vo_gpu: make it possible to load multiple hwdec interop drivers

Make the VO<->decoder interface capable of supporting multiple hwdec
APIs at once. The main gain is that this simplifies autoprobing a lot.
Before this change, it could happen that the VO loaded the "wrong" hwdec
API, and the decoder was stuck with the choice (breaking hw decoding).
With the change applied, the VO simply loads all available APIs, so
autoprobing trickery is left entirely to the decoder.

In the past, we were quite careful about not accidentally loading the
wrong interop drivers. This was in part to make sure autoprobing works,
but also because libva had this obnoxious bug of dumping garbage to
stderr when using the API. libva was fixed, so this is not a problem
anymore.

The --opengl-hwdec-interop option is changed in various ways (again...),
and renamed to --gpu-hwdec-interop. It does not have much use anymore,
other than debugging. It's notable that the order in the hwdec interop
array ra_hwdec_drivers[] still matters if multiple drivers support the
same image formats, so the option can explicitly force one, if that
should ever be necessary, or more likely, for debugging. One example are
the ra_hwdec_d3d11egl and ra_hwdec_d3d11eglrgb drivers, which both
support d3d11 input.

vo_gpu now always loads the interop lazily by default, but when it does,
it loads them all. vo_opengl_cb now always loads them when the GL
context handle is initialized. I don't expect that this causes any
problems.

It's now possible to do things like changing between vdpau and nvdec
decoding at runtime.

This is also preparation for cleaning up vd_lavc.c hwdec autoprobing.
It's another reason why hwdec_devices_request_all() does not take a
hwdec type anymore.

16 files changed, 207 insertions, 197 deletions

diff --git a/DOCS/interface-changes.rst b/DOCS/interface-changes.rst
index 494b769b8b..312fe9552a 100644
--- a/DOCS/interface-changes.rst
+++ b/DOCS/interface-changes.rst
@@ -48,6 +48,15 @@ Interface changes
     - remove deprecated --demuxer-max-packets
     - remove most of the deprecated audio and video filters
     - remove the deprecated --balance option/property
+    - rename the --opengl-hwdec-interop option to --gpu-hwdec-interop, and
+      change some of its semantics: extend it take the strings "auto" and
+      "all". "all" loads all backends. "auto" behaves like "all" for
+      vo_opengl_cb, while on vo_gpu it loads nothing, but allows on demand
+      loading by the decoder. The empty string as option value behaves like
+      "auto". Old --hwdec values do not work anymore.
+      This option is hereby declared as unstable and may change any time - its
+      old use is deprecated, and it has very little use outside of debugging
+      now.
  --- mpv 0.27.0 ---
     - drop previously deprecated --field-dominance option
     - drop previously deprecated "osd" command
diff --git a/DOCS/man/input.rst b/DOCS/man/input.rst
index 75cb9526c6..4e81a3ce2d 100644
--- a/DOCS/man/input.rst
+++ b/DOCS/man/input.rst
@@ -1433,7 +1433,7 @@ Property list
     This is known only once the VO has opened (and possibly later). With some
     VOs (like ``opengl``), this might be never known in advance, but only when
     the decoder attempted to create the hw decoder successfully. (Using
-    ``--opengl-hwdec-interop`` can load it eagerly.) If there are multiple
+    ``--gpu-hwdec-interop`` can load it eagerly.) If there are multiple
     drivers loaded, they will be separated by ``,``.
 
     If no VO is active or no interop driver is known, this property is
@@ -1443,9 +1443,6 @@ Property list
     multiple interop drivers for the same hardware decoder, depending on
     platform and VO.
 
-    This is somewhat similar to the ``--opengl-hwdec-interop`` option, but
-    it returns the actually loaded backend, not the value of this option.
-
 ``video-format``
     Video format as string.
 
diff --git a/DOCS/man/options.rst b/DOCS/man/options.rst
index f3564cffb6..f01d40e563 100644
--- a/DOCS/man/options.rst
+++ b/DOCS/man/options.rst
@@ -816,34 +816,33 @@ Video
         frame glitches or discoloration, and you have ``--hwdec`` turned on,
         the first thing you should try is disabling it.
 
