_frame_free, hwfc, 0);
+    if (!avbuf)
+        goto fail;
+
+    return avbuf;
+
+fail:
+    vulkan_frame_free(hwfc, (uint8_t *)f);
+    return NULL;
+}
+
+static void vulkan_frames_uninit(AVHWFramesContext *hwfc)
+{
+    VulkanFramesPriv *fp = hwfc->internal->priv;
+
+    free_exec_ctx(hwfc->device_ctx, &fp->cmd);
+}
+
+static int vulkan_frames_init(AVHWFramesContext *hwfc)
+{
+    int err;
+    AVVkFrame *f;
+    AVVulkanFramesContext *hwctx = hwfc->hwctx;
+    VulkanFramesPriv *fp = hwfc->internal->priv;
+    AVVulkanDeviceContext *dev_hwctx = hwfc->device_ctx->hwctx;
+    VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
+
+    if (hwfc->pool)
+        return 0;
+
+    /* Default pool flags */
+    hwctx->tiling = hwctx->tiling ? hwctx->tiling : p->use_linear_images ?
+                    VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
+
+    hwctx->usage |= DEFAULT_USAGE_FLAGS;
+
+    err = create_exec_ctx(hwfc->device_ctx, &fp->cmd,
+                          dev_hwctx->queue_family_tx_index);
+    if (err)
+        return err;
+
+    /* Test to see if allocation will fail */
+    err = create_frame(hwfc, &f, hwctx->tiling, hwctx->usage,
+                       hwctx->create_pnext);
+    if (err) {
+        free_exec_ctx(hwfc->device_ctx, &p->cmd);
+        return err;
+    }
+
+    vulkan_frame_free(hwfc, (uint8_t *)f);
+
+    hwfc->internal->pool_internal = av_buffer_pool_init2(sizeof(AVVkFrame),
+                                                         hwfc, vulkan_pool_alloc,
+                                                         NULL);
+    if (!hwfc->internal->pool_internal) {
+        free_exec_ctx(hwfc->device_ctx, &p->cmd);
+        return AVERROR(ENOMEM);
+    }
+
+    return 0;
+}
+
+static int vulkan_get_buffer(AVHWFramesContext *hwfc, AVFrame *frame)
+{
+    frame->buf[0] = av_buffer_pool_get(hwfc->pool);
+    if (!frame->buf[0])
+        return AVERROR(ENOMEM);
+
+    frame->data[0] = frame->buf[0]->data;
+    frame->format  = AV_PIX_FMT_VULKAN;
+    frame->width   = hwfc->width;
+    frame->height  = hwfc->height;
+
+    return 0;
+}
+
+static int vulkan_transfer_get_formats(AVHWFramesContext *hwfc,
+                                       enum AVHWFrameTransferDirection dir,
+                                       enum AVPixelFormat **formats)
+{
+    enum AVPixelFormat *fmts = av_malloc_array(2, sizeof(*fmts));
+    if (!fmts)
+        return AVERROR(ENOMEM);
+
+    fmts[0] = hwfc->sw_format;
+    fmts[1] = AV_PIX_FMT_NONE;
+
+    *formats = fmts;
+    return 0;
+}
+
+typedef struct VulkanMapping {
+    AVVkFrame *frame;
+    int flags;
+} VulkanMapping;
+
+static void vulkan_unmap_frame(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
+{
+    VulkanMapping *map = hwmap->priv;
+    AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
+    const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
+
+    /* Check if buffer needs flushing */
+    if ((map->flags & AV_HWFRAME_MAP_WRITE) &&
+        !(map->frame->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
+        VkResult ret;
+        VkMappedMemoryRange flush_ranges[AV_NUM_DATA_POINTERS] = { { 0 } };
+
+        for (int i = 0; i < planes; i++) {
+            flush_ranges[i].sType  = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
+            flush_ranges[i].memory = map->frame->mem[i];
+            flush_ranges[i].size   = VK_WHOLE_SIZE;
+        }
+
+        ret = vkFlushMappedMemoryRanges(hwctx->act_dev, planes,
+                                        flush_ranges);
+        if (ret != VK_SUCCESS) {
+            av_log(hwfc, AV_LOG_ERROR, "Failed to flush memory: %s\n",
+                   vk_ret2str(ret));
+        }
+    }
+
+    for (int i = 0; i < planes; i++)
+        vkUnmapMemory(hwctx->act_dev, map->frame->mem[i]);
+
+    av_free(map);
+}
+
+static int vulkan_map_frame_to_mem(AVHWFramesContext *hwfc, AVFrame *dst,
+                                   const AVFrame *src, int flags)
+{
+    VkResult ret;
+    int err, mapped_mem_count = 0;
+    AVVkFrame *f = (AVVkFrame *)src->data[0];
+    AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
+    const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
+
+    VulkanMapping *map = av_mallocz(sizeof(VulkanMapping));
+    if (!map)
+        return AVERROR(EINVAL);
+
+    if (src->format != AV_PIX_FMT_VULKAN) {
+        av_log(hwfc, AV_LOG_ERROR, "Cannot map from pixel format %s!\n",
+               av_get_pix_fmt_name(src->format));
+        err = AVERROR(EINVAL);
+        goto fail;
+    }
+
+    if (!(f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) ||
+        !(f->tiling == VK_IMAGE_TILING_LINEAR)) {
+        av_log(hwfc, AV_LOG_ERROR, "Unable to map frame, not host visible "
+               "and linear!\n");
+        err = AVERROR(EINVAL);
+        goto fail;
+    }
+
+    dst->width  = src->width;
+    dst->height = src->height;
+
+    for (int i = 0; i < planes; i++) {
+        ret = vkMapMemory(hwctx->act_dev, f->mem[i], 0,
+                          VK_WHOLE_SIZE, 0, (void **)&dst->data[i]);
+        if (ret != VK_SUCCESS) {
+            av_log(hwfc, AV_LOG_ERROR, "Failed to map image memory: %s\n",
+                vk_ret2str(ret));
+            err = AVERROR_EXTERNAL;
+            goto fail;
+        }
+        mapped_mem_count++;
+    }
+
+    /* Check if the memory contents matter */
+    if (((flags & AV_HWFRAME_MAP_READ) || !(flags & AV_HWFRAME_MAP_OVERWRITE)) &&
+        !(f->flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)) {
+        VkMappedMemoryRange map_mem_ranges[AV_NUM_DATA_POINTERS] = { { 0 } };
+        for (int i = 0; i < planes; i++) {
+            map_mem_ranges[i].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
+            map_mem_ranges[i].size = VK_WHOLE_SIZE;
+            map_mem_ranges[i].memory = f->mem[i];
+        }
+
+        ret = vkInvalidateMappedMemoryRanges(hwctx->act_dev, planes,
+                                             map_mem_ranges);
+        if (ret != VK_SUCCESS) {
+            av_log(hwfc, AV_LOG_ERROR, "Failed to invalidate memory: %s\n",
+                   vk_ret2str(ret));
+            err = AVERROR_EXTERNAL;
+            goto fail;
+        }
+    }
+
+    for (int i = 0; i < planes; i++) {
+        VkImageSubresource sub = {
+            .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+        };
+        VkSubresourceLayout layout;
+        vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
+        dst->linesize[i] = layout.rowPitch;
+    }
+
+    map->frame = f;
+    map->flags = flags;
+
+    err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src,
+                                &vulkan_unmap_frame, map);
+    if (err < 0)
+        goto fail;
+
+    return 0;
+
+fail:
+    for (int i = 0; i < mapped_mem_count; i++)
+        vkUnmapMemory(hwctx->act_dev, f->mem[i]);
+
+    av_free(map);
+    return err;
+}
+
+#if CONFIG_LIBDRM
+static void vulkan_unmap_from(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
+{
+    VulkanMapping *map = hwmap->priv;
+    AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
+    const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
+
+    for (int i = 0; i < planes; i++) {
+        vkDestroyImage(hwctx->act_dev, map->frame->img[i], hwctx->alloc);
+        vkFreeMemory(hwctx->act_dev, map->frame->mem[i], hwctx->alloc);
