vo_opengl: add dxinterop backend

WGL_NV_DX_interop is widely supported by Nvidia and AMD drivers. It
allows a texture to be shared between Direct3D and WGL, so that
rendering can be done with WGL and presentation can be done with
Direct3D. This should allow us to work around some persistent WGL
issues, such as dropped frames with some driver/OS combos, drivers that
buffer frames to increase performance at the cost of latency, and the
inability to disable exclusive fullscreen mode when using WGL to render
to a fullscreen window.

The addition of a DX_interop backend might also enable some cool
Direct3D-specific enhancements in the future, such as using the
GetPresentStatistics API to get accurate frame presentation timestamps.

Note that due to a driver bug, this backend is currently broken on
Intel. It will appear to work as long as the window is not resized too
often, but after a few changes of size it will be unable to share the
newly created renderbuffer with GL. See:
https://software.intel.com/en-us/forums/graphics-driver-bug-reporting/topic/562051

7 files changed, 689 insertions, 1 deletions

diff --git a/DOCS/man/vo.rst b/DOCS/man/vo.rst
index f4f78c491f..697584ca98 100644
--- a/DOCS/man/vo.rst
+++ b/DOCS/man/vo.rst
@@ -795,6 +795,9 @@ Available video output drivers are:
             Direct3D11 through the OpenGL ES translation layer ANGLE. This
             supports almost everything the ``win`` backend does, except ICC
             profiles, high bit depth video input, and the ``nnedi3`` prescaler.
+        dxinterop (experimental)
+            Win32, using WGL for rendering and Direct3D 9Ex for presentation.
+            Works on Nvidia and AMD only.
         x11
             X11/GLX
         wayland
diff --git a/video/out/opengl/common.c b/video/out/opengl/common.c
index a93ecf9d0b..00cd535301 100644
--- a/video/out/opengl/common.c
+++ b/video/out/opengl/common.c
@@ -289,6 +289,22 @@ static const struct gl_functions gl_functions[] = {
             {0}
         },
     },
+#if HAVE_GL_DXINTEROP
+    {
+        .extension = "WGL_NV_DX_interop",
+        .provides = MPGL_CAP_DXINTEROP,
+        .functions = (const struct gl_function[]) {
+            DEF_FN_NAME(DXSetResourceShareHandleNV, "wglDXSetResourceShareHandleNV"),
+            DEF_FN_NAME(DXOpenDeviceNV, "wglDXOpenDeviceNV"),
+            DEF_FN_NAME(DXCloseDeviceNV, "wglDXCloseDeviceNV"),
+            DEF_FN_NAME(DXRegisterObjectNV, "wglDXRegisterObjectNV"),
+            DEF_FN_NAME(DXUnregisterObjectNV, "wglDXUnregisterObjectNV"),
+            DEF_FN_NAME(DXLockObjectsNV, "wglDXLockObjectsNV"),
+            DEF_FN_NAME(DXUnlockObjectsNV, "wglDXUnlockObjectsNV"),
+            {0}
+        },
+    },
+#endif
     // Apple Packed YUV Formats
     // For gl_hwdec_vda.c
     // http://www.opengl.org/registry/specs/APPLE/rgb_422.txt
@@ -527,6 +543,7 @@ extern const struct mpgl_driver mpgl_driver_cocoa;
 extern const struct mpgl_driver mpgl_driver_wayland;
 extern const struct mpgl_driver mpgl_driver_w32;
 extern const struct mpgl_driver mpgl_driver_angle;
+extern const struct mpgl_driver mpgl_driver_dxinterop;
 extern const struct mpgl_driver mpgl_driver_rpi;
 
 static const struct mpgl_driver *const backends[] = {
@@ -542,6 +559,9 @@ static const struct mpgl_driver *const backends[] = {
 #if HAVE_GL_WIN32
     &mpgl_driver_w32,
 #endif
+#if HAVE_GL_DXINTEROP
+    &mpgl_driver_dxinterop,
+#endif
 #if HAVE_GL_WAYLAND
     &mpgl_driver_wayland,
 #endif
diff --git a/video/out/opengl/common.h b/video/out/opengl/common.h
index 8d5bbdbb6d..537e785e3b 100644
--- a/video/out/opengl/common.h
+++ b/video/out/opengl/common.h
@@ -63,6 +63,7 @@ enum {
     MPGL_CAP_1D_TEX             = (1 << 14),
     MPGL_CAP_3D_TEX             = (1 << 15),
     MPGL_CAP_DEBUG              = (1 << 16),
+    MPGL_CAP_DXINTEROP          = (1 << 17),    // WGL_NV_DX_interop
     MPGL_CAP_SW                 = (1 << 30),    // indirect or sw renderer
 };
 
@@ -128,6 +129,9 @@ typedef struct MPGLContext {
     // Windows-specific hack. See vo_opengl dwmflush suboption.
     int dwm_flush_opt;
 
+    // Flip the rendered image vertically. This is useful for dxinterop.