-``--opengl-hwdec-interop=<name>``
-    This is useful for the ``gpu`` and ``opengl-cb`` VOs for creating the
-    hardware decoding OpenGL interop context, but without actually enabling
-    hardware decoding itself (like ``--hwdec`` does).
-
-    If set to an empty string (default), the ``--hwdec`` option is used.
-
-    For ``gpu``, if set, do not create the interop context on demand, but
-    when the VO is created.
-
-    For ``opengl-cb``, if set, load the interop context as soon as the OpenGL
-    context is created. Since ``opengl-cb`` has no on-demand loading, this
-    allows enabling hardware decoding at runtime at all, without having
-    to temporarily set the ``hwdec`` option just during OpenGL context
-    initialization with ``mpv_opengl_cb_init_gl()``.
-
-    See ``--opengl-hwdec-interop=help`` for accepted values. This lists the
-    interop backend, with the ``--hwdec`` alias after it in ``[...]``. Consider
-    all values except the proper interop backend name, ``auto``, and ``no`` as
-    silently deprecated and subject to change. Also, if you use this in
-    application code (e.g. via libmpv), any value other than ``auto`` and ``no``
-    should be avoided, as backends can change.
-
-    Currently the option sets a single value. It is possible that the option
-    type changes to a list in the future.
-
-    The old alias ``--hwdec-preload`` has different behavior if the option value
-    is ``no``.
+``--gpu-hwdec-interop=<auto|all|no|name>``
+    This option is for troubleshooting hwdec interop issues. Since it's a
+    debugging option, its semantics may change at any time.
+
+    This is useful for the ``gpu`` and ``opengl-cb`` VOs for selecting which
+    hwdec interop context to use exactly. Effectively it also can be used
+    to block loading of certain backends.
+
+    If set to ``auto`` (default), the behavior depends on the VO: for ``gpu``,
+    it does nothing, and the interop context is loaded on demand (when the
+    decoder probes for ``--hwdec`` support). For ``opengl-cb``, which has
+    has no on-demand loading, this is equivalent to ``all``.
+
+    The empty string is equivalent to ``auto``.
+
+    If set to ``all``, it attempts to load all interop contexts at GL context
+    creation time.
+
+    Other than that, a specific backend can be set, and the list of them can
+    be queried with ``help`` (mpv CLI only).
+
+    Runtime changes to this are ignored (the current option value is used
+    whenever the renderer is created).
+
+    The old aliases ``--opengl-hwdec-interop`` and ``--hwdec-preload`` are
+    barely related to this anymore, but will be somewhat compatible in some
+    cases.
 
 ``--hwdec-image-format=<name>``
     Set the internal pixel format used by hardware decoding via ``--hwdec``
diff --git a/libmpv/opengl_cb.h b/libmpv/opengl_cb.h
index 0658a2e509..ccafbf5839 100644
--- a/libmpv/opengl_cb.h
+++ b/libmpv/opengl_cb.h
@@ -116,17 +116,18 @@ extern "C" {
  *
  * While "normal" mpv loads the OpenGL hardware decoding interop on demand,
  * this can't be done with opengl_cb for internal technical reasons. Instead,
- * make it load the interop at load time by setting the
- * "opengl-hwdec-interop"="auto" option before calling mpv_opengl_cb_init_gl()
- * ("hwdec-preload" in older mpv releases).
+ * it loads them by default, even if hardware decoding is not going to be used.
+ * In older mpv relases, this had to be done by setting the
+ * "opengl-hwdec-interop" or "hwdec-preload" options before calling
+ * mpv_opengl_cb_init_gl(). You can still use the newer "gpu-hwdec-interop"
+ * option to prevent loading of interop, or to load only a specific interop.
  *
  * There may be certain requirements on the OpenGL implementation:
  * - Windows: ANGLE is required (although in theory GL/DX interop could be used)
  * - Intel/Linux: EGL is required, and also a glMPGetNativeDisplay() callback
  *                must be provided (see sections below)
- * - nVidia/Linux: GLX is required (if you force "cuda", it should work on EGL
- *                 as well, if you have recent enough drivers and the
- *                 "hwaccel" option is set to "cuda" as well)
+ * - nVidia/Linux: Both GLX and EGL should work (GLX is required if vdpau is
+ *                 used, e.g. due to old drivers.)
  * - OSX: CGL is required (CGLGetCurrentContext() returning non-NULL)
  * - iOS: EAGL is required (EAGLContext.currentContext returning non-nil)
  *
diff --git a/options/options.c b/options/options.c
index 348cc2f5d2..c6837a03e8 100644
--- a/options/options.c
+++ b/options/options.c
@@ -55,8 +55,6 @@
 #include "video/out/drm_common.h"
 #endif
 
-#include "video/out/gpu/hwdec.h"
-
 static void print_version(struct mp_log *log)
 {
     mp_print_version(log, true);
@@ -184,9 +182,6 @@ static const m_option_t mp_vo_opt_list[] = {
 #if HAVE_DRM
     OPT_SUBSTRUCT("", drm_opts, drm_conf, 0),
 #endif
-    OPT_STRING_VALIDATE("opengl-hwdec-interop", gl_hwdec_interop, 0,
-                        ra_hwdec_validate_opt),
-    OPT_REPLACED("hwdec-preload", "opengl-hwdec-interop"),
     {0}
 };
 
diff --git a/options/options.h b/options/options.h
index e2e1220fda..8bf8f1ec6a 100644
--- a/options/options.h
+++ b/options/options.h
@@ -50,10 +50,8 @@ typedef struct mp_vo_opts {
 
     char *mmcss_profile;
 
-    // vo_wayland, vo_drm
+    // vo_drm
     struct sws_opts *sws_opts;
-    // vo_opengl, vo_opengl_cb
-    char *gl_hwdec_interop;
     // vo_drm
     struct drm_opts *drm_opts;
 } mp_vo_opts;
diff --git a/player/command.c b/player/command.c
index 569272c680..9017278286 100644
--- a/player/command.c
+++ b/player/command.c
@@ -2475,14 +2475,10 @@ static int mp_property_hwdec_interop(void *ctx, struct m_property *prop,
     if (!mpctx->video_out || !mpctx->video_out->hwdec_devs)
         return M_PROPERTY_UNAVAILABLE;
 
