video: rewrite filtering glue code

Get rid of the old vf.c code. Replace it with a generic filtering
framework, which can potentially handle more than just --vf. At least
reimplementing --af with this code is planned.

This changes some --vf semantics (including runtime behavior and the
"vf" command). The most important ones are listed in interface-changes.

vf_convert.c is renamed to f_swscale.c. It is now an internal filter
that can not be inserted by the user manually.

f_lavfi.c is a refactor of player/lavfi.c. The latter will be removed
once --lavfi-complex is reimplemented on top of f_lavfi.c. (which is
conceptually easy, but a big mess due to the data flow changes).

The existing filters are all changed heavily. The data flow of the new
filter framework is different. Especially EOF handling changes - EOF is
now a "frame" rather than a state, and must be passed through exactly
once.

Another major thing is that all filters must support dynamic format
changes. The filter reconfig() function goes away. (This sounds complex,
but since all filters need to handle EOF draining anyway, they can use
the same code, and it removes the mess with reconfig() having to predict
the output format, which completely breaks with libavfilter anyway.)

In addition, there is no automatic format negotiation or conversion.
libavfilter's primitive and insufficient API simply doesn't allow us to
do this in a reasonable way. Instead, filters can use f_autoconvert as
sub-filter, and tell it which formats they support. This filter will in
turn add actual conversion filters, such as f_swscale, to perform
necessary format changes.

vf_vapoursynth.c uses the same basic principle of operation as before,
but with worryingly different details in data flow. Still appears to
work.

The hardware deint filters (vf_vavpp.c, vf_d3d11vpp.c, vf_vdpaupp.c) are
heavily changed. Fortunately, they all used refqueue.c, which is for
sharing the data flow logic (especially for managing future/past
surfaces and such). It turns out it can be used to factor out most of
the data flow. Some of these filters accepted software input. Instead of
having ad-hoc upload code in each filter, surface upload is now
delegated to f_autoconvert, which can use f_hwupload to perform this.

Exporting VO capabilities is still a big mess (mp_stream_info stuff).

The D3D11 code drops the redundant image formats, and all code uses the
hw_subfmt (sw_format in FFmpeg) instead. Although that too seems to be a
big mess for now.

f_async_queue is unused.

21 files changed, 4632 insertions, 0 deletions

diff --git a/filters/f_auto_filters.c b/filters/f_auto_filters.c
new file mode 100644
index 0000000000..eac6f745ca
--- /dev/null
+++ b/filters/f_auto_filters.c
@@ -0,0 +1,244 @@
+#include <math.h>
+
+#include "common/common.h"
+#include "common/msg.h"
+#include "options/m_config.h"
+#include "options/options.h"
+#include "video/mp_image.h"
+
+#include "f_auto_filters.h"
+#include "f_swscale.h"
+#include "f_utils.h"
+#include "filter.h"
+#include "filter_internal.h"
+#include "user_filters.h"
+
+struct deint_priv {
+    struct mp_subfilter sub;
+    int prev_imgfmt;
+    int prev_setting;
+    struct m_config_cache *opts;
+};
+
+static void deint_process(struct mp_filter *f)
+{
+    struct deint_priv *p = f->priv;
+
+    if (!mp_subfilter_read(&p->sub))
+        return;
+
+    struct mp_frame frame = p->sub.frame;
+
+    if (mp_frame_is_signaling(frame)) {
+        mp_subfilter_continue(&p->sub);
+        return;
+    }
+
+    if (frame.type != MP_FRAME_VIDEO) {