+        vkDestroySemaphore(hwctx->act_dev, map->frame->sem[i], hwctx->alloc);
+    }
+
+    av_freep(&map->frame);
+}
+
+static const struct {
+    uint32_t drm_fourcc;
+    VkFormat vk_format;
+} vulkan_drm_format_map[] = {
+    { DRM_FORMAT_R8,       VK_FORMAT_R8_UNORM       },
+    { DRM_FORMAT_R16,      VK_FORMAT_R16_UNORM      },
+    { DRM_FORMAT_GR88,     VK_FORMAT_R8G8_UNORM     },
+    { DRM_FORMAT_RG88,     VK_FORMAT_R8G8_UNORM     },
+    { DRM_FORMAT_GR1616,   VK_FORMAT_R16G16_UNORM   },
+    { DRM_FORMAT_RG1616,   VK_FORMAT_R16G16_UNORM   },
+    { DRM_FORMAT_ARGB8888, VK_FORMAT_B8G8R8A8_UNORM },
+    { DRM_FORMAT_XRGB8888, VK_FORMAT_B8G8R8A8_UNORM },
+    { DRM_FORMAT_ABGR8888, VK_FORMAT_R8G8B8A8_UNORM },
+    { DRM_FORMAT_XBGR8888, VK_FORMAT_R8G8B8A8_UNORM },
+};
+
+static inline VkFormat drm_to_vulkan_fmt(uint32_t drm_fourcc)
+{
+    for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
+        if (vulkan_drm_format_map[i].drm_fourcc == drm_fourcc)
+            return vulkan_drm_format_map[i].vk_format;
+    return VK_FORMAT_UNDEFINED;
+}
+
+static int vulkan_map_from_drm_frame_desc(AVHWFramesContext *hwfc, AVVkFrame **frame,
+                                          AVDRMFrameDescriptor *desc)
+{
+    int err = 0;
+    VkResult ret;
+    AVVkFrame *f;
+    AVHWDeviceContext *ctx = hwfc->device_ctx;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+    VulkanDevicePriv *p = ctx->internal->priv;
+    const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
+    const AVPixFmtDescriptor *fmt_desc = av_pix_fmt_desc_get(hwfc->sw_format);
+    const int has_modifiers = p->extensions & EXT_DRM_MODIFIER_FLAGS;
+    VkSubresourceLayout plane_data[AV_NUM_DATA_POINTERS];
+    VkBindImageMemoryInfo bind_info[AV_NUM_DATA_POINTERS];
+    VkExternalMemoryHandleTypeFlagBits htype = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT;
+
+    VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdPropertiesKHR);
+
+    for (int i = 0; i < desc->nb_layers; i++) {
+        if (desc->layers[i].nb_planes > 1) {
+            av_log(ctx, AV_LOG_ERROR, "Cannot import DMABUFS with more than 1 "
+                                      "plane per layer!\n");
+            return AVERROR(EINVAL);
+        }
+
+        if (drm_to_vulkan_fmt(desc->layers[i].format) == VK_FORMAT_UNDEFINED) {
+            av_log(ctx, AV_LOG_ERROR, "Unsupported DMABUF layer format %#08x!\n",
+                   desc->layers[i].format);
+            return AVERROR(EINVAL);
+        }
+    }
+
+    if (!(f = av_vk_frame_alloc())) {
+        av_log(ctx, AV_LOG_ERROR, "Unable to allocate memory for AVVkFrame!\n");
+        err = AVERROR(ENOMEM);
+        goto fail;
+    }
+
+    for (int i = 0; i < desc->nb_objects; i++) {
+        VkMemoryFdPropertiesKHR fdmp = {
+            .sType = VK_STRUCTURE_TYPE_MEMORY_FD_PROPERTIES_KHR,
+        };
+        VkMemoryRequirements req = {
+            .size = desc->objects[i].size,
+        };
+        VkImportMemoryFdInfoKHR idesc = {
+            .sType      = VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR,
+            .handleType = htype,
+            .fd         = desc->objects[i].fd,
+        };
+
+        ret = pfn_vkGetMemoryFdPropertiesKHR(hwctx->act_dev, htype,
+                                             desc->objects[i].fd, &fdmp);
+        if (ret != VK_SUCCESS) {
+            av_log(hwfc, AV_LOG_ERROR, "Failed to get FD properties: %s\n",
+                   vk_ret2str(ret));
+            err = AVERROR_EXTERNAL;
+            goto fail;
+        }
+
+        req.memoryTypeBits = fdmp.memoryTypeBits;
+
+        err = alloc_mem(ctx, &req, 0x0, &idesc, &f->flags, &f->mem[i]);
+        if (err)
+            return err;
+
+        f->size[i] = desc->objects[i].size;
+    }
+
+    f->tiling = has_modifiers ? VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT :
+                desc->objects[0].format_modifier == DRM_FORMAT_MOD_LINEAR ?
+                VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
+
+    for (int i = 0; i < desc->nb_layers; i++) {
+        VkImageDrmFormatModifierExplicitCreateInfoEXT drm_info = {
+            .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_EXPLICIT_CREATE_INFO_EXT,
+            .drmFormatModifier = desc->objects[0].format_modifier,
+            .drmFormatModifierPlaneCount = desc->layers[i].nb_planes,
+            .pPlaneLayouts = (const VkSubresourceLayout *)&plane_data,
+        };
+
+        VkExternalMemoryImageCreateInfo einfo = {
+            .sType       = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO,
+            .pNext       = has_modifiers ? &drm_info : NULL,
+            .handleTypes = htype,
+        };
+
+        VkSemaphoreCreateInfo sem_spawn = {
+            .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
+        };
+
+        const int p_w = i > 0 ? AV_CEIL_RSHIFT(hwfc->width, fmt_desc->log2_chroma_w) : hwfc->width;
+        const int p_h = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, fmt_desc->log2_chroma_h) : hwfc->height;
+
+        VkImageCreateInfo image_create_info = {
+            .sType         = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
+            .pNext         = &einfo,
+            .imageType     = VK_IMAGE_TYPE_2D,
+            .format        = drm_to_vulkan_fmt(desc->layers[i].format),
+            .extent.width  = p_w,
+            .extent.height = p_h,
+            .extent.depth  = 1,
+            .mipLevels     = 1,
+            .arrayLayers   = 1,
+            .flags         = VK_IMAGE_CREATE_ALIAS_BIT,
+            .tiling        = f->tiling,
+            .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED, /* specs say so */
+            .usage         = DEFAULT_USAGE_FLAGS,
+            .sharingMode   = VK_SHARING_MODE_EXCLUSIVE,
+            .samples       = VK_SAMPLE_COUNT_1_BIT,
+        };
+
+        for (int j = 0; j < desc->layers[i].nb_planes; j++) {
+            plane_data[j].offset     = desc->layers[i].planes[j].offset;
+            plane_data[j].rowPitch   = desc->layers[i].planes[j].pitch;
+            plane_data[j].size       = 0; /* The specs say so for all 3 */
+            plane_data[j].arrayPitch = 0;
+            plane_data[j].depthPitch = 0;
+        }
+
+        /* Create image */
+        ret = vkCreateImage(hwctx->act_dev, &image_create_info,
+                            hwctx->alloc, &f->img[i]);
+        if (ret != VK_SUCCESS) {
+            av_log(ctx, AV_LOG_ERROR, "Image creation failure: %s\n",
+                   vk_ret2str(ret));
+            err = AVERROR(EINVAL);
+            goto fail;
+        }
+
+        ret = vkCreateSemaphore(hwctx->act_dev, &sem_spawn,
+                                hwctx->alloc, &f->sem[i]);
+        if (ret != VK_SUCCESS) {
+            av_log(hwctx, AV_LOG_ERROR, "Failed to create semaphore: %s\n",
+                   vk_ret2str(ret));
+            return AVERROR_EXTERNAL;
+        }
+
+        /* We'd import a semaphore onto the one we created using
+         * vkImportSemaphoreFdKHR but unfortunately neither DRM nor VAAPI
+         * offer us anything we could import and sync with, so instead
+         * leave the semaphore unsignalled and enjoy the validation spam. */
+
+        f->layout[i] = image_create_info.initialLayout;
+        f->access[i] = 0x0;
+
