[FFmpeg-devel,v6,08/12] swscale/graph: add new high-level scaler dispatch mechanism

Message ID 20241112095154.483778-9-ffmpeg@haasn.xyz
State New
Headers
Series swscale: introduce new, dynamic scaling API |

Checks

Context Check Description
yinshiyou/make_loongarch64 fail Make failed

Commit Message

Niklas Haas Nov. 12, 2024, 9:50 a.m. UTC
From: Niklas Haas <git@haasn.dev>

This interface has been designed from the ground up to serve as a new
framework for dispatching various scaling operations at a high level. This
will eventually replace the old ad-hoc system of using cascaded contexts,
as well as allowing us to plug in more dynamic scaling passes requiring
intermediate steps, such as colorspace conversions, etc.

The starter implementation merely piggybacks off the existing sws_init() and
sws_scale(), functions, though it does bring the immediate improvement of
splitting up cascaded functions and pre/post conversion functions into
separate filter passes, which allows them to e.g. be executed in parallel
even when the main scaler is required to be single threaded. Additionally,
a dedicated (multi-threaded) noop memcpy pass substantially improves
throughput of that fast path.

Follow-up commits will eventually expand this to move all of the scaling
decision logic into the graph init function, and also eliminate some of the
current special cases.

Sponsored-by: Sovereign Tech Fund
Signed-off-by: Niklas Haas <git@haasn.dev>
---
 libswscale/Makefile |   1 +
 libswscale/graph.c  | 597 ++++++++++++++++++++++++++++++++++++++++++++
 libswscale/graph.h  | 122 +++++++++
 3 files changed, 720 insertions(+)
 create mode 100644 libswscale/graph.c
 create mode 100644 libswscale/graph.h
  

Comments

Niklas Haas Nov. 12, 2024, 11:57 a.m. UTC | #1
On Tue, 12 Nov 2024 10:50:42 +0100 Niklas Haas <ffmpeg@haasn.xyz> wrote:
> From: Niklas Haas <git@haasn.dev>
>
> This interface has been designed from the ground up to serve as a new
> framework for dispatching various scaling operations at a high level. This
> will eventually replace the old ad-hoc system of using cascaded contexts,
> as well as allowing us to plug in more dynamic scaling passes requiring
> intermediate steps, such as colorspace conversions, etc.
>
> The starter implementation merely piggybacks off the existing sws_init() and
> sws_scale(), functions, though it does bring the immediate improvement of
> splitting up cascaded functions and pre/post conversion functions into
> separate filter passes, which allows them to e.g. be executed in parallel
> even when the main scaler is required to be single threaded. Additionally,
> a dedicated (multi-threaded) noop memcpy pass substantially improves
> throughput of that fast path.
>
> Follow-up commits will eventually expand this to move all of the scaling
> decision logic into the graph init function, and also eliminate some of the
> current special cases.

FWIW, I am currently working on another branch that pulls all per slice
state out of SwsContext and into a separate struct, so we can simplify our
threading wrappers in the new API.
  