+    bool flip_v;
+
     // For free use by the mpgl_driver.
     void *priv;
 } MPGLContext;
@@ -261,6 +265,21 @@ struct GL {
     void (GLAPIENTRY *VDPAUMapSurfacesNV)(GLsizei, const GLvdpauSurfaceNV *);
     void (GLAPIENTRY *VDPAUUnmapSurfacesNV)(GLsizei, const GLvdpauSurfaceNV *);
 
+#if HAVE_GL_WIN32
+    // The HANDLE type might not be present on non-Win32
+    BOOL (GLAPIENTRY *DXSetResourceShareHandleNV)(void *dxObject,
+        HANDLE shareHandle);
+    HANDLE (GLAPIENTRY *DXOpenDeviceNV)(void *dxDevice);
+    BOOL (GLAPIENTRY *DXCloseDeviceNV)(HANDLE hDevice);
+    HANDLE (GLAPIENTRY *DXRegisterObjectNV)(HANDLE hDevice, void *dxObject,
+        GLuint name, GLenum type, GLenum access);
+    BOOL (GLAPIENTRY *DXUnregisterObjectNV)(HANDLE hDevice, HANDLE hObject);
+    BOOL (GLAPIENTRY *DXLockObjectsNV)(HANDLE hDevice, GLint count,
+        HANDLE *hObjects);
+    BOOL (GLAPIENTRY *DXUnlockObjectsNV)(HANDLE hDevice, GLint count,
+        HANDLE *hObjects);
+#endif
+
     GLint (GLAPIENTRY *GetVideoSync)(GLuint *);
     GLint (GLAPIENTRY *WaitVideoSync)(GLint, GLint, unsigned int *);
 
diff --git a/video/out/opengl/dxinterop.c b/video/out/opengl/dxinterop.c
new file mode 100644
index 0000000000..04518714d1
--- /dev/null
+++ b/video/out/opengl/dxinterop.c
@@ -0,0 +1,636 @@
+/*
+ * This file is part of mpv.
+ *
+ * mpv is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * mpv 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with mpv.  If not, see <http://www.gnu.org/licenses/>.
+ *
+ * You can alternatively redistribute this file 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.
+ */
+
+#include <windows.h>
+#include <initguid.h>
+#include <d3d9.h>
+#include <dwmapi.h>
+#include "video/out/w32_common.h"
+#include "common.h"
+
+// For WGL_ACCESS_WRITE_DISCARD_NV, etc.
+#include <GL/wglext.h>
+
+// mingw-w64 header typo?
+#ifndef IDirect3DSwapChain9Ex_GetBackBuffer
+#define IDirect3DSwapChain9Ex_GetBackBuffer IDirect3DSwapChain9EX_GetBackBuffer
+#endif
+
+struct priv {
+    HMODULE d3d9_dll;
+    HRESULT (WINAPI *Direct3DCreate9Ex)(UINT SDKVersion, IDirect3D9Ex **ppD3D);
+
+    // Direct3D9 device and resources
+    IDirect3D9Ex *d3d9ex;
+    IDirect3DDevice9Ex *device;
+    HANDLE device_h;
+    IDirect3DSwapChain9Ex *swapchain;
+    IDirect3DSurface9 *backbuffer;
+    IDirect3DSurface9 *rtarget;
+    HANDLE rtarget_h;
+
+    // OpenGL offscreen context
+    HWND os_wnd;
+    HDC os_dc;
+    HGLRC os_ctx;
+
+    // OpenGL resources
+    GLuint framebuffer;
+    GLuint texture;
+
+    // Is the shared framebuffer currently bound?
+    bool fb_bound;
+    // Is the shared texture currently attached?
+    bool tex_attached;
+    // Did we lose the device?