+        /* TODO: Fix to support more than 1 plane per layer */
+        bind_info[i].sType  = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO;
+        bind_info[i].pNext  = NULL;
+        bind_info[i].image  = f->img[i];
+        bind_info[i].memory = f->mem[desc->layers[i].planes[0].object_index];
+        bind_info[i].memoryOffset = desc->layers[i].planes[0].offset;
+    }
+
+    /* Bind the allocated memory to the images */
+    ret = vkBindImageMemory2(hwctx->act_dev, planes, bind_info);
+    if (ret != VK_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Failed to bind memory: %s\n",
+               vk_ret2str(ret));
+        return AVERROR_EXTERNAL;
+    }
+
+    *frame = f;
+
+    return 0;
+
+fail:
+    for (int i = 0; i < planes; i++) {
+        vkDestroyImage(hwctx->act_dev, f->img[i], hwctx->alloc);
+        vkFreeMemory(hwctx->act_dev, f->mem[i], hwctx->alloc);
+        vkDestroySemaphore(hwctx->act_dev, f->sem[i], hwctx->alloc);
+    }
+
+    av_free(f);
+
+    return err;
+}
+
+static int vulkan_map_from_drm(AVHWFramesContext *hwfc, AVFrame *dst,
+                               const AVFrame *src, int flags)
+{
+    int err = 0;
+    AVVkFrame *f;
+    VulkanMapping *map = NULL;
+
+    err = vulkan_map_from_drm_frame_desc(hwfc, &f,
+                                         (AVDRMFrameDescriptor *)src->data[0]);
+    if (err)
+        return err;
+
+    /* The unmapping function will free this */
+    dst->data[0] = (uint8_t *)f;
+    dst->width   = src->width;
+    dst->height  = src->height;
+
+    map = av_mallocz(sizeof(VulkanMapping));
+    if (!map)
+        goto fail;
+
+    map->frame = f;
+    map->flags = flags;
+
+    err = ff_hwframe_map_create(dst->hw_frames_ctx, dst, src,
+                                &vulkan_unmap_from, map);
+    if (err < 0)
+        goto fail;
+
+    av_log(hwfc, AV_LOG_DEBUG, "Mapped DRM object to Vulkan!\n");
+
+    return 0;
+
+fail:
+    vulkan_frame_free(hwfc->device_ctx->hwctx, (uint8_t *)f);
+    av_free(map);
+    return err;
+}
+
+#if CONFIG_VAAPI
+static int vulkan_map_from_vaapi(AVHWFramesContext *dst_fc,
+                                 AVFrame *dst, const AVFrame *src,
+                                 int flags)
+{
+    int err;
+    AVFrame *tmp = av_frame_alloc();
+    AVHWFramesContext *vaapi_fc = (AVHWFramesContext*)src->hw_frames_ctx->data;
+    AVVAAPIDeviceContext *vaapi_ctx = vaapi_fc->device_ctx->hwctx;
+    VASurfaceID surface_id = (VASurfaceID)(uintptr_t)src->data[3];
+
+    if (!tmp)
+        return AVERROR(ENOMEM);
+
+    /* We have to sync since like the previous comment said, no semaphores */
+    vaSyncSurface(vaapi_ctx->display, surface_id);
+
+    tmp->format = AV_PIX_FMT_DRM_PRIME;
+
+    err = av_hwframe_map(tmp, src, flags);
+    if (err < 0)
+        goto fail;
+
+    err = vulkan_map_from_drm(dst_fc, dst, tmp, flags);
+    if (err < 0)
+        goto fail;
+
+    err = ff_hwframe_map_replace(dst, src);
+
+fail:
+    av_frame_free(&tmp);
+    return err;
+}
+#endif
+#endif
+
+#if CONFIG_CUDA
+static int vulkan_export_to_cuda(AVHWFramesContext *hwfc,
+                                 AVBufferRef *cuda_hwfc,
+                                 const AVFrame *frame)
+{
+    int err;
+    VkResult ret;
+    AVVkFrame *dst_f;
+    AVVkFrameInternal *dst_int;
+    AVHWDeviceContext *ctx = hwfc->device_ctx;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+    const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
+    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
+    VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdKHR);
+    VK_LOAD_PFN(hwctx->inst, vkGetSemaphoreFdKHR);
+
+    AVHWFramesContext *cuda_fc = (AVHWFramesContext*)cuda_hwfc->data;
+    AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
+    AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
+    AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
+    CudaFunctions *cu = cu_internal->cuda_dl;
+    CUarray_format cufmt = desc->comp[0].depth > 8 ? CU_AD_FORMAT_UNSIGNED_INT16 :
+                                                     CU_AD_FORMAT_UNSIGNED_INT8;
+
+    dst_f = (AVVkFrame *)frame->data[0];
+
+    dst_int = dst_f->internal;
+    if (!dst_int || !dst_int->cuda_fc_ref) {
+        if (!dst_f->internal)
+            dst_f->internal = dst_int = av_mallocz(sizeof(*dst_f->internal));
+
+        if (!dst_int) {
+            err = AVERROR(ENOMEM);
+            goto fail;
+        }
+
+        dst_int->cuda_fc_ref = av_buffer_ref(cuda_hwfc);
+        if (!dst_int->cuda_fc_ref) {
+            err = AVERROR(ENOMEM);
+            goto fail;
+        }
+
+        for (int i = 0; i < planes; i++) {
+            CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC tex_desc = {
+                .offset = 0,
+                .arrayDesc = {
+                    .Width  = i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
+                                    : hwfc->width,
+                    .Height = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
+                                    : hwfc->height,
+                    .Depth = 0,
+                    .Format = cufmt,
+                    .NumChannels = 1 + ((planes == 2) && i),
+                    .Flags = 0,
+                },
+                .numLevels = 1,
+            };
+            CUDA_EXTERNAL_MEMORY_HANDLE_DESC ext_desc = {
+                .type = CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD,
+                .size = dst_f->size[i],
+            };
+            VkMemoryGetFdInfoKHR export_info = {
+                .sType      = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
+                .memory     = dst_f->mem[i],
+                .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR,
+            };
+            VkSemaphoreGetFdInfoKHR sem_export = {
+                .sType = VK_STRUCTURE_TYPE_SEMAPHORE_GET_FD_INFO_KHR,
+                .semaphore = dst_f->sem[i],
+                .handleType = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT,
+            };
+            CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC ext_sem_desc = {
+                .type = CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD,
+            };
+
+            ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
+                                       &ext_desc.handle.fd);
+            if (ret != VK_SUCCESS) {
+                av_log(hwfc, AV_LOG_ERROR, "Unable to export the image as a FD!\n");
+                err = AVERROR_EXTERNAL;
+                goto fail;
+            }
+
+            ret = CHECK_CU(cu->cuImportExternalMemory(&dst_int->ext_mem[i], &ext_desc));
+            if (ret < 0) {
+                err = AVERROR_EXTERNAL;
+                goto fail;
+            }
+
+            ret = CHECK_CU(cu->cuExternalMemoryGetMappedMipmappedArray(&dst_int->cu_mma[i],
+                                                                       dst_int->ext_mem[i],
+                                                                       &tex_desc));
+            if (ret < 0) {
+                err = AVERROR_EXTERNAL;
+                goto fail;
+            }
+
+            ret = CHECK_CU(cu->cuMipmappedArrayGetLevel(&dst_int->cu_array[i],
+                                                        dst_int->cu_mma[i], 0));
+            if (ret < 0) {