-    struct mp_hwdec_ctx *hwctx =
-        hwdec_devices_get_first(mpctx->video_out->hwdec_devs);
-
-    const char *name = hwctx ? hwctx->driver_name : NULL;
-    if (!name && hwctx)
-        name = m_opt_choice_str(mp_hwdec_names, hwctx->type);
-
-    return m_property_strdup_ro(action, arg, name);
+    char *names = hwdec_devices_get_names(mpctx->video_out->hwdec_devs);
+    int res = m_property_strdup_ro(action, arg, names);
+    talloc_free(names);
+    return res;
 }
 
 /// Helper to set vo flags.
diff --git a/video/decode/vd_lavc.c b/video/decode/vd_lavc.c
index 2362e9aa73..21e86b91b2 100644
--- a/video/decode/vd_lavc.c
+++ b/video/decode/vd_lavc.c
@@ -370,7 +370,7 @@ static struct mp_hwdec_ctx *hwdec_create_dev(struct dec_video *vd,
         return hwdec->create_dev(vd->global, vd->log, autoprobe);
     if (vd->hwdec_devs) {
         int type = hwdec->interop_type ? hwdec->interop_type : hwdec->type;
-        hwdec_devices_request(vd->hwdec_devs, type);
+        hwdec_devices_request_all(vd->hwdec_devs);
         return hwdec_devices_get(vd->hwdec_devs, type);
     }
     return NULL;
diff --git a/video/hwdec.c b/video/hwdec.c
index df0cf88839..31df1522ce 100644
--- a/video/hwdec.c
+++ b/video/hwdec.c
@@ -8,9 +8,10 @@
 struct mp_hwdec_devices {
     pthread_mutex_t lock;
 
-    struct mp_hwdec_ctx *hwctx;
+    struct mp_hwdec_ctx **hwctxs;
+    int num_hwctxs;
 
-    void (*load_api)(void *ctx, enum hwdec_type type);
+    void (*load_api)(void *ctx);
     void *load_api_ctx;
 };
 
@@ -25,7 +26,7 @@ void hwdec_devices_destroy(struct mp_hwdec_devices *devs)
 {
     if (!devs)
         return;
-    assert(!devs->hwctx); // must have been hwdec_devices_remove()ed
+    assert(!devs->num_hwctxs); // must have been hwdec_devices_remove()ed
     assert(!devs->load_api); // must have been unset
     pthread_mutex_destroy(&devs->lock);
     talloc_free(devs);
@@ -36,8 +37,12 @@ struct mp_hwdec_ctx *hwdec_devices_get(struct mp_hwdec_devices *devs,
 {
     struct mp_hwdec_ctx *res = NULL;
     pthread_mutex_lock(&devs->lock);
-    if (devs->hwctx && devs->hwctx->type == type)
-        res = devs->hwctx;
+    for (int n = 0; n < devs->num_hwctxs; n++) {
+        if (devs->hwctxs[n]->type == type) {
+            res = devs->hwctxs[n];
+            break;
+        }
+    }
     pthread_mutex_unlock(&devs->lock);
     return res;
 }
@@ -45,7 +50,7 @@ struct mp_hwdec_ctx *hwdec_devices_get(struct mp_hwdec_devices *devs,
 struct mp_hwdec_ctx *hwdec_devices_get_first(struct mp_hwdec_devices *devs)
 {
     pthread_mutex_lock(&devs->lock);
-    struct mp_hwdec_ctx *res = devs->hwctx;
+    struct mp_hwdec_ctx *res = devs->num_hwctxs ? devs->hwctxs[0] : NULL;
     pthread_mutex_unlock(&devs->lock);
     return res;
 }
@@ -53,42 +58,55 @@ struct mp_hwdec_ctx *hwdec_devices_get_first(struct mp_hwdec_devices *devs)
 void hwdec_devices_add(struct mp_hwdec_devices *devs, struct mp_hwdec_ctx *ctx)
 {
     pthread_mutex_lock(&devs->lock);
-    // We support only 1 device; ignore the rest.
-    if (!devs->hwctx)
-        devs->hwctx = ctx;
+    MP_TARRAY_APPEND(devs, devs->hwctxs, devs->num_hwctxs, ctx);
     pthread_mutex_unlock(&devs->lock);
 }
 
 void hwdec_devices_remove(struct mp_hwdec_devices *devs, struct mp_hwdec_ctx *ctx)
 {
     pthread_mutex_lock(&devs->lock);
-    if (devs->hwctx == ctx)
-        devs->hwctx = NULL;
+    for (int n = 0; n < devs->num_hwctxs; n++) {
+        if (devs->hwctxs[n] == ctx) {
+            MP_TARRAY_REMOVE_AT(devs->hwctxs, devs->num_hwctxs, n);
+            break;
+        }
+    }
     pthread_mutex_unlock(&devs->lock);
 }
 
 void hwdec_devices_set_loader(struct mp_hwdec_devices *devs,
-    void (*load_api)(void *ctx, enum hwdec_type type), void *load_api_ctx)
+    void (*load_api)(void *ctx), void *load_api_ctx)
 {
     devs->load_api = load_api;
     devs->load_api_ctx = load_api_ctx;
 }
 