+        MP_ERR(f, "video input required!\n");
+        mp_filter_internal_mark_failed(f);
+        return;
+    }
+
+    m_config_cache_update(p->opts);
+    struct filter_opts *opts = p->opts->opts;
+
+    if (!opts->deinterlace)
+        mp_subfilter_destroy(&p->sub);
+
+    struct mp_image *img = frame.data;
+
+    if (img->imgfmt == p->prev_imgfmt && p->prev_setting == opts->deinterlace) {
+        mp_subfilter_continue(&p->sub);
+        return;
+    }
+
+    if (!mp_subfilter_drain_destroy(&p->sub))
+        return;
+
+    assert(!p->sub.filter);
+
+    p->prev_imgfmt = img->imgfmt;
+    p->prev_setting = opts->deinterlace;
+    if (!p->prev_setting) {
+        mp_subfilter_continue(&p->sub);
+        return;
+    }
+
+    if (img->imgfmt == IMGFMT_VDPAU) {
+        char *args[] = {"deint", "yes", NULL};
+        p->sub.filter =
+            mp_create_user_filter(f, MP_OUTPUT_CHAIN_VIDEO, "vdpaupp", args);
+    } else if (img->imgfmt == IMGFMT_VAAPI) {
+        p->sub.filter =
+            mp_create_user_filter(f, MP_OUTPUT_CHAIN_VIDEO, "vavpp", NULL);
+    } else if (img->imgfmt == IMGFMT_D3D11) {
+        p->sub.filter =
+            mp_create_user_filter(f, MP_OUTPUT_CHAIN_VIDEO, "d3d11vpp", NULL);
+    } else if (mp_sws_supports_input(img->imgfmt)) {
+        char *args[] = {"mode", "send_field", "deint", "interlaced", NULL};
+        p->sub.filter =
+            mp_create_user_filter(f, MP_OUTPUT_CHAIN_VIDEO, "yadif", args);
+    } else {
+        MP_ERR(f, "no deinterlace filter available for this format\n");
+        mp_subfilter_continue(&p->sub);
+        return;
+    }
+
+    if (!p->sub.filter)
+        MP_ERR(f, "creating deinterlacer failed\n");
+
+    mp_subfilter_continue(&p->sub);
+}
+
+static void deint_reset(struct mp_filter *f)
+{
+    struct deint_priv *p = f->priv;
+
+    mp_subfilter_reset(&p->sub);
+}
+
+static void deint_destroy(struct mp_filter *f)
+{
+    struct deint_priv *p = f->priv;
+
+    mp_subfilter_reset(&p->sub);
+    TA_FREEP(&p->sub.filter);
+}
+
+static const struct mp_filter_info deint_filter = {
+    .name = "deint",
+    .priv_size = sizeof(struct deint_priv),
+    .process = deint_process,
+    .reset = deint_reset,
+    .destroy = deint_destroy,
+};
+
+struct mp_filter *mp_deint_create(struct mp_filter *parent)
+{
+    struct mp_filter *f = mp_filter_create(parent, &deint_filter);
+    if (!f)
+        return NULL;
+
+    struct deint_priv *p = f->priv;
+
+    p->sub.in = mp_filter_add_pin(f, MP_PIN_IN, "in");
+    p->sub.out = mp_filter_add_pin(f, MP_PIN_OUT, "out");
+
+    p->opts = m_config_cache_alloc(f, f->global, &filter_conf);
+
+    return f;
+}
+
+struct rotate_priv {
+    struct mp_subfilter sub;
+    int prev_rotate;
+    int prev_imgfmt;
+    int target_rotate;
+};
+
+static void rotate_process(struct mp_filter *f)
+{
+    struct rotate_priv *p = f->priv;
+
+    if (!mp_subfilter_read(&p->sub))
+        return;
+
+    struct mp_frame frame = p->sub.frame;
+
+    if (mp_frame_is_signaling(frame)) {
+        mp_subfilter_continue(&p->sub);
+        return;
+    }
+
+    if (frame.type != MP_FRAME_VIDEO) {
+        MP_ERR(f, "video input required!\n");
+        return;
+    }
+
+    struct mp_image *img = frame.data;
+
+    if (img->params.rotate == p->prev_rotate &&
+        img->imgfmt == p->prev_imgfmt)
+    {
+        img->params.rotate = p->target_rotate;
+        mp_subfilter_continue(&p->sub);
+        return;
+    }
+
+    if (!mp_subfilter_drain_destroy(&p->sub))
+        return;
+
+    assert(!p->sub.filter);
+