+    bool lost_device;
+
+    // Requested and current parameters
+    int requested_swapinterval;
+    int width, height, swapinterval;
+
+    void (GLAPIENTRY *real_gl_bind_framebuffer)(GLenum, GLuint);
+};
+
+static __thread struct MPGLContext *current_ctx;
+
+static void pump_message_loop(void)
+{
+    // We have a hidden window on this thread (for the OpenGL context,) so pump
+    // its message loop at regular intervals to be safe
+    MSG message;
+    while (PeekMessageW(&message, NULL, 0, 0, PM_REMOVE))
+        DispatchMessageW(&message);
+}
+
+static void *w32gpa(const GLubyte *procName)
+{
+    HMODULE oglmod;
+    void *res = wglGetProcAddress(procName);
+    if (res)
+        return res;
+    oglmod = GetModuleHandleW(L"opengl32.dll");
+    return GetProcAddress(oglmod, procName);
+}
+
+static int os_ctx_create(struct MPGLContext *ctx)
+{
+    static const wchar_t os_wnd_class[] = L"mpv offscreen gl";
+    struct priv *p = ctx->priv;
+    HGLRC legacy_context = NULL;
+
+    RegisterClassExW(&(WNDCLASSEXW) {
+        .cbSize = sizeof(WNDCLASSEXW),
+        .style = CS_OWNDC,
+        .lpfnWndProc = DefWindowProc,
+        .hInstance = GetModuleHandleW(NULL),
+        .lpszClassName = os_wnd_class,
+    });
+
+    // Create a hidden window for an offscreen OpenGL context. It might also be
+    // possible to use the VO window, but MSDN recommends against drawing to
+    // the same window with flip mode present and other APIs, so play it safe.
+    p->os_wnd = CreateWindowExW(0, os_wnd_class, os_wnd_class, 0, 0, 0, 200,
+        200, NULL, NULL, GetModuleHandleW(NULL), NULL);
+    p->os_dc = GetDC(p->os_wnd);
+    if (!p->os_dc) {
+        MP_FATAL(ctx->vo, "Couldn't create window for offscreen rendering\n");
+        goto fail;
+    }
+
+    // Choose a pixel format. It probably doesn't matter what this is because
+    // the primary framebuffer will not be used.
+    PIXELFORMATDESCRIPTOR pfd = {
+        .nSize = sizeof pfd,
+        .nVersion = 1,
+        .dwFlags = PFD_DRAW_TO_WINDOW | PFD_SUPPORT_OPENGL | PFD_DOUBLEBUFFER,
+        .iPixelType = PFD_TYPE_RGBA,
+        .cColorBits = 24,
+        .iLayerType = PFD_MAIN_PLANE,
+    };
+    int pf = ChoosePixelFormat(p->os_dc, &pfd);
+    if (!pf) {
+        MP_FATAL(ctx->vo, "Couldn't choose pixelformat for offscreen rendering\n");
+        goto fail;
+    }
+    SetPixelFormat(p->os_dc, pf, &pfd);
+
+    legacy_context = wglCreateContext(p->os_dc);
+    if (!legacy_context || !wglMakeCurrent(p->os_dc, legacy_context)) {
+        MP_FATAL(ctx->vo, "Couldn't create GL context for offscreen rendering\n");
+        goto fail;
+    }
+
+    const char *(GLAPIENTRY *wglGetExtensionsStringARB)(HDC hdc)
+        = w32gpa((const GLubyte*)"wglGetExtensionsStringARB");
+    if (!wglGetExtensionsStringARB) {
+        MP_FATAL(ctx->vo, "The OpenGL driver does not support OpenGL 3.x\n");
+        goto fail;
+    }
+
+    const char *wgl_exts = wglGetExtensionsStringARB(p->os_dc);
+    if (!strstr(wgl_exts, "WGL_ARB_create_context")) {
+        MP_FATAL(ctx->vo, "The OpenGL driver does not support OpenGL 3.x\n");
+        goto fail;
+    }
+
+    HGLRC (GLAPIENTRY *wglCreateContextAttribsARB)(HDC hDC, HGLRC hShareContext,
+                                                   const int *attribList)
+        = w32gpa((const GLubyte*)"wglCreateContextAttribsARB");
+    if (!wglCreateContextAttribsARB) {
+        MP_FATAL(ctx->vo, "The OpenGL driver does not support OpenGL 3.x\n");
+        goto fail;
+    }
+
+    int attribs[] = {
+        WGL_CONTEXT_MAJOR_VERSION_ARB, 3,
+        WGL_CONTEXT_MINOR_VERSION_ARB, 0,
+        WGL_CONTEXT_FLAGS_ARB, 0,
+        WGL_CONTEXT_PROFILE_MASK_ARB, WGL_CONTEXT_CORE_PROFILE_BIT_ARB,
+        0
+    };
+
+    p->os_ctx = wglCreateContextAttribsARB(p->os_dc, 0, attribs);
+    if (!p->os_ctx) {
+        // NVidia, instead of ignoring WGL_CONTEXT_FLAGS_ARB, will error out if
+        // it's present on pre-3.2 contexts.