-void hwdec_devices_request(struct mp_hwdec_devices *devs, enum hwdec_type type)
+void hwdec_devices_request_all(struct mp_hwdec_devices *devs)
 {
     if (devs->load_api && !hwdec_devices_get_first(devs))
-        devs->load_api(devs->load_api_ctx, type);
+        devs->load_api(devs->load_api_ctx);
 }
 
 void *hwdec_devices_load(struct mp_hwdec_devices *devs, enum hwdec_type type)
 {
     if (!devs)
         return NULL;
-    hwdec_devices_request(devs, type);
+    hwdec_devices_request_all(devs);
     struct mp_hwdec_ctx *hwctx = hwdec_devices_get(devs, type);
     return hwctx ? hwctx->ctx : NULL;
 }
 
+char *hwdec_devices_get_names(struct mp_hwdec_devices *devs)
+{
+    char *res = NULL;
+    for (int n = 0; n < devs->num_hwctxs; n++) {
+        if (res)
+            ta_xstrdup_append(&res, ",");
+        ta_xstrdup_append(&res, devs->hwctxs[n]->driver_name);
+    }
+    return res;
+}
+
 #if HAVE_D3D_HWACCEL
 extern const struct hwcontext_fns hwcontext_fns_d3d11;
 #endif
diff --git a/video/hwdec.h b/video/hwdec.h
index 2b89c3247c..3a551ef6f7 100644
--- a/video/hwdec.h
+++ b/video/hwdec.h
@@ -87,18 +87,18 @@ struct mp_hwdec_ctx *hwdec_devices_get_first(struct mp_hwdec_devices *devs);
 void hwdec_devices_add(struct mp_hwdec_devices *devs, struct mp_hwdec_ctx *ctx);
 
 // Remove this from the list of internal devices. Idempotent/ignores entries
-// not added yet.
+// not added yet. This is not thread-safe.
 void hwdec_devices_remove(struct mp_hwdec_devices *devs, struct mp_hwdec_ctx *ctx);
 
 // Can be used to enable lazy loading of an API with hwdec_devices_request().
 // If used at all, this must be set/unset during initialization/uninitialization,
 // as concurrent use with hwdec_devices_request() is a race condition.
 void hwdec_devices_set_loader(struct mp_hwdec_devices *devs,
-    void (*load_api)(void *ctx, enum hwdec_type type), void *load_api_ctx);
+    void (*load_api)(void *ctx), void *load_api_ctx);
 
-// Cause VO to lazily load the requested device, and will block until this is
-// done (even if not available).
-void hwdec_devices_request(struct mp_hwdec_devices *devs, enum hwdec_type type);
+// Cause VO to lazily load all devices, and will block until this is done (even
+// if not available).
+void hwdec_devices_request_all(struct mp_hwdec_devices *devs);
 
 // Convenience function:
 // - return NULL if devs==NULL
@@ -107,6 +107,9 @@ void hwdec_devices_request(struct mp_hwdec_devices *devs, enum hwdec_type type);
 // - then return the mp_hwdec_ctx.ctx field
 void *hwdec_devices_load(struct mp_hwdec_devices *devs, enum hwdec_type type);
 
+// Return "," concatenated list (for introspection/debugging). Use talloc_free().
+char *hwdec_devices_get_names(struct mp_hwdec_devices *devs);
+
 struct mp_image;
 
 // Per AV_HWDEVICE_TYPE_* functions, queryable via hwdec_get_hwcontext_fns().
diff --git a/video/out/gpu/hwdec.c b/video/out/gpu/hwdec.c
index 12f4dd7f64..4661e875bf 100644
--- a/video/out/gpu/hwdec.c
+++ b/video/out/gpu/hwdec.c
@@ -40,7 +40,7 @@ extern const struct ra_hwdec_driver ra_hwdec_cuda_nvdec;
 extern const struct ra_hwdec_driver ra_hwdec_rpi_overlay;
 extern const struct ra_hwdec_driver ra_hwdec_drmprime_drm;
 
-static const struct ra_hwdec_driver *const mpgl_hwdec_drivers[] = {
+const struct ra_hwdec_driver *const ra_hwdec_drivers[] = {
 #if HAVE_VAAPI_EGL
     &ra_hwdec_vaegl,
 #endif
@@ -76,11 +76,11 @@ static const struct ra_hwdec_driver *const mpgl_hwdec_drivers[] = {
     NULL
 };
 
-static struct ra_hwdec *load_hwdec_driver(struct mp_log *log, struct ra *ra,
-                                          struct mpv_global *global,
-                                          struct mp_hwdec_devices *devs,
-                                          const struct ra_hwdec_driver *drv,
-                                          bool is_auto)
+struct ra_hwdec *ra_hwdec_load_driver(struct ra *ra, struct mp_log *log,
+                                      struct mpv_global *global,
+                                      struct mp_hwdec_devices *devs,
+                                      const struct ra_hwdec_driver *drv,
+                                      bool is_auto)
 {
     struct ra_hwdec *hwdec = talloc(NULL, struct ra_hwdec);
     *hwdec = (struct ra_hwdec) {
@@ -101,81 +101,30 @@ static struct ra_hwdec *load_hwdec_driver(struct mp_log *log, struct ra *ra,
     return hwdec;
 }
 