+    int rotate = p->prev_rotate = img->params.rotate;
+    p->target_rotate = rotate;
+    p->prev_imgfmt = img->imgfmt;
+
+    struct mp_stream_info *info = mp_filter_find_stream_info(f);
+    if (rotate == 0 || (info && info->rotate90 && !(rotate % 90))) {
+        mp_subfilter_continue(&p->sub);
+        return;
+    }
+
+    if (mp_sws_supports_input(img->imgfmt)) {
+        MP_ERR(f, "Video rotation with this format not supported\n");
+        mp_subfilter_continue(&p->sub);
+        return;
+    }
+
+    double angle = rotate / 360.0 * M_PI * 2;
+    char *args[] = {"angle", mp_tprintf(30, "%f", angle),
+                    "ow", mp_tprintf(30, "rotw(%f)", angle),
+                    "oh", mp_tprintf(30, "roth(%f)", angle),
+                    NULL};
+    p->sub.filter =
+        mp_create_user_filter(f, MP_OUTPUT_CHAIN_VIDEO, "rotate", args);
+
+    if (p->sub.filter) {
+        MP_INFO(f, "Inserting rotation filter.\n");
+        p->target_rotate = 0;
+    } else {
+        MP_ERR(f, "could not create rotation filter\n");
+    }
+
+    mp_subfilter_continue(&p->sub);
+}
+
+static void rotate_reset(struct mp_filter *f)
+{
+    struct rotate_priv *p = f->priv;
+
+    mp_subfilter_reset(&p->sub);
+}
+
+static void rotate_destroy(struct mp_filter *f)
+{
+    struct rotate_priv *p = f->priv;
+
+    mp_subfilter_reset(&p->sub);
+    TA_FREEP(&p->sub.filter);
+}
+
+static const struct mp_filter_info rotate_filter = {
+    .name = "autorotate",
+    .priv_size = sizeof(struct rotate_priv),
+    .process = rotate_process,
+    .reset = rotate_reset,
+    .destroy = rotate_destroy,
+};
+
+struct mp_filter *mp_autorotate_create(struct mp_filter *parent)
+{
+    struct mp_filter *f = mp_filter_create(parent, &rotate_filter);
+    if (!f)
+        return NULL;
+
+    struct rotate_priv *p = f->priv;
+    p->prev_rotate = -1;
+
+    p->sub.in = mp_filter_add_pin(f, MP_PIN_IN, "in");
+    p->sub.out = mp_filter_add_pin(f, MP_PIN_OUT, "out");
+
+    return f;
+}
diff --git a/filters/f_auto_filters.h b/filters/f_auto_filters.h
new file mode 100644
index 0000000000..5f1a99f636
--- /dev/null
+++ b/filters/f_auto_filters.h
@@ -0,0 +1,10 @@
+#pragma once
+
+#include "filter.h"
+
+// A filter which inserts the required deinterlacing filter based on the
+// hardware decode mode and the deinterlace user option.
+struct mp_filter *mp_deint_create(struct mp_filter *parent);
+
+// Rotate according to mp_image.rotate and VO capabilities.
+struct mp_filter *mp_autorotate_create(struct mp_filter *parent);
diff --git a/filters/f_autoconvert.c b/filters/f_autoconvert.c
new file mode 100644
index 0000000000..687a846ae5
--- /dev/null
+++ b/filters/f_autoconvert.c
@@ -0,0 +1,288 @@
+#include "config.h"
+
+#include "common/common.h"
+#include "common/msg.h"
+#include "video/hwdec.h"
+#include "video/mp_image.h"
+
+#include "f_autoconvert.h"
+#include "f_hwtransfer.h"
+#include "f_swscale.h"
+#include "f_utils.h"
+#include "filter.h"
+#include "filter_internal.h"
+
+struct priv {
+    struct mp_log *log;
+
+    struct mp_subfilter sub;
+
+    bool force_update;
+
+    int *imgfmts;
+    int *subfmts;
+    int num_imgfmts;
+
+    // Enable special conversion for the final stage before the VO.
+    bool vo_convert;
+
+    // sws state
+    int in_imgfmt, in_subfmt;
+
+    struct mp_autoconvert public;
+};
+
+// Dummy filter for bundling sub-conversion filters.
+static const struct mp_filter_info convert_filter = {
+    .name = "convert",
+};
+
+// For hw decoding: thing which can convert between underlying surface formats.