+        // Remove it from attribs and retry the context creation.
+        attribs[6] = attribs[7] = 0;
+        p->os_ctx = wglCreateContextAttribsARB(p->os_dc, 0, attribs);
+    }
+    if (!p->os_ctx) {
+        MP_FATAL(ctx->vo, "Couldn't create GL 3.x context for offscreen rendering\n");
+        goto fail;
+    }
+
+    wglMakeCurrent(p->os_dc, NULL);
+    wglDeleteContext(legacy_context);
+    legacy_context = NULL;
+
+    if (!wglMakeCurrent(p->os_dc, p->os_ctx)) {
+        MP_FATAL(ctx->vo, "Couldn't create GL 3.x context for offscreen rendering\n");
+        goto fail;
+    }
+
+    mpgl_load_functions(ctx->gl, w32gpa, wgl_exts, ctx->vo->log);
+    if (!(ctx->gl->mpgl_caps & MPGL_CAP_DXINTEROP)) {
+        MP_FATAL(ctx->vo, "WGL_NV_DX_interop is not supported\n");
+        goto fail;
+    }
+
+    return 0;
+fail:
+    if (legacy_context) {
+        wglMakeCurrent(p->os_dc, NULL);
+        wglDeleteContext(legacy_context);
+    }
+    return -1;
+}
+
+static void os_ctx_destroy(MPGLContext *ctx)
+{
+    struct priv *p = ctx->priv;
+
+    if (p->os_ctx) {
+        wglMakeCurrent(p->os_dc, NULL);
+        wglDeleteContext(p->os_ctx);
+    }
+    if (p->os_dc)
+        ReleaseDC(p->os_wnd, p->os_dc);
+    if (p->os_wnd)
+        DestroyWindow(p->os_wnd);
+}
+
+static void try_attach_texture(MPGLContext *ctx)
+{
+    struct priv *p = ctx->priv;
+    struct GL *gl = ctx->gl;
+
+    if (p->fb_bound && !p->tex_attached) {
+        gl->FramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0,
+            GL_TEXTURE_2D, p->texture, 0);
+        p->tex_attached = true;
+    }
+}
+
+static int d3d_size_dependent_create(MPGLContext *ctx)
+{
+    struct priv *p = ctx->priv;
+    struct GL *gl = ctx->gl;
+    HRESULT hr;
+
+    IDirect3DSwapChain9 *sw9;
+    hr = IDirect3DDevice9Ex_GetSwapChain(p->device, 0, &sw9);
+    if (FAILED(hr)) {
+        MP_FATAL(ctx->vo, "Couldn't get swap chain\n");
+        return -1;
+    }
+
+    hr = IDirect3DSwapChain9_QueryInterface(sw9, &IID_IDirect3DSwapChain9Ex,
+        (void**)&p->swapchain);
+    if (FAILED(hr)) {
+        IDirect3DSwapChain9_Release(sw9);
+        MP_FATAL(ctx->vo, "Couldn't get swap chain\n");
+        return -1;
+    }
+    IDirect3DSwapChain9_Release(sw9);
+
+    hr = IDirect3DSwapChain9Ex_GetBackBuffer(p->swapchain, 0,
+        D3DBACKBUFFER_TYPE_MONO, &p->backbuffer);
+    if (FAILED(hr)) {
+        MP_FATAL(ctx->vo, "Couldn't get backbuffer\n");
+        return -1;
+    }
+
+    // Get the format of the backbuffer
+    D3DSURFACE_DESC bb_desc = { 0 };
+    IDirect3DSurface9_GetDesc(p->backbuffer, &bb_desc);
+
+    MP_VERBOSE(ctx->vo, "DX_interop backbuffer size: %ux%u\n",
+        (unsigned)bb_desc.Width, (unsigned)bb_desc.Height);
+    MP_VERBOSE(ctx->vo, "DX_interop backbuffer format: %u\n",
+        (unsigned)bb_desc.Format);
+
+    // Note: This backend has only been tested on an 8-bit display. It's
+    // unknown whether this code is enough to support other formats or if more
+    // work is needed.