-struct ra_hwdec *ra_hwdec_load_api(struct mp_log *log, struct ra *ra,
-                                   struct mpv_global *g,
-                                   struct mp_hwdec_devices *devs,
-                                   enum hwdec_type api)
-{
-    bool is_auto = HWDEC_IS_AUTO(api);
-    for (int n = 0; mpgl_hwdec_drivers[n]; n++) {
-        const struct ra_hwdec_driver *drv = mpgl_hwdec_drivers[n];
-        if ((is_auto || api == drv->api) && !drv->testing_only) {
-            struct ra_hwdec *r = load_hwdec_driver(log, ra, g, devs, drv, is_auto);
-            if (r)
-                return r;
-        }
-    }
-    return NULL;
-}
-
-// Load by option name.
-struct ra_hwdec *ra_hwdec_load(struct mp_log *log, struct ra *ra,
-                               struct mpv_global *g,
-                               struct mp_hwdec_devices *devs,
-                               const char *name)
-{
-    int g_hwdec_api;
-    mp_read_option_raw(g, "hwdec", &m_option_type_choice, &g_hwdec_api);
-    if (!name || !name[0])
-        name = m_opt_choice_str(mp_hwdec_names, g_hwdec_api);
-
-    int api_id = HWDEC_NONE;
-    for (int n = 0; mp_hwdec_names[n].name; n++) {
-        if (name && strcmp(mp_hwdec_names[n].name, name) == 0)
-            api_id = mp_hwdec_names[n].value;
-    }
-
-    for (int n = 0; mpgl_hwdec_drivers[n]; n++) {
-        const struct ra_hwdec_driver *drv = mpgl_hwdec_drivers[n];
-        if (name && strcmp(drv->name, name) == 0) {
-            struct ra_hwdec *r = load_hwdec_driver(log, ra, g, devs, drv, false);
-            if (r)
-                return r;
-        }
-    }
-
-    return ra_hwdec_load_api(log, ra, g, devs, api_id);
-}
-
 int ra_hwdec_validate_opt(struct mp_log *log, const m_option_t *opt,
                           struct bstr name, struct bstr param)
 {
     bool help = bstr_equals0(param, "help");
     if (help)
         mp_info(log, "Available hwdecs:\n");
-    for (int n = 0; mpgl_hwdec_drivers[n]; n++) {
-        const struct ra_hwdec_driver *drv = mpgl_hwdec_drivers[n];
-        const char *api_name = m_opt_choice_str(mp_hwdec_names, drv->api);
+    for (int n = 0; ra_hwdec_drivers[n]; n++) {
+        const struct ra_hwdec_driver *drv = ra_hwdec_drivers[n];
         if (help) {
-            mp_info(log, "    %s [%s]\n", drv->name, api_name);
-        } else if (bstr_equals0(param, drv->name) ||
-                   bstr_equals0(param, api_name))
-        {
+            mp_info(log, "    %s\n", drv->name);
+        } else if (bstr_equals0(param, drv->name)) {
             return 1;
         }
     }
     if (help) {
-        mp_info(log, "    auto (loads best)\n"
-                     "    (other --hwdec values)\n"
-                     "Setting an empty string means use --hwdec.\n");
+        mp_info(log, "    auto (behavior depends on context)\n"
+                     "    no (do not load any and block loading on demand)\n");
         return M_OPT_EXIT;
     }
     if (!param.len)
         return 1; // "" is treated specially
-    for (int n = 0; mp_hwdec_names[n].name; n++) {
-        if (bstr_equals0(param, mp_hwdec_names[n].name))
-            return 1;
-    }
+    if (bstr_equals0(param, "all") || bstr_equals0(param, "auto") ||
+        bstr_equals0(param, "no"))
+        return 1;
     mp_fatal(log, "No hwdec backend named '%.*s' found!\n", BSTR_P(param));
     return M_OPT_INVALID;
 }
diff --git a/video/out/gpu/hwdec.h b/video/out/gpu/hwdec.h
index 20bbaae9eb..0ff6056bfd 100644
--- a/video/out/gpu/hwdec.h
+++ b/video/out/gpu/hwdec.h
@@ -104,15 +104,13 @@ struct ra_hwdec_driver {
                          struct mp_rect *src, struct mp_rect *dst, bool newframe);
 };
 