+                err = AVERROR_EXTERNAL;
+                goto fail;
+            }
+
+            ret = pfn_vkGetSemaphoreFdKHR(hwctx->act_dev, &sem_export,
+                                          &ext_sem_desc.handle.fd);
+            if (ret != VK_SUCCESS) {
+                av_log(ctx, AV_LOG_ERROR, "Failed to export semaphore: %s\n",
+                       vk_ret2str(ret));
+                err = AVERROR_EXTERNAL;
+                goto fail;
+            }
+
+            ret = CHECK_CU(cu->cuImportExternalSemaphore(&dst_int->cu_sem[i],
+                                                         &ext_sem_desc));
+            if (ret < 0) {
+                err = AVERROR_EXTERNAL;
+                goto fail;
+            }
+        }
+    }
+
+    return 0;
+
+fail:
+    return err;
+}
+
+static int vulkan_transfer_data_from_cuda(AVHWFramesContext *hwfc,
+                                          AVFrame *dst, const AVFrame *src)
+{
+    int err;
+    VkResult ret;
+    CUcontext dummy;
+    AVVkFrame *dst_f;
+    AVVkFrameInternal *dst_int;
+    const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
+    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
+
+    AVHWFramesContext *cuda_fc = (AVHWFramesContext*)src->hw_frames_ctx->data;
+    AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
+    AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
+    AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
+    CudaFunctions *cu = cu_internal->cuda_dl;
+    CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS s_w_par[AV_NUM_DATA_POINTERS] = { 0 };
+    CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS s_s_par[AV_NUM_DATA_POINTERS] = { 0 };
+
+    ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
+    if (ret < 0) {
+        err = AVERROR_EXTERNAL;
+        goto fail;
+    }
+
+    dst_f = (AVVkFrame *)dst->data[0];
+
+    ret = vulkan_export_to_cuda(hwfc, src->hw_frames_ctx, dst);
+    if (ret < 0) {
+        goto fail;
+    }
+    dst_int = dst_f->internal;
+
+    ret = CHECK_CU(cu->cuWaitExternalSemaphoresAsync(dst_int->cu_sem, s_w_par,
+                                                     planes, cuda_dev->stream));
+    if (ret < 0) {
+        err = AVERROR_EXTERNAL;
+        goto fail;
+    }
+
+    for (int i = 0; i < planes; i++) {
+        CUDA_MEMCPY2D cpy = {
+            .srcMemoryType = CU_MEMORYTYPE_DEVICE,
+            .srcDevice     = (CUdeviceptr)src->data[i],
+            .srcPitch      = src->linesize[i],
+            .srcY          = 0,
+
+            .dstMemoryType = CU_MEMORYTYPE_ARRAY,
+            .dstArray      = dst_int->cu_array[i],
+            .WidthInBytes  = (i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
+                                    : hwfc->width) * desc->comp[i].step,
+            .Height        = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
+                                   : hwfc->height,
+        };
+
+        ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream));
+        if (ret < 0) {
+            err = AVERROR_EXTERNAL;
+            goto fail;
+        }
+    }
+
+    ret = CHECK_CU(cu->cuSignalExternalSemaphoresAsync(dst_int->cu_sem, s_s_par,
+                                                       planes, cuda_dev->stream));
+    if (ret < 0) {
+        err = AVERROR_EXTERNAL;
+        goto fail;
+    }
+
+    CHECK_CU(cu->cuCtxPopCurrent(&dummy));
+
+    av_log(hwfc, AV_LOG_VERBOSE, "Transfered CUDA image to Vulkan!\n");
+
+    return 0;
+
+fail:
+    CHECK_CU(cu->cuCtxPopCurrent(&dummy));
+    vulkan_free_internal(dst_int);
+    dst_f->internal = NULL;
+    av_buffer_unref(&dst->buf[0]);
+    return err;
+}
+#endif
+
+static int vulkan_map_to(AVHWFramesContext *hwfc, AVFrame *dst,
+                         const AVFrame *src, int flags)
+{
+    av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
+
+    switch (src->format) {
+#if CONFIG_LIBDRM
+#if CONFIG_VAAPI
+    case AV_PIX_FMT_VAAPI:
+        if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
+            return vulkan_map_from_vaapi(hwfc, dst, src, flags);
+#endif
+    case AV_PIX_FMT_DRM_PRIME:
+        if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
+            return vulkan_map_from_drm(hwfc, dst, src, flags);
+#endif
+    default:
+        return AVERROR(ENOSYS);
+    }
+}
+
+#if CONFIG_LIBDRM
+typedef struct VulkanDRMMapping {
+    AVDRMFrameDescriptor drm_desc;
+    AVVkFrame *source;
+} VulkanDRMMapping;
+
+static void vulkan_unmap_to_drm(AVHWFramesContext *hwfc, HWMapDescriptor *hwmap)
+{
+    AVDRMFrameDescriptor *drm_desc = hwmap->priv;
+
+    for (int i = 0; i < drm_desc->nb_objects; i++)
+        close(drm_desc->objects[i].fd);
+
+    av_free(drm_desc);
+}
+
+static inline uint32_t vulkan_fmt_to_drm(VkFormat vkfmt)
+{
+    for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
+        if (vulkan_drm_format_map[i].vk_format == vkfmt)
+            return vulkan_drm_format_map[i].drm_fourcc;
+    return DRM_FORMAT_INVALID;
+}
+
+static int vulkan_map_to_drm(AVHWFramesContext *hwfc, AVFrame *dst,
+                             const AVFrame *src, int flags)
+{
+    int err = 0;
+    VkResult ret;
+    AVVkFrame *f = (AVVkFrame *)src->data[0];
+    VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
+    AVVulkanDeviceContext *hwctx = hwfc->device_ctx->hwctx;
+    const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
+    VK_LOAD_PFN(hwctx->inst, vkGetMemoryFdKHR);
+    VkImageDrmFormatModifierPropertiesEXT drm_mod = {
+        .sType = VK_STRUCTURE_TYPE_IMAGE_DRM_FORMAT_MODIFIER_PROPERTIES_EXT,
+    };
+
+    AVDRMFrameDescriptor *drm_desc = av_mallocz(sizeof(*drm_desc));
+    if (!drm_desc)
+        return AVERROR(ENOMEM);
+
+    err = ff_hwframe_map_create(src->hw_frames_ctx, dst, src, &vulkan_unmap_to_drm, drm_desc);
+    if (err < 0)
+        goto end;
+
+    if (p->extensions & EXT_DRM_MODIFIER_FLAGS) {
+        VK_LOAD_PFN(hwctx->inst, vkGetImageDrmFormatModifierPropertiesEXT);
+        ret = pfn_vkGetImageDrmFormatModifierPropertiesEXT(hwctx->act_dev, f->img[0],
+                                                           &drm_mod);
+        if (ret != VK_SUCCESS) {
+            av_log(hwfc, AV_LOG_ERROR, "Failed to retrieve DRM format modifier!\n");
+            err = AVERROR_EXTERNAL;
+            goto end;
+        }
+    }
+
+    for (int i = 0; (i < planes) && (f->mem[i]); i++) {
+        VkMemoryGetFdInfoKHR export_info = {
+            .sType      = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR,
+            .memory     = f->mem[i],
+            .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT,
+        };
+
+        ret = pfn_vkGetMemoryFdKHR(hwctx->act_dev, &export_info,
+                                   &drm_desc->objects[i].fd);
+        if (ret != VK_SUCCESS) {
+            av_log(hwfc, AV_LOG_ERROR, "Unable to export the image as a FD!\n");
+            err = AVERROR_EXTERNAL;
+            goto end;
+        }
+
+        drm_desc->nb_objects++;
+        drm_desc->objects[i].size = f->size[i];
+        drm_desc->objects[i].format_modifier = drm_mod.drmFormatModifier;
+    }
+
+    drm_desc->nb_layers = planes;
+    for (int i = 0; i < drm_desc->nb_layers; i++) {
+        VkSubresourceLayout layout;
+        VkImageSubresource sub = {
+            .aspectMask = p->extensions & EXT_DRM_MODIFIER_FLAGS ?