+    switch (bb_desc.Format) {
+    case D3DFMT_X1R5G5B5: case D3DFMT_A1R5G5B5:
+        ctx->depth_r = ctx->depth_g = ctx->depth_b = 5;
+        break;
+    case D3DFMT_R5G6B5:
+        ctx->depth_r = 5; ctx->depth_g = 6; ctx->depth_b = 5;
+        break;
+    case D3DFMT_R8G8B8: case D3DFMT_A8R8G8B8: case D3DFMT_X8R8G8B8:
+    case D3DFMT_A8B8G8R8: case D3DFMT_X8B8G8R8: default:
+        ctx->depth_r = ctx->depth_g = ctx->depth_b = 8;
+        break;
+    case D3DFMT_A2R10G10B10: case D3DFMT_A2B10G10R10:
+        ctx->depth_r = ctx->depth_g = ctx->depth_b = 10;
+        break;
+    }
+
+    // Create a rendertarget with the same format as the backbuffer for
+    // rendering from OpenGL
+    HANDLE share_handle = NULL;
+    hr = IDirect3DDevice9Ex_CreateRenderTarget(p->device, bb_desc.Width,
+        bb_desc.Height, bb_desc.Format, D3DMULTISAMPLE_NONE, 0, FALSE,
+        &p->rtarget, &share_handle);
+    if (FAILED(hr)) {
+        MP_FATAL(ctx->vo, "Couldn't create rendertarget\n");
+        return -1;
+    }
+
+    // Register the share handle with WGL_NV_DX_interop. Nvidia does not
+    // require the use of share handles, but Intel does.
+    if (share_handle)
+        gl->DXSetResourceShareHandleNV(p->rtarget, share_handle);
+
+    // Create the OpenGL-side texture
+    gl->GenTextures(1, &p->texture);
+    p->tex_attached = false;
+
+    // Now share the rendertarget with OpenGL as a texture
+    p->rtarget_h = gl->DXRegisterObjectNV(p->device_h, p->rtarget, p->texture,
+        GL_TEXTURE_2D, WGL_ACCESS_WRITE_DISCARD_NV);
+    if (!p->rtarget_h) {
+        MP_FATAL(ctx->vo, "Couldn't share rendertarget with GL: 0x%08x\n",
+            (unsigned)GetLastError());
+        return -1;
+    }
+
+    // Lock the rendertarget for use from OpenGL. This will only be unlocked in
+    // swap_buffers() when it is blitted to the real Direct3D backbuffer.
+    if (!gl->DXLockObjectsNV(p->device_h, 1, &p->rtarget_h)) {
+        MP_FATAL(ctx->vo, "Couldn't lock rendertarget\n");
+        return -1;
+    }
+
+    // Only attach the shared texture if the shared framebuffer is bound. If
+    // it's not, the texture will be attached when glBindFramebuffer is called.
+    try_attach_texture(ctx);
+
+    return 0;
+}
+
+static void d3d_size_dependent_destroy(MPGLContext *ctx)
+{
+    struct priv *p = ctx->priv;
+    struct GL *gl = ctx->gl;
+
+    if (p->rtarget_h) {
+        gl->DXUnlockObjectsNV(p->device_h, 1, &p->rtarget_h);
+        gl->DXUnregisterObjectNV(p->device_h, p->rtarget_h);
+    }
+    if (p->texture)
+        gl->DeleteTextures(1, &p->texture);
+    if (p->rtarget)
+        IDirect3DSurface9_Release(p->rtarget);
+    if (p->backbuffer)
+        IDirect3DSurface9_Release(p->backbuffer);
+    if (p->swapchain)
+        IDirect3DSwapChain9Ex_Release(p->swapchain);
+}
+
+static void fill_presentparams(MPGLContext *ctx, D3DPRESENT_PARAMETERS *pparams)
+{
+    struct priv *p = ctx->priv;
+
+    // Present intervals other than IMMEDIATE and ONE don't seem to work. It's
+    // possible that they're not compatible with FLIPEX.