+// The filter detects the needed target format from struct mp_hwdec_ctx.
+struct subfmt_conv {
+    int hw_imgfmt;
+    struct mp_filter *(*create)(struct mp_filter *parent);
+};
+
+static const struct subfmt_conv subfmt_converters[] = {
+#if HAVE_D3D_HWACCEL
+    {IMGFMT_D3D11, vf_d3d11_create_outconv},
+#endif
+    {0}
+};
+
+void mp_autoconvert_clear(struct mp_autoconvert *c)
+{
+    struct priv *p = c->f->priv;
+
+    p->num_imgfmts = 0;
+}
+
+void mp_autoconvert_add_imgfmt(struct mp_autoconvert *c, int imgfmt, int subfmt)
+{
+    struct priv *p = c->f->priv;
+
+    MP_TARRAY_GROW(p, p->imgfmts, p->num_imgfmts);
+    MP_TARRAY_GROW(p, p->subfmts, p->num_imgfmts);
+
+    p->imgfmts[p->num_imgfmts] = imgfmt;
+    p->subfmts[p->num_imgfmts] = subfmt;
+
+    p->num_imgfmts += 1;
+    p->force_update = true;
+}
+
+void mp_autoconvert_add_vo_hwdec_subfmts(struct mp_autoconvert *c,
+                                         struct mp_hwdec_devices *devs)
+{
+    struct priv *p = c->f->priv;
+    assert(devs);
+
+    int prev_format = 0;
+
+    for (int n = 0; ; n++) {
+        struct mp_hwdec_ctx *ctx = hwdec_devices_get_n(devs, n);
+        if (!ctx)
+            break;
+        if (!ctx->hw_imgfmt || !ctx->supported_formats)
+            continue;
+        // Very hacky: don't let d3d11-egl-rgb overwrite d3d11-egl
+        if (ctx->hw_imgfmt == prev_format)
+            continue;
+        prev_format = ctx->hw_imgfmt;
+        // Stupidity: VOs export imgfmt only, so subfmt is always 0. Remove it
+        // to fix it up.
+        for (int i = 0; i < p->num_imgfmts; i++) {
+            if (p->imgfmts[i] != ctx->hw_imgfmt)
+                continue;
+
+            int count = p->num_imgfmts;
+            MP_TARRAY_REMOVE_AT(p->imgfmts, count, i);
+            count = p->num_imgfmts;
+            MP_TARRAY_REMOVE_AT(p->subfmts, count, i);
+            p->num_imgfmts -= 1;
+            break;
+        }
+        for (int i = 0; ctx->supported_formats[i]; i++)
+            mp_autoconvert_add_imgfmt(c, ctx->hw_imgfmt, ctx->supported_formats[i]);
+    }
+
+    p->vo_convert = true;
+}
+
+static void handle_video_frame(struct mp_filter *f)
+{
+    struct priv *p = f->priv;
+
+    struct mp_frame frame = p->sub.frame;
+    if (frame.type != MP_FRAME_VIDEO) {
+        MP_ERR(p, "video input required!\n");
+        mp_filter_internal_mark_failed(f);
+        return;
+    }
+
+    struct mp_image *img = frame.data;
+
+    if (p->force_update)
+        p->in_imgfmt = p->in_subfmt = 0;
+
+    if (img->imgfmt == p->in_imgfmt && img->params.hw_subfmt == p->in_subfmt) {
+        mp_subfilter_continue(&p->sub);
+        return;
+    }
+
+    if (!mp_subfilter_drain_destroy(&p->sub)) {
+        p->in_imgfmt = p->in_subfmt = 0;
+        return;
+    }
+
+    p->in_imgfmt = img->params.imgfmt;
+    p->in_subfmt = img->params.hw_subfmt;
+    p->force_update = false;
+
+    bool different_subfmt = false;
+
+    for (int n = 0; n < p->num_imgfmts; n++) {
+        bool samefmt = img->params.imgfmt == p->imgfmts[n];
+        bool samesubffmt = img->params.hw_subfmt == p->subfmts[n];
+        if (samefmt && !samesubffmt)
+            different_subfmt = true;
+        if (samefmt && (samesubffmt || !p->subfmts[n])) {
+            mp_subfilter_continue(&p->sub);
+            return;
+        }
+    }
+
+    struct mp_stream_info *info = mp_filter_find_stream_info(f);
+
+    struct mp_filter *conv = mp_filter_create(f, &convert_filter);
+    mp_filter_add_pin(conv, MP_PIN_IN, "in");
+    mp_filter_add_pin(conv, MP_PIN_OUT, "out");
+
+    struct mp_filter *filters[2] = {0};
+    bool need_sws = true;
+
+    int *fmts = p->imgfmts;
+    int num_fmts = p->num_imgfmts;
+
+    // Source is sw, all targets are hw -> try to upload.