+                          VK_IMAGE_ASPECT_MEMORY_PLANE_0_BIT_EXT :
+                          VK_IMAGE_ASPECT_COLOR_BIT,
+        };
+        VkFormat plane_vkfmt = av_vkfmt_from_pixfmt(hwfc->sw_format)[i];
+
+        drm_desc->layers[i].format    = vulkan_fmt_to_drm(plane_vkfmt);
+        drm_desc->layers[i].nb_planes = 1;
+
+        if (drm_desc->layers[i].format == DRM_FORMAT_INVALID) {
+            av_log(hwfc, AV_LOG_ERROR, "Cannot map to DRM layer, unsupported!\n");
+            err = AVERROR_PATCHWELCOME;
+            goto end;
+        }
+
+        drm_desc->layers[i].planes[0].object_index = FFMIN(i, drm_desc->nb_objects - 1);
+
+        if (f->tiling != VK_IMAGE_TILING_OPTIMAL)
+            continue;
+
+        vkGetImageSubresourceLayout(hwctx->act_dev, f->img[i], &sub, &layout);
+        drm_desc->layers[i].planes[0].offset       = layout.offset;
+        drm_desc->layers[i].planes[0].pitch        = layout.rowPitch;
+    }
+
+    dst->width   = src->width;
+    dst->height  = src->height;
+    dst->data[0] = (uint8_t *)drm_desc;
+
+    av_log(hwfc, AV_LOG_VERBOSE, "Mapped AVVkFrame to a DRM object!\n");
+
+    return 0;
+
+end:
+    av_free(drm_desc);
+    return err;
+}
+
+#if CONFIG_VAAPI
+static int vulkan_map_to_vaapi(AVHWFramesContext *hwfc, AVFrame *dst,
+                               const AVFrame *src, int flags)
+{
+    int err;
+    AVFrame *tmp = av_frame_alloc();
+    if (!tmp)
+        return AVERROR(ENOMEM);
+
+    tmp->format = AV_PIX_FMT_DRM_PRIME;
+
+    err = vulkan_map_to_drm(hwfc, tmp, src, flags);
+    if (err < 0)
+        goto fail;
+
+    err = av_hwframe_map(dst, tmp, flags);
+    if (err < 0)
+        goto fail;
+
+    err = ff_hwframe_map_replace(dst, src);
+
+fail:
+    av_frame_free(&tmp);
+    return err;
+}
+#endif
+#endif
+
+static int vulkan_map_from(AVHWFramesContext *hwfc, AVFrame *dst,
+                           const AVFrame *src, int flags)
+{
+    av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
+
+    switch (dst->format) {
+#if CONFIG_LIBDRM
+    case AV_PIX_FMT_DRM_PRIME:
+        if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
+            return vulkan_map_to_drm(hwfc, dst, src, flags);
+#if CONFIG_VAAPI
+    case AV_PIX_FMT_VAAPI:
+        if (p->extensions & EXT_EXTERNAL_DMABUF_MEMORY)
+            return vulkan_map_to_vaapi(hwfc, dst, src, flags);
+#endif
+#endif
+    default:
+        return vulkan_map_frame_to_mem(hwfc, dst, src, flags);
+    }
+}
+
+typedef struct ImageBuffer {
+    VkBuffer buf;
+    VkDeviceMemory mem;
+    VkMemoryPropertyFlagBits flags;
+} ImageBuffer;
+
+static void free_buf(AVHWDeviceContext *ctx, ImageBuffer *buf)
+{
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+    if (!buf)
+        return;
+
+    vkDestroyBuffer(hwctx->act_dev, buf->buf, hwctx->alloc);
+    vkFreeMemory(hwctx->act_dev, buf->mem, hwctx->alloc);
+}
+
+static int create_buf(AVHWDeviceContext *ctx, ImageBuffer *buf, int height,
+                      int *stride, VkBufferUsageFlags usage,
+                      VkMemoryPropertyFlagBits flags, void *create_pnext,
+                      void *alloc_pnext)
+{
+    int err;
+    VkResult ret;
+    VkMemoryRequirements req;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+    VulkanDevicePriv *p = ctx->internal->priv;
+
+    VkBufferCreateInfo buf_spawn = {
+        .sType       = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
+        .pNext       = create_pnext,
+        .usage       = usage,
+        .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
+    };
+
+    *stride = FFALIGN(*stride, p->props.limits.optimalBufferCopyRowPitchAlignment);
+    buf_spawn.size = height*(*stride);
+
+    ret = vkCreateBuffer(hwctx->act_dev, &buf_spawn, NULL, &buf->buf);
+    if (ret != VK_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Failed to create buffer: %s\n",
+               vk_ret2str(ret));
+        return AVERROR_EXTERNAL;
+    }
+
+    vkGetBufferMemoryRequirements(hwctx->act_dev, buf->buf, &req);
+
+    err = alloc_mem(ctx, &req, flags, alloc_pnext, &buf->flags, &buf->mem);
+    if (err)
+        return err;
+
+    ret = vkBindBufferMemory(hwctx->act_dev, buf->buf, buf->mem, 0);
+    if (ret != VK_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Failed to bind memory to buffer: %s\n",
+               vk_ret2str(ret));
+        free_buf(ctx, buf);
+        return AVERROR_EXTERNAL;
+    }
+
+    return 0;
+}
+
+static int map_buffers(AVHWDeviceContext *ctx, ImageBuffer *buf, uint8_t *mem[],
+                       int nb_buffers, int invalidate)
+{
+    VkResult ret;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+    VkMappedMemoryRange invalidate_ctx[AV_NUM_DATA_POINTERS];
+    int invalidate_count = 0;
+
+    for (int i = 0; i < nb_buffers; i++) {
+        ret = vkMapMemory(hwctx->act_dev, buf[i].mem, 0,
+                          VK_WHOLE_SIZE, 0, (void **)&mem[i]);
+        if (ret != VK_SUCCESS) {
+            av_log(ctx, AV_LOG_ERROR, "Failed to map buffer memory: %s\n",
+                   vk_ret2str(ret));
+            return AVERROR_EXTERNAL;
+        }
+    }
+
+    if (!invalidate)
+        return 0;
+
+    for (int i = 0; i < nb_buffers; i++) {
+        const VkMappedMemoryRange ival_buf = {
+            .sType  = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
+            .memory = buf[i].mem,
+            .size   = VK_WHOLE_SIZE,
+        };
+        if (buf[i].flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
+            continue;
+        invalidate_ctx[invalidate_count++] = ival_buf;
+    }
+
+    if (invalidate_count) {
+        ret = vkInvalidateMappedMemoryRanges(hwctx->act_dev, invalidate_count,
+                                             invalidate_ctx);
+        if (ret != VK_SUCCESS)
+            av_log(ctx, AV_LOG_WARNING, "Failed to invalidate memory: %s\n",
+                   vk_ret2str(ret));
+    }
+
+    return 0;
+}
+
+static int unmap_buffers(AVHWDeviceContext *ctx, ImageBuffer *buf,
+                         int nb_buffers, int flush)
+{
+    int err = 0;
+    VkResult ret;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+    VkMappedMemoryRange flush_ctx[AV_NUM_DATA_POINTERS];
+    int flush_count = 0;