+    UINT presentation_interval;
+    switch (p->requested_swapinterval) {
+    case 0:  presentation_interval = D3DPRESENT_INTERVAL_IMMEDIATE; break;
+    case 1:  presentation_interval = D3DPRESENT_INTERVAL_ONE;       break;
+    default: presentation_interval = D3DPRESENT_INTERVAL_ONE;       break;
+    }
+
+    *pparams = (D3DPRESENT_PARAMETERS) {
+        .Windowed = TRUE,
+        // The length of the backbuffer queue shouldn't affect latency because
+        // swap_buffers() always uses the backbuffer at the head of the queue
+        // and presents it immediately. MSDN says there is a performance
+        // penalty for having a short backbuffer queue and this seems to be
+        // true, at least on Nvidia, where less than four backbuffers causes
+        // very high CPU usage. Use six to be safe.
+        .BackBufferCount = 6,
+        .SwapEffect = D3DSWAPEFFECT_FLIPEX,
+        // Automatically get the backbuffer format from the display format. The
+        // size of the backbuffer is automatically determined too.
+        .BackBufferFormat = D3DFMT_UNKNOWN,
+        .PresentationInterval = presentation_interval,
+        .hDeviceWindow = vo_w32_hwnd(ctx->vo),
+    };
+}
+
+static int d3d_create(MPGLContext *ctx)
+{
+    struct priv *p = ctx->priv;
+    struct GL *gl = ctx->gl;
+    HRESULT hr;
+
+    p->d3d9_dll = LoadLibraryW(L"d3d9.dll");
+    if (!p->d3d9_dll) {
+        MP_FATAL(ctx->vo, "\"d3d9.dll\" not found\n");
+        return -1;
+    }
+
+    // WGL_NV_dx_interop requires Direct3D 9Ex on WDDM systems. Direct3D 9Ex
+    // also enables flip mode present for efficient rendering with the DWM.
+    p->Direct3DCreate9Ex = (void*)GetProcAddress(p->d3d9_dll,
+        "Direct3DCreate9Ex");
+    if (!p->Direct3DCreate9Ex) {
+        MP_FATAL(ctx->vo, "Direct3D 9Ex not supported\n");
+        return -1;
+    }
+
+    hr = p->Direct3DCreate9Ex(D3D_SDK_VERSION, &p->d3d9ex);
+    if (FAILED(hr)) {
+        MP_FATAL(ctx->vo, "Couldn't create Direct3D9Ex\n");
+        return -1;
+    }
+
+    D3DPRESENT_PARAMETERS pparams;
+    fill_presentparams(ctx, &pparams);
+
+    hr = IDirect3D9Ex_CreateDeviceEx(p->d3d9ex, D3DADAPTER_DEFAULT,
+        D3DDEVTYPE_HAL, vo_w32_hwnd(ctx->vo),
+        D3DCREATE_HARDWARE_VERTEXPROCESSING | D3DCREATE_PUREDEVICE |
+        D3DCREATE_FPU_PRESERVE | D3DCREATE_MULTITHREADED |
+        D3DCREATE_NOWINDOWCHANGES,
+        &pparams, NULL, &p->device);
+    if (FAILED(hr)) {
+        MP_FATAL(ctx->vo, "Couldn't create device\n");
+        return -1;
+    }
+
+    // mpv expects frames to be presented right after swap_buffers() returns
+    IDirect3DDevice9Ex_SetMaximumFrameLatency(p->device, 1);
+
+    // Register the Direct3D device with WGL_NV_dx_interop
+    p->device_h = gl->DXOpenDeviceNV(p->device);
+    if (!p->device_h) {
+        MP_FATAL(ctx->vo, "Couldn't open Direct3D from GL\n");
+        return -1;
+    }
+
+    return 0;
+}
+
+static void d3d_destroy(MPGLContext *ctx)
+{
+    struct priv *p = ctx->priv;
+    struct GL *gl = ctx->gl;
+
+    if (p->device_h)
+        gl->DXCloseDeviceNV(p->device_h);
+    if (p->device)
+        IDirect3DDevice9Ex_Release(p->device);
+    if (p->d3d9ex)
+        IDirect3D9Ex_Release(p->d3d9ex);
+    if (p->d3d9_dll)