-struct ra_hwdec *ra_hwdec_load_api(struct mp_log *log, struct ra *ra,
-                                   struct mpv_global *g,
-                                   struct mp_hwdec_devices *devs,
-                                   enum hwdec_type api);
-
-struct ra_hwdec *ra_hwdec_load(struct mp_log *log, struct ra *ra,
-                               struct mpv_global *g,
-                               struct mp_hwdec_devices *devs,
-                               const char *name);
+extern const struct ra_hwdec_driver *const ra_hwdec_drivers[];
+
+struct ra_hwdec *ra_hwdec_load_driver(struct ra *ra, struct mp_log *log,
+                                      struct mpv_global *global,
+                                      struct mp_hwdec_devices *devs,
+                                      const struct ra_hwdec_driver *drv,
+                                      bool is_auto);
 
 int ra_hwdec_validate_opt(struct mp_log *log, const m_option_t *opt,
                           struct bstr name, struct bstr param);
diff --git a/video/out/gpu/video.c b/video/out/gpu/video.c
index 4cdf20dfae..29af4b92b1 100644
--- a/video/out/gpu/video.c
+++ b/video/out/gpu/video.c
@@ -277,8 +277,12 @@ struct gl_video {
     struct cached_file *files;
     int num_files;
 
-    struct ra_hwdec *hwdec;
+    bool hwdec_interop_loading_done;
+    struct ra_hwdec **hwdecs;
+    int num_hwdecs;
+
     struct ra_hwdec_mapper *hwdec_mapper;
+    struct ra_hwdec *hwdec_overlay;
     bool hwdec_active;
 
     bool dsi_warned;
@@ -401,6 +405,10 @@ const struct m_sub_options gl_video_conf = {
         OPT_INTRANGE("gpu-tex-pad-y", tex_pad_y, 0, 0, 4096),
         OPT_SUBSTRUCT("", icc_opts, mp_icc_conf, 0),
         OPT_STRING("gpu-shader-cache-dir", shader_cache_dir, 0),
+        OPT_STRING_VALIDATE("gpu-hwdec-interop", hwdec_interop, 0,
+                             ra_hwdec_validate_opt),
+        OPT_REPLACED("opengl-hwdec-interop", "gpu-hwdec-interop"),
+        OPT_REPLACED("hwdec-preload", "opengl-hwdec-interop"),
         OPT_REPLACED("hdr-tone-mapping", "tone-mapping"),
         OPT_REPLACED("opengl-shaders", "glsl-shaders"),
         OPT_REPLACED("opengl-shader", "glsl-shader"),
@@ -803,18 +811,27 @@ static void init_video(struct gl_video *p)
 {
     p->use_integer_conversion = false;
 
-    if (p->hwdec && ra_hwdec_test_format(p->hwdec, p->image_params.imgfmt)) {
-        if (p->hwdec->driver->overlay_frame) {
+    struct ra_hwdec *hwdec = NULL;
+    for (int n = 0; n < p->num_hwdecs; n++) {
+        if (ra_hwdec_test_format(p->hwdecs[n], p->image_params.imgfmt)) {
+            hwdec = p->hwdecs[n];
+            break;
+        }
+    }
+
+    if (hwdec) {
+        if (hwdec->driver->overlay_frame) {
             MP_WARN(p, "Using HW-overlay mode. No GL filtering is performed "
                        "on the video!\n");
+            p->hwdec_overlay = hwdec;
         } else {
-            p->hwdec_mapper = ra_hwdec_mapper_create(p->hwdec, &p->image_params);
+            p->hwdec_mapper = ra_hwdec_mapper_create(hwdec, &p->image_params);
             if (!p->hwdec_mapper)
                 MP_ERR(p, "Initializing texture for hardware decoding failed.\n");
         }
         if (p->hwdec_mapper)
             p->image_params = p->hwdec_mapper->dst_params;
-        const char **exts = p->hwdec->glsl_extensions;
+        const char **exts = hwdec->glsl_extensions;
         for (int n = 0; exts && exts[n]; n++)
             gl_sc_enable_extension(p->sc, (char *)exts[n]);
         p->hwdec_active = true;
@@ -957,8 +974,8 @@ static void unref_current_image(struct gl_video *p)
 // lead to flickering artifacts.
 static void unmap_overlay(struct gl_video *p)
 {
-    if (p->hwdec_active && p->hwdec->driver->overlay_frame)
-        p->hwdec->driver->overlay_frame(p->hwdec, NULL, NULL, NULL, true);
+    if (p->hwdec_overlay)
+        p->hwdec_overlay->driver->overlay_frame(p->hwdec_overlay, NULL, NULL, NULL, true);
 }
 
 static void uninit_video(struct gl_video *p)
@@ -981,6 +998,7 @@ static void uninit_video(struct gl_video *p)
     p->real_image_params = (struct mp_image_params){0};
     p->image_params = p->real_image_params;
     p->hwdec_active = false;
+    p->hwdec_overlay = NULL;
     ra_hwdec_mapper_free(&p->hwdec_mapper);
 }
 
@@ -3005,15 +3023,15 @@ void gl_video_render_frame(struct gl_video *p, struct vo_frame *frame,
         p->ra->fns->clear(p->ra, fbo.tex, color, &target_rc);
     }
 
-    if (p->hwdec_active && p->hwdec->driver->overlay_frame) {
+    if (p->hwdec_overlay) {
         if (has_frame) {
-            float *color = p->hwdec->overlay_colorkey;
+            float *color = p->hwdec_overlay->overlay_colorkey;
             p->ra->fns->clear(p->ra, fbo.tex, color, &p->dst_rect);
         }
 