+    bool sw_to_hw = !IMGFMT_IS_HWACCEL(img->imgfmt);
+    for (int n = 0; n < num_fmts; n++)
+        sw_to_hw &= IMGFMT_IS_HWACCEL(fmts[n]);
+
+    if (sw_to_hw && num_fmts > 0) {
+        // We can probably use this! Very lazy and very approximate.
+        struct mp_hwupload *upload = mp_hwupload_create(conv, fmts[0]);
+        if (upload) {
+            MP_INFO(p, "HW-uploading to %s\n", mp_imgfmt_to_name(fmts[0]));
+            filters[1] = upload->f;
+            fmts = upload->upload_fmts;
+            num_fmts = upload->num_upload_fmts;
+        }
+    } else if (p->vo_convert && different_subfmt && info && info->hwdec_devs) {
+        for (int n = 0; subfmt_converters[n].hw_imgfmt; n++) {
+            if (subfmt_converters[n].hw_imgfmt == img->imgfmt) {
+                MP_INFO(p, "Using HW sub-conversion.\n");
+                filters[1] = subfmt_converters[n].create(conv);
+                if (filters[1]) {
+                    need_sws = false;
+                    break;
+                }
+            }
+        }
+    }
+
+    if (need_sws) {
+        // Create a new conversion filter.
+        struct mp_sws_filter *sws = mp_sws_filter_create(conv);
+        if (!sws) {
+            MP_ERR(p, "error creating conversion filter\n");
+            return;
+        }
+
+        int out = mp_sws_find_best_out_format(img->imgfmt,  fmts, num_fmts);
+        if (!out) {
+            MP_ERR(p, "can't find video conversion for %s/%s\n",
+                   mp_imgfmt_to_name(img->imgfmt),
+                   mp_imgfmt_to_name(img->params.hw_subfmt));
+            talloc_free(conv);
+            mp_filter_internal_mark_failed(f);
+            return;
+        }
+
+        if (out == img->imgfmt) {
+            // Can happen if hwupload goes to same format.
+            talloc_free(sws->f);
+        } else {
+            sws->out_format = out;
+            MP_INFO(p, "Converting %s -> %s\n", mp_imgfmt_to_name(img->imgfmt),
+                    mp_imgfmt_to_name(sws->out_format));
+            filters[0] = sws->f;
+        }
+    }
+
+    mp_chain_filters(conv->ppins[0], conv->ppins[1], filters, 2);
+
+    p->sub.filter = conv;
+    mp_subfilter_continue(&p->sub);
+}
+
+static void process(struct mp_filter *f)
+{
+    struct priv *p = f->priv;
+
+    if (!mp_subfilter_read(&p->sub))
+        return;
+
+    struct mp_frame frame = p->sub.frame;
+
+    if (!mp_frame_is_signaling(frame)) {
+        if (p->num_imgfmts) {
+            handle_video_frame(f);
+            return;
+        }
+    }
+
+    mp_subfilter_continue(&p->sub);
+}
+
+static void reset(struct mp_filter *f)
+{
+    struct priv *p = f->priv;
+
+    mp_subfilter_reset(&p->sub);
+}
+
+static void destroy(struct mp_filter *f)
+{
+    struct priv *p = f->priv;
+
+    mp_subfilter_reset(&p->sub);
+    TA_FREEP(&p->sub.filter);
+}
+
+static const struct mp_filter_info autoconvert_filter = {
+    .name = "autoconvert",
+    .priv_size = sizeof(struct priv),
+    .process = process,
+    .reset = reset,
+    .destroy = destroy,
+};
+
+struct mp_autoconvert *mp_autoconvert_create(struct mp_filter *parent)
+{
+    struct mp_filter *f = mp_filter_create(parent, &autoconvert_filter);
+    if (!f)
+        return NULL;
+
+    mp_filter_add_pin(f, MP_PIN_IN, "in");
+    mp_filter_add_pin(f, MP_PIN_OUT, "out");
+
+    struct priv *p = f->priv;
+    p->public.f = f;
+    p->log = f->log;
+    p->sub.in = f->ppins[0];
+    p->sub.out = f->ppins[1];
+
+    return &p->public;
+}
diff --git a/filters/f_autoconvert.h b/filters/f_autoconvert.h
new file mode 100644
index 0000000000..72af21a0df
--- /dev/null
+++ b/filters/f_autoconvert.h
@@ -0,0 +1,39 @@
+#pragma once
+
+#include "filter.h"
+
+// A filter which automatically creates and uses a conversion filter based on
+// the filter settings, or passes through data unchanged if no conversion is
+// required.