+        FreeLibrary(p->d3d9_dll);
+}
+
+static void dxinterop_uninit(MPGLContext *ctx)
+{
+    d3d_size_dependent_destroy(ctx);
+    d3d_destroy(ctx);
+    os_ctx_destroy(ctx);
+    vo_w32_uninit(ctx->vo);
+    DwmEnableMMCSS(FALSE);
+    pump_message_loop();
+}
+
+static GLAPIENTRY void dxinterop_bind_framebuffer(GLenum target,
+    GLuint framebuffer)
+{
+    if (!current_ctx)
+        return;
+    struct priv *p = current_ctx->priv;
+
+    // Keep track of whether the shared framebuffer is bound
+    if (target == GL_FRAMEBUFFER || target == GL_DRAW_FRAMEBUFFER)
+        p->fb_bound = (framebuffer == 0);
+
+    // Pretend the shared framebuffer is the primary framebuffer
+    if (framebuffer == 0)
+        framebuffer = p->framebuffer;
+
+    p->real_gl_bind_framebuffer(target, framebuffer);
+
+    // Attach the shared texture if it is not attached already
+    try_attach_texture(current_ctx);
+}
+
+static void dxinterop_reset(struct MPGLContext *ctx)
+{
+    struct priv *p = ctx->priv;
+    HRESULT hr;
+
+    // Check if the device actually needs to be reset
+    if (ctx->vo->dwidth == p->width && ctx->vo->dheight == p->height &&
+        p->requested_swapinterval == p->swapinterval && !p->lost_device)
+        return;
+
+    d3d_size_dependent_destroy(ctx);
+
+    D3DPRESENT_PARAMETERS pparams;
+    fill_presentparams(ctx, &pparams);
+
+    hr = IDirect3DDevice9Ex_ResetEx(p->device, &pparams, NULL);
+    if (FAILED(hr)) {
+        MP_FATAL(ctx->vo, "Couldn't reset device\n");
+        return;
+    }
+
+    if (d3d_size_dependent_create(ctx) < 0) {
+        MP_FATAL(ctx->vo, "Couldn't recreate Direct3D objects after reset\n");
+        return;
+    }
+
+    MP_VERBOSE(ctx->vo, "Direct3D device reset\n");
+    p->width = ctx->vo->dwidth;
+    p->height = ctx->vo->dheight;
+    p->swapinterval = p->requested_swapinterval;
+    p->lost_device = false;
+}
+
+static int GLAPIENTRY dxinterop_swap_interval(int interval)
+{
+    if (!current_ctx)
+        return 0;
+    struct priv *p = current_ctx->priv;
+
+    p->requested_swapinterval = interval;
+    dxinterop_reset(current_ctx);
+    return 1;
+}
+
+static int dxinterop_init(struct MPGLContext *ctx, int flags)
+{
+    struct priv *p = ctx->priv;
+    struct GL *gl = ctx->gl;
+
+    p->requested_swapinterval = 1;
+
+    if (!vo_w32_init(ctx->vo))
+        goto fail;
+    if (os_ctx_create(ctx) < 0)
+        goto fail;
+
+    // Create the shared framebuffer
+    gl->GenFramebuffers(1, &p->framebuffer);
+
+    // Hook glBindFramebuffer to return the shared framebuffer instead of the
+    // primary one
+    current_ctx = ctx;
+    p->real_gl_bind_framebuffer = gl->BindFramebuffer;
+    gl->BindFramebuffer = dxinterop_bind_framebuffer;
+
+    gl->SwapInterval = dxinterop_swap_interval;
+
+    if (d3d_create(ctx) < 0)
+        goto fail;
+    if (d3d_size_dependent_create(ctx) < 0)
+        goto fail;
+
+    // Bind the shared framebuffer. This will also attach the shared texture.
+    gl->BindFramebuffer(GL_FRAMEBUFFER, 0);
+
+    // The OpenGL and Direct3D coordinate systems are flipped vertically
+    // relative to each other. Flip the video during rendering so it can be
+    // copied to the Direct3D backbuffer with a simple (and fast) StretchRect.