+
+    if (flush) {
+        for (int i = 0; i < nb_buffers; i++) {
+            const VkMappedMemoryRange flush_buf = {
+                .sType  = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE,
+                .memory = buf[i].mem,
+                .size   = VK_WHOLE_SIZE,
+            };
+            if (buf[i].flags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
+                continue;
+            flush_ctx[flush_count++] = flush_buf;
+        }
+    }
+
+    if (flush_count) {
+        ret = vkFlushMappedMemoryRanges(hwctx->act_dev, flush_count, flush_ctx);
+        if (ret != VK_SUCCESS) {
+            av_log(ctx, AV_LOG_ERROR, "Failed to flush memory: %s\n",
+                    vk_ret2str(ret));
+            err = AVERROR_EXTERNAL; /* We still want to try to unmap them */
+        }
+    }
+
+    for (int i = 0; i < nb_buffers; i++)
+        vkUnmapMemory(hwctx->act_dev, buf[i].mem);
+
+    return err;
+}
+
+static int transfer_image_buf(AVHWDeviceContext *ctx, AVVkFrame *frame,
+                              ImageBuffer *buffer, const int *buf_stride, int w,
+                              int h, enum AVPixelFormat pix_fmt, int to_buf)
+{
+    VkResult ret;
+    AVVulkanDeviceContext *hwctx = ctx->hwctx;
+    VulkanDevicePriv *s = ctx->internal->priv;
+    VkPipelineStageFlagBits sem_wait_dst[AV_NUM_DATA_POINTERS];
+
+    const int planes = av_pix_fmt_count_planes(pix_fmt);
+    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(pix_fmt);
+
+    VkCommandBufferBeginInfo cmd_start = {
+        .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
+        .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
+    };
+
+    VkImageMemoryBarrier img_bar[AV_NUM_DATA_POINTERS] = { 0 };
+
+    VkSubmitInfo s_info = {
+        .sType                = VK_STRUCTURE_TYPE_SUBMIT_INFO,
+        .commandBufferCount   = 1,
+        .pCommandBuffers      = &s->cmd.buf,
+        .pSignalSemaphores    = frame->sem,
+        .pWaitSemaphores      = frame->sem,
+        .pWaitDstStageMask    = sem_wait_dst,
+        .signalSemaphoreCount = planes,
+        .waitSemaphoreCount   = planes,
+    };
+
+    ret = vkBeginCommandBuffer(s->cmd.buf, &cmd_start);
+    if (ret != VK_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Unable to init command buffer: %s\n",
+               vk_ret2str(ret));
+        return AVERROR_EXTERNAL;
+    }
+
+    /* Change the image layout to something more optimal for transfers */
+    for (int i = 0; i < planes; i++) {
+        img_bar[i].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
+        img_bar[i].srcAccessMask = 0x0;
+        img_bar[i].dstAccessMask = to_buf ? VK_ACCESS_TRANSFER_READ_BIT :
+                                            VK_ACCESS_TRANSFER_WRITE_BIT;
+        img_bar[i].oldLayout = frame->layout[i];
+        img_bar[i].newLayout = to_buf ? VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL :
+                                        VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
+        img_bar[i].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+        img_bar[i].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+        img_bar[i].image = frame->img[i];
+        img_bar[i].subresourceRange.levelCount = 1;
+        img_bar[i].subresourceRange.layerCount = 1;
+        img_bar[i].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
+
+        sem_wait_dst[i] = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
+
+        frame->layout[i] = img_bar[i].newLayout;
+        frame->access[i] = img_bar[i].dstAccessMask;
+    }
+
+    vkCmdPipelineBarrier(s->cmd.buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+                         VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
+                         0, NULL, 0, NULL, planes, img_bar);
+
+    /* Schedule a copy for each plane */
+    for (int i = 0; i < planes; i++) {
+        const int p_w = i > 0 ? AV_CEIL_RSHIFT(w, desc->log2_chroma_w) : w;
+        const int p_h = i > 0 ? AV_CEIL_RSHIFT(h, desc->log2_chroma_h) : h;
+        VkBufferImageCopy buf_reg = {
+            .bufferOffset = 0,
+            /* Buffer stride isn't in bytes, it's in samples, the implementation
+             * uses the image's VkFormat to know how many bytes per sample
+             * the buffer has. So we have to convert by dividing. Stupid.
+             * Won't work with YUVA or other planar formats with alpha. */
+            .bufferRowLength = buf_stride[i] / desc->comp[i].step,
+            .bufferImageHeight = p_h,
+            .imageSubresource.layerCount = 1,
+            .imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+            .imageOffset = { 0, 0, 0, },
+            .imageExtent = { p_w, p_h, 1, },
+        };
+
+        if (to_buf)
+            vkCmdCopyImageToBuffer(s->cmd.buf, frame->img[i], frame->layout[i],
+                                   buffer[i].buf, 1, &buf_reg);
+        else
+            vkCmdCopyBufferToImage(s->cmd.buf, buffer[i].buf, frame->img[i],
+                                   frame->layout[i], 1, &buf_reg);
+    }
+
+    ret = vkEndCommandBuffer(s->cmd.buf);
+    if (ret != VK_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Unable to finish command buffer: %s\n",
+               vk_ret2str(ret));
+        return AVERROR_EXTERNAL;
+    }
+
+    /* Wait for the download/upload to finish if uploading, otherwise the
+     * semaphore will take care of synchronization when uploading */
+    ret = vkQueueSubmit(s->cmd.queue, 1, &s_info, s->cmd.fence);
+    if (ret != VK_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "Unable to submit command buffer: %s\n",
+               vk_ret2str(ret));
+        return AVERROR_EXTERNAL;
+    } else {
+        vkWaitForFences(hwctx->act_dev, 1, &s->cmd.fence, VK_TRUE, UINT64_MAX);
+        vkResetFences(hwctx->act_dev, 1, &s->cmd.fence);
+    }
+
+    return 0;
+}
+
+/* Technically we can use VK_EXT_external_memory_host to upload and download,
+ * however the alignment requirements make this unfeasible as both the pointer
+ * and the size of each plane need to be aligned to the minimum alignment
+ * requirement, which on all current implementations (anv, radv) is 4096.