+struct mp_autoconvert {
+    // f->pins[0] is input, f->pins[1] is output
+    struct mp_filter *f;
+};
+
+// (to free this, free the filter itself, mp_autoconvert.f)
+struct mp_autoconvert *mp_autoconvert_create(struct mp_filter *parent);
+
+// Add the imgfmt as allowed video image format, and error on non-video frames.
+// Each call adds to the list of allowed formats. Before the first call, all
+// formats are allowed (even non-video).
+// subfmt can be used to specify underlying surface formats for hardware formats,
+// otherwise must be 0.
+void mp_autoconvert_add_imgfmt(struct mp_autoconvert *c, int imgfmt, int subfmt);
+
+// Add the formats supported by the hwdec interops (or essentially refine them),
+// and trigger conversion if hw_subfmts mismatch. This is mostly a hack for
+// D3D11/ANGLE (which supports NV12 only).
+// Must be called mp_autoconvert_add_imgfmt(), and overrides them where formats
+// collide.
+struct mp_hwdec_devices;
+void mp_autoconvert_add_vo_hwdec_subfmts(struct mp_autoconvert *c,
+                                         struct mp_hwdec_devices *devs);
+
+// Reset set of allowed formats back to initial state. (This does not flush
+// any frames or remove currently active filters, although to get reasonable
+// behavior, you need to readd all previously allowed formats, or reset the
+// filter.)
+void mp_autoconvert_clear(struct mp_autoconvert *c);
+
+// vf_d3d11vpp.c
+struct mp_filter *vf_d3d11_create_outconv(struct mp_filter *parent);
diff --git a/filters/f_hwtransfer.c b/filters/f_hwtransfer.c
new file mode 100644
index 0000000000..6ffda567ae
--- /dev/null
+++ b/filters/f_hwtransfer.c
@@ -0,0 +1,299 @@
+#include <libavutil/buffer.h>
+#include <libavutil/hwcontext.h>
+#include <libavutil/mem.h>
+
+#include "video/fmt-conversion.h"
+#include "video/hwdec.h"
+#include "video/mp_image.h"
+#include "video/mp_image_pool.h"
+
+#include "f_hwtransfer.h"
+#include "filter_internal.h"
+
+struct priv {
+    AVBufferRef *av_device_ctx;
+
+    AVBufferRef *hw_pool;
+
+    int last_input_fmt;
+    int last_upload_fmt;
+    int last_sw_fmt;
+
+    struct mp_hwupload public;
+};
+
+static bool update_format_decision(struct priv *p, int input_fmt)
+{
+    struct mp_hwupload *u = &p->public;
+
+    if (!input_fmt)
+        return false;
+
+    if (input_fmt == p->last_input_fmt)
+        return true;
+
+    p->last_input_fmt = 0;
+
+    int res = mp_imgfmt_select_best_list(u->upload_fmts, u->num_upload_fmts,
+                                         input_fmt);
+
+    if (!res)
+        return false;
+
+    // Find which sw format we should use.