-        p->hwdec->driver->overlay_frame(p->hwdec, frame->current,
-                                        &p->src_rect, &p->dst_rect,
-                                        frame->frame_id != p->image.id);
+        p->hwdec_overlay->driver->overlay_frame(p->hwdec_overlay, frame->current,
+                                                &p->src_rect, &p->dst_rect,
+                                                frame->frame_id != p->image.id);
 
         if (frame->current)
             p->osd_pts = frame->current->pts;
@@ -3507,6 +3525,10 @@ void gl_video_uninit(struct gl_video *p)
 
     uninit_video(p);
 
+    for (int n = 0; n < p->num_hwdecs; n++)
+        ra_hwdec_uninit(p->hwdecs[n]);
+    p->num_hwdecs = 0;
+
     gl_sc_destroy(p->sc);
 
     ra_tex_free(p->ra, &p->lut_3d_texture);
@@ -3561,8 +3583,10 @@ bool gl_video_check_format(struct gl_video *p, int mp_format)
     if (ra_get_imgfmt_desc(p->ra, mp_format, &desc) &&
         is_imgfmt_desc_supported(p, &desc))
         return true;
-    if (p->hwdec && ra_hwdec_test_format(p->hwdec, mp_format))
-        return true;
+    for (int n = 0; n < p->num_hwdecs; n++) {
+        if (ra_hwdec_test_format(p->hwdecs[n], mp_format))
+            return true;
+    }
     return false;
 }
 
@@ -3782,13 +3806,6 @@ void gl_video_set_ambient_lux(struct gl_video *p, int lux)
     }
 }
 
-void gl_video_set_hwdec(struct gl_video *p, struct ra_hwdec *hwdec)
-{
-    unref_current_image(p);
-    ra_hwdec_mapper_free(&p->hwdec_mapper);
-    p->hwdec = hwdec;
-}
-
 static void *gl_video_dr_alloc_buffer(struct gl_video *p, size_t size)
 {
     struct ra_buf_params params = {
@@ -3845,3 +3862,47 @@ struct mp_image *gl_video_get_image(struct gl_video *p, int imgfmt, int w, int h
         gl_video_dr_free_buffer(p, ptr);
     return res;
 }
+
+static void load_add_hwdec(struct gl_video *p, struct mp_hwdec_devices *devs,
+                           const struct ra_hwdec_driver *drv, bool is_auto)
+{
+    struct ra_hwdec *hwdec =
+        ra_hwdec_load_driver(p->ra, p->log, p->global, devs, drv, is_auto);
+    if (hwdec)
+        MP_TARRAY_APPEND(p, p->hwdecs, p->num_hwdecs, hwdec);
+}
+
+void gl_video_load_hwdecs(struct gl_video *p, struct mp_hwdec_devices *devs,
+                          bool load_all_by_default)
+{
+    char *type = p->opts.hwdec_interop;
+    if (!type || !type[0] || strcmp(type, "auto") == 0) {
+        if (!load_all_by_default)
+            return;
+        type = "all";
+    }
+    if (strcmp(type, "no") == 0) {
+        // do nothing, just block further loading
+    } else if (strcmp(type, "all") == 0) {
+        for (int n = 0; ra_hwdec_drivers[n]; n++)
+            load_add_hwdec(p, devs, ra_hwdec_drivers[n], true);
+    } else {
+        for (int n = 0; ra_hwdec_drivers[n]; n++) {
+            const struct ra_hwdec_driver *drv = ra_hwdec_drivers[n];
+            if (strcmp(type, drv->name) == 0) {
+                load_add_hwdec(p, devs, drv, false);
+                break;
+            }
+        }
+    }
+    p->hwdec_interop_loading_done = true;
+}
+
+void gl_video_load_hwdecs_all(struct gl_video *p, struct mp_hwdec_devices *devs)
+{
+    if (!p->hwdec_interop_loading_done) {
+        for (int n = 0; ra_hwdec_drivers[n]; n++)
+            load_add_hwdec(p, devs, ra_hwdec_drivers[n], true);
+        p->hwdec_interop_loading_done = true;
+    }
+}
diff --git a/video/out/gpu/video.h b/video/out/gpu/video.h
index b87390f4d1..78f8828f99 100644
--- a/video/out/gpu/video.h
+++ b/video/out/gpu/video.h
@@ -139,6 +139,7 @@ struct gl_video_opts {
     struct mp_icc_opts *icc_opts;
     int early_flush;
     char *shader_cache_dir;
+    char *hwdec_interop;
 };
 
 extern const struct m_sub_options gl_video_conf;
@@ -176,8 +177,10 @@ struct mp_colorspace gl_video_get_output_colorspace(struct gl_video *p);
 void gl_video_reset(struct gl_video *p);
 bool gl_video_showing_interpolated_frame(struct gl_video *p);
 