+ * If the requirement gets relaxed (unlikely) this can easily be implemented. */
+static int vulkan_transfer_data_from_mem(AVHWFramesContext *hwfc, AVFrame *dst,
+                                         const AVFrame *src)
+{
+    int err = 0;
+    AVFrame tmp;
+    AVVkFrame *f = (AVVkFrame *)dst->data[0];
+    AVHWDeviceContext *dev_ctx = hwfc->device_ctx;
+    ImageBuffer buf[AV_NUM_DATA_POINTERS] = { { 0 } };
+    const int planes = av_pix_fmt_count_planes(src->format);
+    int log2_chroma = av_pix_fmt_desc_get(src->format)->log2_chroma_h;
+
+    if ((src->format != AV_PIX_FMT_NONE && !av_vkfmt_from_pixfmt(src->format))) {
+        av_log(hwfc, AV_LOG_ERROR, "Unsupported source pixel format!\n");
+        return AVERROR(EINVAL);
+    }
+
+    if (src->width > hwfc->width || src->height > hwfc->height)
+        return AVERROR(EINVAL);
+
+    /* For linear, host visiable images */
+    if (f->tiling == VK_IMAGE_TILING_LINEAR &&
+        f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
+        AVFrame *map = av_frame_alloc();
+        if (!map)
+            return AVERROR(ENOMEM);
+        map->format = src->format;
+
+        err = vulkan_map_frame_to_mem(hwfc, map, dst, AV_HWFRAME_MAP_WRITE);
+        if (err)
+            goto end;
+
+        err = av_frame_copy(map, src);
+        av_frame_free(&map);
+        goto end;
+    }
+
+    /* Create buffers */
+    for (int i = 0; i < planes; i++) {
+        int h = src->height;
+        int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
+
+        tmp.linesize[i] = src->linesize[i];
+        err = create_buf(dev_ctx, &buf[i], p_height,
+                         &tmp.linesize[i], VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
+                         VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL, NULL);
+        if (err)
+            goto end;
+    }
+
+    /* Map, copy image to buffer, unmap */
+    if ((err = map_buffers(dev_ctx, buf, tmp.data, planes, 0)))
+        goto end;
+
+    av_image_copy(tmp.data, tmp.linesize, (const uint8_t **)src->data,
+                  src->linesize, src->format, src->width, src->height);
+
+    if ((err = unmap_buffers(dev_ctx, buf, planes, 1)))
+        goto end;
+
+    /* Copy buffers to image */
+    err = transfer_image_buf(dev_ctx, f, buf, tmp.linesize,
+                             src->width, src->height, src->format, 0);
+
+end:
+    for (int i = 0; i < planes; i++)
+        free_buf(dev_ctx, &buf[i]);
+
+    return err;
+}
+
+static int vulkan_transfer_data_to(AVHWFramesContext *hwfc, AVFrame *dst,
+                                        const AVFrame *src)
+{
+    av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
+
+    switch (src->format) {
+#if CONFIG_CUDA
+    case AV_PIX_FMT_CUDA:
+        if ((p->extensions & EXT_EXTERNAL_FD_MEMORY) &&
+            (p->extensions & EXT_EXTERNAL_FD_SEM))
+            return vulkan_transfer_data_from_cuda(hwfc, dst, src);
+#endif
+    default:
+        if (src->hw_frames_ctx)
+            return AVERROR(ENOSYS);
+        else
+            return vulkan_transfer_data_from_mem(hwfc, dst, src);
+    }
+}
+
+#if CONFIG_CUDA
+static int vulkan_transfer_data_to_cuda(AVHWFramesContext *hwfc, AVFrame *dst,
+                                      const AVFrame *src)
+{
+    int err;
+    VkResult ret;
+    CUcontext dummy;
+    AVVkFrame *dst_f;
+    AVVkFrameInternal *dst_int;
+    const int planes = av_pix_fmt_count_planes(hwfc->sw_format);
+    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hwfc->sw_format);
+
+    AVHWFramesContext *cuda_fc = (AVHWFramesContext*)dst->hw_frames_ctx->data;
+    AVHWDeviceContext *cuda_cu = cuda_fc->device_ctx;
+    AVCUDADeviceContext *cuda_dev = cuda_cu->hwctx;
+    AVCUDADeviceContextInternal *cu_internal = cuda_dev->internal;
+    CudaFunctions *cu = cu_internal->cuda_dl;
+
+    ret = CHECK_CU(cu->cuCtxPushCurrent(cuda_dev->cuda_ctx));
+    if (ret < 0) {
+        err = AVERROR_EXTERNAL;
+        goto fail;
+    }
+
+    dst_f = (AVVkFrame *)src->data[0];
+
+    err = vulkan_export_to_cuda(hwfc, dst->hw_frames_ctx, src);
+    if (err < 0) {
+        goto fail;
+    }
+
+    dst_int = dst_f->internal;
+
+    for (int i = 0; i < planes; i++) {
+        CUDA_MEMCPY2D cpy = {
+            .dstMemoryType = CU_MEMORYTYPE_DEVICE,
+            .dstDevice     = (CUdeviceptr)dst->data[i],
+            .dstPitch      = dst->linesize[i],
+            .dstY          = 0,
+
+            .srcMemoryType = CU_MEMORYTYPE_ARRAY,
+            .srcArray      = dst_int->cu_array[i],
+            .WidthInBytes  = (i > 0 ? AV_CEIL_RSHIFT(hwfc->width, desc->log2_chroma_w)
+                                    : hwfc->width) * desc->comp[i].step,
+            .Height        = i > 0 ? AV_CEIL_RSHIFT(hwfc->height, desc->log2_chroma_h)
+                                   : hwfc->height,
+        };
+
+        ret = CHECK_CU(cu->cuMemcpy2DAsync(&cpy, cuda_dev->stream));
+        if (ret < 0) {
+            err = AVERROR_EXTERNAL;
+            goto fail;
+        }
+    }
+
+    CHECK_CU(cu->cuCtxPopCurrent(&dummy));
+
+    av_log(hwfc, AV_LOG_VERBOSE, "Transfered Vulkan image to CUDA!\n");
+
+    return 0;
+
+fail:
+    CHECK_CU(cu->cuCtxPopCurrent(&dummy));
+    vulkan_free_internal(dst_int);
+    dst_f->internal = NULL;
+    av_buffer_unref(&dst->buf[0]);
+    return err;
+}
+#endif
+
+static int vulkan_transfer_data_to_mem(AVHWFramesContext *hwfc, AVFrame *dst,
+                                       const AVFrame *src)
+{
+    int err = 0;
+    AVFrame tmp;
+    AVVkFrame *f = (AVVkFrame *)src->data[0];
+    AVHWDeviceContext *dev_ctx = hwfc->device_ctx;
+    ImageBuffer buf[AV_NUM_DATA_POINTERS] = { { 0 } };
+    const int planes = av_pix_fmt_count_planes(dst->format);
+    int log2_chroma = av_pix_fmt_desc_get(dst->format)->log2_chroma_h;
+
+    if (dst->width > hwfc->width || dst->height > hwfc->height)
+        return AVERROR(EINVAL);
+
+    /* For linear, host visiable images */
+    if (f->tiling == VK_IMAGE_TILING_LINEAR &&
+        f->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
+        AVFrame *map = av_frame_alloc();
+        if (!map)
+            return AVERROR(ENOMEM);
+        map->format = dst->format;
+
+        err = vulkan_map_frame_to_mem(hwfc, map, src, AV_HWFRAME_MAP_READ);
+        if (err)
+            return err;
+
+        err = av_frame_copy(dst, map);
+        av_frame_free(&map);
+        return err;
+    }
+
+    /* Create buffers */
+    for (int i = 0; i < planes; i++) {
+        int h = dst->height;
+        int p_height = i > 0 ? AV_CEIL_RSHIFT(h, log2_chroma) : h;
+
+        tmp.linesize[i] = dst->linesize[i];
+        err = create_buf(dev_ctx, &buf[i], p_height,
+                         &tmp.linesize[i], VK_BUFFER_USAGE_TRANSFER_DST_BIT,
+                         VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, NULL, NULL);
+    }
+
+    /* Copy image to buffer */
+    if ((err = transfer_image_buf(dev_ctx, f, buf, tmp.linesize,
+                                  dst->width, dst->height, dst->format, 1)))
+        goto end;
+
+    /* Map, copy buffer to frame, unmap */
+    if ((err = map_buffers(dev_ctx, buf, tmp.data, planes, 1)))
+        goto end;
+
+    av_image_copy(dst->data, dst->linesize, (const uint8_t **)tmp.data,
+                  tmp.linesize, dst->format, dst->width, dst->height);
+
+    err = unmap_buffers(dev_ctx, buf, planes, 0);
+
+end:
+    for (int i = 0; i < planes; i++)
+        free_buf(dev_ctx, &buf[i]);
+
+    return err;
+}
+
+static int vulkan_transfer_data_from(AVHWFramesContext *hwfc, AVFrame *dst,
+                                     const AVFrame *src)
+{
+    av_unused VulkanDevicePriv *p = hwfc->device_ctx->internal->priv;
+
+    switch (dst->format) {
+#if CONFIG_CUDA
+    case AV_PIX_FMT_CUDA:
+        if ((p->extensions & EXT_EXTERNAL_FD_MEMORY) &&
+            (p->extensions & EXT_EXTERNAL_FD_SEM))
+            return vulkan_transfer_data_to_cuda(hwfc, dst, src);
+#endif
+    default:
+        if (dst->hw_frames_ctx)
+            return AVERROR(ENOSYS);
+        else
+            return vulkan_transfer_data_to_mem(hwfc, dst, src);
+    }
+}
+
+AVVkFrame *av_vk_frame_alloc(void)
+{
+    return av_mallocz(sizeof(AVVkFrame));
+}
+
+const HWContextType ff_hwcontext_type_vulkan = {
+    .type                   = AV_HWDEVICE_TYPE_VULKAN,
+    .name                   = "Vulkan",
+
+    .device_hwctx_size      = sizeof(AVVulkanDeviceContext),
+    .device_priv_size       = sizeof(VulkanDevicePriv),
+    .frames_hwctx_size      = sizeof(AVVulkanFramesContext),
+    .frames_priv_size       = sizeof(VulkanFramesPriv),
+
+    .device_init            = &vulkan_device_init,
+    .device_create          = &vulkan_device_create,
+    .device_derive          = &vulkan_device_derive,
+
+    .frames_get_constraints = &vulkan_frames_get_constraints,
+    .frames_init            = vulkan_frames_init,
+    .frames_get_buffer      = vulkan_get_buffer,
+    .frames_uninit          = vulkan_frames_uninit,
+
+    .transfer_get_formats   = vulkan_transfer_get_formats,
+    .transfer_data_to       = vulkan_transfer_data_to,
+    .transfer_data_from     = vulkan_transfer_data_from,
+
+    .map_to                 = vulkan_map_to,
+    .map_from               = vulkan_map_from,
+
+    .pix_fmts = (const enum AVPixelFormat []) {
+        AV_PIX_FMT_VULKAN,
+        AV_PIX_FMT_NONE
+    },
+};
diff --git a/libavutil/hwcontext_vulkan.h b/libavutil/hwcontext_vulkan.h
new file mode 100644 (file)
index 0000000..01d6d98
--- /dev/null
@@ -0,0 +1,158 @@
+/*
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#ifndef AVUTIL_HWCONTEXT_VULKAN_H
+#define AVUTIL_HWCONTEXT_VULKAN_H
+
+#include <vulkan/vulkan.h>
+
+/**
+ * @file
+ * API-specific header for AV_HWDEVICE_TYPE_VULKAN.