+    // NOTE: if there are ever any hw APIs that actually do expensive
+    // conversions on mismatching format uploads, we should probably first look
+    // which sw format is preferred?
+    int index = -1;
+    for (int n = 0; n < u->num_upload_fmts; n++) {
+        if (u->upload_fmts[n] == res)
+            index = n;
+    }
+
+    if (index < 0)
+        return false;
+
+    for (int n = 0; n < u->num_fmts; n++) {
+        if (u->fmt_upload_index[n] >= index &&
+            index  < u->fmt_upload_index[n] + u->fmt_upload_num[n])
+        {
+            p->last_input_fmt = input_fmt;
+            p->last_upload_fmt = u->upload_fmts[index];
+            p->last_sw_fmt = u->fmts[n];
+            MP_INFO(u->f, "upload %s -> %s\n",
+                    mp_imgfmt_to_name(p->last_sw_fmt),
+                    mp_imgfmt_to_name(p->last_input_fmt));
+            return true;
+        }
+    }
+
+    return false;
+}
+
+int mp_hwupload_find_upload_format(struct mp_hwupload *u, int imgfmt)
+{
+    struct priv *p = u->f->priv;
+
+    if (!update_format_decision(p, imgfmt))
+        return 0;
+    return p->last_upload_fmt;
+}
+
+static void process(struct mp_filter *f)
+{
+    struct priv *p = f->priv;
+
+    if (!mp_pin_can_transfer_data(f->ppins[1], f->ppins[0]))
+        return;
+
+    struct mp_frame frame = mp_pin_out_read(f->ppins[0]);
+    if (mp_frame_is_signaling(frame)) {
+        mp_pin_in_write(f->ppins[1], frame);
+        return;
+    }
+    if (frame.type != MP_FRAME_VIDEO) {
+        MP_ERR(f, "unsupported frame type\n");
+        goto error;
+    }
+    struct mp_image *src = frame.data;
+
+    // As documented, just pass though HW frames.
+    if (IMGFMT_IS_HWACCEL(src->imgfmt)) {
+        mp_pin_in_write(f->ppins[1], frame);
+        return;
+    }
+
+    if (src->w % 2 || src->h % 2) {
+        MP_ERR(f, "non-mod 2 input frames unsupported\n");
+        goto error;
+    }
+
+    if (!update_format_decision(p, src->imgfmt)) {
+        MP_ERR(f, "no hw upload format found\n");
+        goto error;
+    }
+
+    if (!mp_update_av_hw_frames_pool(&p->hw_pool, p->av_device_ctx,
+                                     p->public.hw_imgfmt, p->last_sw_fmt,
+                                     src->w, src->h))
+    {
+        MP_ERR(f, "failed to create frame pool\n");
+        goto error;
+    }
+
+    struct mp_image *dst = mp_av_pool_image_hw_upload(p->hw_pool, src);
+    if (!dst)
+        goto error;
+
+    mp_frame_unref(&frame);
+    mp_pin_in_write(f->ppins[1], MAKE_FRAME(MP_FRAME_VIDEO, dst));
+
+    return;
+
+error:
+    mp_frame_unref(&frame);
+    MP_ERR(f, "failed to upload frame\n");
+    mp_filter_internal_mark_failed(f);
+}
+
+static void destroy(struct mp_filter *f)
+{
+    struct priv *p = f->priv;
+
+    av_buffer_unref(&p->hw_pool);
+    av_buffer_unref(&p->av_device_ctx);
+}
+
+static const struct mp_filter_info hwupload_filter = {
+    .name = "hwupload",
+    .priv_size = sizeof(struct priv),
+    .process = process,
+    .destroy = destroy,
+};
+
+// The VO layer might have restricted format support. It might actually
+// work if this is input to a conversion filter anyway, but our format
+// negotiation is too stupid and non-existent to detect this.
+// So filter out all not explicitly supported formats.