+    ctx->flip_v = true;
+
+    DwmEnableMMCSS(TRUE);
+
+    return 0;
+fail:
+    dxinterop_uninit(ctx);
+    return -1;
+}
+
+static int dxinterop_reconfig(struct MPGLContext *ctx)
+{
+    vo_w32_config(ctx->vo);
+    return 0;
+}
+
+static void dxinterop_swap_buffers(MPGLContext *ctx)
+{
+    struct priv *p = ctx->priv;
+    struct GL *gl = ctx->gl;
+    HRESULT hr;
+
+    pump_message_loop();
+
+    if (!gl->DXUnlockObjectsNV(p->device_h, 1, &p->rtarget_h)) {
+        MP_FATAL(ctx->vo, "Couldn't unlock rendertarget for present\n");
+        return;
+    }
+
+    // Blit the OpenGL rendertarget to the backbuffer
+    hr = IDirect3DDevice9Ex_StretchRect(p->device, p->rtarget, NULL,
+                                        p->backbuffer, NULL, D3DTEXF_NONE);
+    if (FAILED(hr)) {
+        MP_FATAL(ctx->vo, "Couldn't stretchrect for present\n");
+        return;
+    }
+
+    if (!gl->DXLockObjectsNV(p->device_h, 1, &p->rtarget_h)) {
+        MP_FATAL(ctx->vo, "Couldn't lock rendertarget after stretchrect\n");
+        return;
+    }
+
+    hr = IDirect3DDevice9Ex_PresentEx(p->device, NULL, NULL, NULL, NULL, 0);
+    switch (hr) {
+    case D3DERR_DEVICELOST:
+    case D3DERR_DEVICEHUNG:
+        MP_VERBOSE(ctx->vo, "Direct3D device lost! Resetting.\n");
+        p->lost_device = true;
+        dxinterop_reset(ctx);
+        break;
+    default:
+        if (FAILED(hr))
+            MP_FATAL(ctx->vo, "Couldn't present! 0x%08x\n", (unsigned)hr);
+    }
+}
+
+static int dxinterop_control(MPGLContext *ctx, int *events, int request,
+                             void *arg)
+{
+    int r = vo_w32_control(ctx->vo, events, request, arg);
+    if (*events & VO_EVENT_RESIZE)
+        dxinterop_reset(ctx);
+    return r;
+}
+
+const struct mpgl_driver mpgl_driver_dxinterop = {
+    .name         = "dxinterop",
+    .priv_size    = sizeof(struct priv),
+    .init         = dxinterop_init,
+    .reconfig     = dxinterop_reconfig,
+    .swap_buffers = dxinterop_swap_buffers,
+    .control      = dxinterop_control,
+    .uninit       = dxinterop_uninit,
+};
diff --git a/video/out/vo_opengl.c b/video/out/vo_opengl.c
index eb057d139f..ef6fe53e43 100644
--- a/video/out/vo_opengl.c
+++ b/video/out/vo_opengl.c
@@ -102,7 +102,8 @@ static void resize(struct gl_priv *p)
     struct mp_osd_res osd;
     vo_get_src_dst_rects(vo, &src, &dst, &osd);
 
-    gl_video_resize(p->renderer, vo->dwidth, -vo->dheight, &src, &dst, &osd);
+    int height = p->glctx->flip_v ? vo->dheight : -vo->dheight;
+    gl_video_resize(p->renderer, vo->dwidth, height, &src, &dst, &osd);
 
     vo->want_redraw = true;
 }
diff --git a/wscript b/wscript
index afc0540fee..a5e6914133 100644
--- a/wscript
+++ b/wscript
@@ -663,6 +663,14 @@ video_output_features = [
         'func': check_statement('windows.h', 'wglCreateContext(0)',
                                 lib='opengl32')
     } , {
+        'name': '--gl-dxinterop',
+        'desc': 'OpenGL/DirectX Interop Backend',
+        'deps': [ 'gl-win32' ],
+        'groups': [ 'gl' ],
+        'func': compose_checks(
+            check_statement(['GL/gl.h', 'GL/wglext.h'], 'int i = WGL_ACCESS_WRITE_DISCARD_NV'),
+            check_statement('d3d9.h', 'IDirect3D9Ex *d'))
+    } , {
         'name': '--egl-angle',
         'desc': 'OpenGL Win32 ANGLE Backend',
         'deps_any': [ 'os-win32', 'os-cygwin' ],
diff --git a/wscript_build.py b/wscript_build.py
index 3cd0d9fad0..12199f1e86 100644
--- a/wscript_build.py
+++ b/wscript_build.py
@@ -333,6 +333,7 @@ def build(ctx):
         ( "video/out/opengl/video_shaders.c",    "gl" ),
         ( "video/out/opengl/w32.c",              "gl-win32" ),
         ( "video/out/opengl/angle.c",            "egl-angle" ),
+        ( "video/out/opengl/dxinterop.c",        "gl-dxinterop" ),
         ( "video/out/opengl/wayland.c",          "gl-wayland" ),
         ( "video/out/opengl/x11.c",              "gl-x11" ),
         ( "video/out/opengl/x11egl.c",           "egl-x11" ),