+ *
+ * For user-allocated pools, AVHWFramesContext.pool must return AVBufferRefs
+ * with the data pointer set to an AVVkFrame.
+ */
+
+/**
+ * Main Vulkan context, allocated as AVHWDeviceContext.hwctx.
+ * All of these can be set before init to change what the context uses
+ */
+typedef struct AVVulkanDeviceContext {
+    /**
+     * Custom memory allocator, else NULL
+     */
+    const VkAllocationCallbacks *alloc;
+    /**
+     * Instance
+     */
+    VkInstance inst;
+    /**
+     * Physical device
+     */
+    VkPhysicalDevice phys_dev;
+    /**
+     * Active device
+     */
+    VkDevice act_dev;
+    /**
+     * Queue family index for graphics
+     * @note av_hwdevice_create() will set all 3 queue indices if unset
+     * If there is no dedicated queue for compute or transfer operations,
+     * they will be set to the graphics queue index which can handle both.
+     */
+    int queue_family_index;
+    /**
+     * Queue family index for transfer ops only
+     */
+    int queue_family_tx_index;
+    /**
+     * Queue family index for compute ops
+     */
+    int queue_family_comp_index;
+} AVVulkanDeviceContext;
+
+/**
+ * Allocated as AVHWFramesContext.hwctx, used to set pool-specific options
+ */
+typedef struct AVVulkanFramesContext {
+    /**
+     * Controls the tiling of output frames.
+     */
+    VkImageTiling tiling;
+    /**
+     * Defines extra usage of output frames. This is bitwise OR'd with the
+     * standard usage flags (SAMPLED, STORAGE, TRANSFER_SRC and TRANSFER_DST).
+     */
+    VkImageUsageFlagBits usage;
+    /**
+     * Extension data for image creation. By default, if the extension is
+     * available, this will be chained to a VkImageFormatListCreateInfoKHR.
+     */
+    void *create_pnext;
+    /**
+     * Extension data for memory allocation. Must have as many entries as
+     * the number of planes of the sw_format.
+     * This will be chained to VkExportMemoryAllocateInfo, which is used
+     * to make all pool images exportable to other APIs.
+     */
+    void *alloc_pnext[AV_NUM_DATA_POINTERS];
+} AVVulkanFramesContext;
+
+/*
+ * Frame structure, the VkFormat of the image will always match
+ * the pool's sw_format.
+ * All frames, imported or allocated, will be created with the
+ * VK_IMAGE_CREATE_ALIAS_BIT flag set, so the memory may be aliased if needed.
+ *
+ * @note the size of this structure is not part of the ABI, to allocate
+ * you must use @av_vk_frame_alloc().
+ */
+typedef struct AVVkFrame {
+    /**
+     * Vulkan images to which the memory is bound to.
+     */
+    VkImage img[AV_NUM_DATA_POINTERS];
+
+    /**
+     * Same tiling must be used for all images.
+     */
+    VkImageTiling tiling;
+
+    /**
+     * Memory backing the images. Could be less than the amount of images
+     * if importing from a DRM or VAAPI frame.
+     */
+    VkDeviceMemory mem[AV_NUM_DATA_POINTERS];
+    size_t size[AV_NUM_DATA_POINTERS];
+
+    /**
+     * OR'd flags for all memory allocated
+     */
+    VkMemoryPropertyFlagBits flags;
+
+    /**
+     * Updated after every barrier
+     */
+    VkAccessFlagBits access[AV_NUM_DATA_POINTERS];
+    VkImageLayout layout[AV_NUM_DATA_POINTERS];
+
+    /**
+     * Per-image semaphores. Must not be freed manually. Must be waited on
+     * and signalled at every queue submission.
+     */
+    VkSemaphore sem[AV_NUM_DATA_POINTERS];
+
+    /**
+     * Internal data.
+     */
+    struct AVVkFrameInternal *internal;
+} AVVkFrame;
+
+/**
+ * Allocates a single AVVkFrame and initializes everything as 0.
+ * @note Must be freed via av_free()
+ */
+AVVkFrame *av_vk_frame_alloc(void);
+
+/**
+ * Returns the format of each image up to the number of planes for a given sw_format.
+ */
+const VkFormat *av_vkfmt_from_pixfmt(enum AVPixelFormat p);
+
+#endif /* AVUTIL_HWCONTEXT_VULKAN_H */
index 05dd4a1..64178b7 100644 (file)
@@ -2344,6 +2344,10 @@ static const AVPixFmtDescriptor av_pix_fmt_descriptors[AV_PIX_FMT_NB] = {
         },
         .flags = AV_PIX_FMT_FLAG_PLANAR,
     },
+    [AV_PIX_FMT_VULKAN] = {
+        .name = "vulkan",
+        .flags = AV_PIX_FMT_FLAG_HWACCEL,
+    },
 };
 #if FF_API_PLUS1_MINUS1
 FF_ENABLE_DEPRECATION_WARNINGS
index 37ecebd..e1a5117 100644 (file)
@@ -348,6 +348,13 @@ enum AVPixelFormat {
     AV_PIX_FMT_NV24,      ///< planar YUV 4:4:4, 24bpp, 1 plane for Y and 1 plane for the UV components, which are interleaved (first byte U and the following byte V)
     AV_PIX_FMT_NV42,      ///< as above, but U and V bytes are swapped
 
+    /**
+     * Vulkan hardware images.
+     *
+     * data[0] points to an AVVkFrame
+     */
+    AV_PIX_FMT_VULKAN,
+
     AV_PIX_FMT_NB         ///< number of pixel formats, DO NOT USE THIS if you want to link with shared libav* because the number of formats might differ between versions
 };