+static bool vo_supports(struct mp_hwdec_ctx *ctx, int hw_fmt, int sw_fmt)
+{
+    if (!ctx->hw_imgfmt)
+        return true; // if unset, all formats are allowed
+    if (ctx->hw_imgfmt != hw_fmt)
+        return false;
+
+    for (int i = 0; ctx->supported_formats &&  ctx->supported_formats[i]; i++) {
+        if (ctx->supported_formats[i] == sw_fmt)
+            return true;
+    }
+
+    return false;
+}
+
+static bool probe_formats(struct mp_hwupload *u, int hw_imgfmt)
+{
+    struct priv *p = u->f->priv;
+
+    u->hw_imgfmt = hw_imgfmt;
+    u->num_fmts = 0;
+    u->num_upload_fmts = 0;
+
+    struct mp_stream_info *info = mp_filter_find_stream_info(u->f);
+    if (!info || !info->hwdec_devs) {
+        MP_ERR(u->f, "no hw context\n");
+        return false;
+    }
+
+    struct mp_hwdec_ctx *ctx = NULL;
+    AVHWFramesConstraints *cstr = NULL;
+
+    for (int n = 0; ; n++) {
+        struct mp_hwdec_ctx *cur = hwdec_devices_get_n(info->hwdec_devs, n);
+        if (!cur)
+            break;
+        if (!cur->av_device_ref)
+            continue;
+        cstr = av_hwdevice_get_hwframe_constraints(cur->av_device_ref, NULL);
+        if (!cstr)
+            continue;
+        bool found = false;
+        for (int i = 0; cstr->valid_hw_formats &&
+                        cstr->valid_hw_formats[i] != AV_PIX_FMT_NONE; i++)
+        {
+            found |= cstr->valid_hw_formats[i] == imgfmt2pixfmt(hw_imgfmt);
+        }
+        if (found && (!cur->hw_imgfmt || cur->hw_imgfmt == hw_imgfmt)) {
+            ctx = cur;
+            break;
+        }
+        av_hwframe_constraints_free(&cstr);
+    }
+
+    if (!ctx) {
+        MP_ERR(u->f, "no support for this hw format\n");
+        return false;
+    }
+
+    // Probe for supported formats. This is very roundabout, because the
+    // hwcontext API does not give us this information directly. We resort to
+    // creating temporary AVHWFramesContexts in order to retrieve the list of
+    // supported formats. This should be relatively cheap as we don't create
+    // any real frames (although some backends do for probing info).
+
+    for (int n = 0; cstr->valid_sw_formats &&
+                    cstr->valid_sw_formats[n] != AV_PIX_FMT_NONE; n++)
+    {
+        int imgfmt = pixfmt2imgfmt(cstr->valid_sw_formats[n]);
+        if (!imgfmt)
+            continue;
+
+        MP_VERBOSE(u->f, "looking at format %s\n", mp_imgfmt_to_name(imgfmt));
+
+        // Creates an AVHWFramesContexts with the given parameters.
+        AVBufferRef *frames = NULL;
+        if (!mp_update_av_hw_frames_pool(&frames, ctx->av_device_ref,
+                                        hw_imgfmt, imgfmt, 128, 128))
+        {
+            MP_WARN(u->f, "failed to allocate pool\n");
+            continue;
+        }
+
+        enum AVPixelFormat *fmts;
+        if (av_hwframe_transfer_get_formats(frames,
+                            AV_HWFRAME_TRANSFER_DIRECTION_TO, &fmts, 0) >= 0)
+        {
+            int index = u->num_fmts;
+            MP_TARRAY_APPEND(p, u->fmts, u->num_fmts, imgfmt);
+            MP_TARRAY_GROW(p, u->fmt_upload_index, index);
+            MP_TARRAY_GROW(p, u->fmt_upload_num, index);
+
+            u->fmt_upload_index[index] = u->num_upload_fmts;
+
+            for (int i = 0; fmts[i] != AV_PIX_FMT_NONE; i++) {
+                int fmt = pixfmt2imgfmt(fmts[i]);
+                if (!fmt)
+                    continue;
+                MP_VERBOSE(u->f, "supports %s\n", mp_imgfmt_to_name(fmt));
+                if (vo_supports(ctx, hw_imgfmt, fmt))
+                    MP_TARRAY_APPEND(p, u->upload_fmts, u->num_upload_fmts, fmt);
+            }
+
+            u->fmt_upload_n