-struct ra_hwdec;
-void gl_video_set_hwdec(struct gl_video *p, struct ra_hwdec *hwdec);
+struct mp_hwdec_devices;
+void gl_video_load_hwdecs(struct gl_video *p, struct mp_hwdec_devices *devs,
+                          bool load_all_by_default);
+void gl_video_load_hwdecs_all(struct gl_video *p, struct mp_hwdec_devices *devs);
 
 struct vo;
 void gl_video_configure_queue(struct gl_video *p, struct vo *vo);
diff --git a/video/out/vo_gpu.c b/video/out/vo_gpu.c
index d739cf6bfa..6a971dd94b 100644
--- a/video/out/vo_gpu.c
+++ b/video/out/vo_gpu.c
@@ -50,7 +50,6 @@ struct gpu_priv {
     char *context_type;
     struct ra_ctx_opts opts;
     struct gl_video *renderer;
-    struct ra_hwdec *hwdec;
 
     int events;
 };
@@ -120,23 +119,18 @@ static int reconfig(struct vo *vo, struct mp_image_params *params)
     return 0;
 }
 
-static void request_hwdec_api(struct vo *vo, void *api)
+static void request_hwdec_api(struct vo *vo)
 {
     struct gpu_priv *p = vo->priv;
 
-    if (p->hwdec)
-        return;
-
-    p->hwdec = ra_hwdec_load_api(vo->log, p->ctx->ra, vo->global,
-                                 vo->hwdec_devs, (intptr_t)api);
-    gl_video_set_hwdec(p->renderer, p->hwdec);
+    gl_video_load_hwdecs_all(p->renderer, vo->hwdec_devs);
 }
 
-static void call_request_hwdec_api(void *ctx, enum hwdec_type type)
+static void call_request_hwdec_api(void *ctx)
 {
     // Roundabout way to run hwdec loading on the VO thread.
     // Redirects to request_hwdec_api().
-    vo_control(ctx, VOCTRL_LOAD_HWDEC_API, (void *)(intptr_t)type);
+    vo_control(ctx, VOCTRL_LOAD_HWDEC_API, NULL);
 }
 
 static void get_and_update_icc_profile(struct gpu_priv *p)
@@ -195,7 +189,7 @@ static int control(struct vo *vo, uint32_t request, void *data)
         return true;
     }
     case VOCTRL_LOAD_HWDEC_API:
-        request_hwdec_api(vo, data);
+        request_hwdec_api(vo);
         return true;
     case VOCTRL_UPDATE_RENDER_OPTS: {
         gl_video_configure_queue(p->renderer, vo);
@@ -266,7 +260,6 @@ static void uninit(struct vo *vo)
     struct gpu_priv *p = vo->priv;
 
     gl_video_uninit(p->renderer);
-    ra_hwdec_uninit(p->hwdec);
     if (vo->hwdec_devs) {
         hwdec_devices_set_loader(vo->hwdec_devs, NULL, NULL);
         hwdec_devices_destroy(vo->hwdec_devs);
@@ -298,12 +291,9 @@ static int preinit(struct vo *vo)
     get_and_update_icc_profile(p);
 
     vo->hwdec_devs = hwdec_devices_create();
-
     hwdec_devices_set_loader(vo->hwdec_devs, call_request_hwdec_api, vo);
 
-    p->hwdec = ra_hwdec_load(vo->log, p->ctx->ra, vo->global,
-                             vo->hwdec_devs, vo->opts->gl_hwdec_interop);
-    gl_video_set_hwdec(p->renderer, p->hwdec);
+    gl_video_load_hwdecs(p->renderer, vo->hwdec_devs, false);
 
     return 0;
 
diff --git a/video/out/vo_opengl_cb.c b/video/out/vo_opengl_cb.c
index c903fcd43b..c8dab15bb8 100644
--- a/video/out/vo_opengl_cb.c
+++ b/video/out/vo_opengl_cb.c
@@ -89,7 +89,6 @@ struct mpv_opengl_cb_context {
     GL *gl;
     struct ra_ctx *ra_ctx;
     struct gl_video *renderer;
-    struct ra_hwdec *hwdec;
     struct m_config_cache *vo_opts_cache;
     struct mp_vo_opts *vo_opts;
 };
@@ -198,12 +197,8 @@ int mpv_opengl_cb_init_gl(struct mpv_opengl_cb_context *ctx, const char *exts,
 
     ctx->renderer = gl_video_init(ctx->ra_ctx->ra, ctx->log, ctx->global);
 
-    m_config_cache_update(ctx->vo_opts_cache);
-
     ctx->hwdec_devs = hwdec_devices_create();
-    ctx->hwdec = ra_hwdec_load(ctx->log, ctx->ra_ctx->ra, ctx->global,
-                               ctx->hwdec_devs, ctx->vo_opts->gl_hwdec_interop);
-    gl_video_set_hwdec(ctx->renderer, ctx->hwdec);
+    gl_video_load_hwdecs(ctx->renderer, ctx->hwdec_devs, true);
 
     pthread_mutex_lock(&ctx->lock);
     for (int n = IMGFMT_START; n < IMGFMT_END; n++) {
@@ -238,8 +233,6 @@ int mpv_opengl_cb_uninit_gl(struct mpv_opengl_cb_context *ctx)
 
     gl_video_uninit(ctx->renderer);
     ctx->renderer = NULL;
-    ra_hwdec_uninit(ctx->hwdec);
-    ctx->hwdec = NULL;
     hwdec_devices_destroy(ctx->hwdec_devs);
     ctx->hwdec_devs = NULL;
     ra_gl_ctx_uninit(ctx->ra_ctx);