Patch

diff --git a/libswscale/Makefile b/libswscale/Makefile
index 757997b401..81f32f4dd7 100644
--- a/libswscale/Makefile
+++ b/libswscale/Makefile
@@ -9,6 +9,7 @@  OBJS = alphablend.o                                     \
        hscale.o                                         \
        hscale_fast_bilinear.o                           \
        gamma.o                                          \
+       graph.o                                          \
        half2float.o                                     \
        input.o                                          \
        options.o                                        \
diff --git a/libswscale/graph.c b/libswscale/graph.c
new file mode 100644
index 0000000000..e0388bac61
--- /dev/null
+++ b/libswscale/graph.c
@@ -0,0 +1,597 @@ 
+/*
+ * Copyright (C) 2024 Niklas Haas
+ *
+ * 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 "libavutil/avassert.h"
+#include "libavutil/error.h"
+#include "libavutil/macros.h"
+#include "libavutil/mem.h"
+#include "libavutil/opt.h"
+#include "libavutil/pixdesc.h"
+#include "libavutil/slicethread.h"
+
+#include "libswscale/swscale.h"
+#include "libswscale/utils.h"
+
+#include "swscale_internal.h"
+#include "graph.h"
+
+/* slice_align should be a power of two, or 0 to disable slice threading */
+static SwsPass *pass_add(SwsGraph *graph, void *priv, int w, int h,
+                         sws_filter_run_t run, SwsImg in, SwsImg out,
+                         int slice_align)
+{
+    SwsPass *pass = av_mallocz(sizeof(*pass));
+    int ret;
+
+    pass->graph  = graph;
+    pass->run    = run;
+    pass->input  = in;
+    pass->output = out;
+    pass->priv   = priv;
+    pass->width  = w;
+    pass->height = h;
+
+    if (!slice_align) {
+        pass->slice_h = pass->height;
+        pass->num_slices = 1;
+    } else {
+        pass->slice_h = (pass->height + graph->num_threads - 1) / graph->num_threads;
+        pass->slice_h = FFALIGN(pass->slice_h, slice_align);
+        pass->num_slices = (pass->height + pass->slice_h - 1) / pass->slice_h;
+    }
+
+    ret = av_dynarray_add_nofree(&graph->passes, &graph->num_passes, pass);
+    if (ret < 0)
+        av_freep(&pass);
+    return pass;
+}
+
+/* Set output linesize before calling this */
+static int pass_alloc_output(SwsPass *pass)
+{
+    const int aligned_h = pass->num_slices * pass->slice_h;
+    const int *linesize = pass->output.linesize;
+
+    size_t offset[4];
+    size_t total_size = 0;
+    for (int i = 0; i < 4; i++) {
+        const size_t size = FFABS(linesize[i]) * aligned_h;
+        offset[i] = total_size;
+        total_size = FFALIGN(total_size + size, 16);
+    }
+
+    av_assert0(!pass->buf);
+    pass->buf = av_malloc(total_size);
+    if (!pass->buf)
+        return AVERROR(ENOMEM);
+
+    for (int i = 0; i < 4; i++) {
+        uint8_t *base = pass->buf + offset[i];
+        if (linesize[i] < 0)
+            base -= linesize[i] * (aligned_h - 1);
+        pass->output.data[i] = linesize[i] ? base : NULL;
+    }
+
+    return 0;
+}
+
+static int vshift(enum AVPixelFormat fmt, int plane)
+{
+    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(fmt);
+    const int is_pal = usePal(fmt);
+    const int is_chroma = (plane == 1 || plane == 2) && !is_pal;
+    return is_chroma ? desc->log2_chroma_h : 0;
+}
+
+/* Shift an image vertically by y lines */
+static SwsImg shift_img(const SwsImg *img_base, int y)
+{
+    SwsImg img = *img_base;
+    for (int i = 0; i < 4 && img.data[i]; i++)
+        img.data[i] += (y >> vshift(img.fmt, i)) * img.linesize[i];
+    return img;
+}
+
+static void setup_swscale(const SwsImg *out, const SwsImg *in,
+                          const SwsPass *pass)
+{
+    SwsContext *sws = pass->priv;
+    SwsInternal *c = sws_internal(sws);
+    if (sws->flags & SWS_BITEXACT && sws->dither == SWS_DITHER_ED && c->dither_error[0]) {
+        for (int i = 0; i < 4; i++)
+            memset(c->dither_error[i], 0, sizeof(c->dither_error[0][0]) * (sws->dst_w + 2));
+    }
+
+    if (usePal(sws->src_format))
+        ff_update_palette(c, (const uint32_t *) in->data[1]);
+}
+
+static void run_copy(const SwsImg *out_base, const SwsImg *in_base,
+                     int y, int h, const SwsPass *pass)
+{
+    SwsImg in  = shift_img(in_base,  y);
+    SwsImg out = shift_img(out_base, y);
+
+    for (int i = 0; i < FF_ARRAY_ELEMS(in.data) && in.data[i]; i++) {
+        const int lines = h >> vshift(in.fmt, i);
+        if (in.linesize[i] == out.linesize[i]) {
+            memcpy(out.data[i], in.data[i], lines * out.linesize[i]);
+        } else {
+            const int linesize = FFMIN(out.linesize[i], in.linesize[i]);
+            for (int j = 0; j < lines; j++) {
+                memcpy(out.data[i], in.data[i], linesize);
+                in.data[i]  += in.linesize[i];
+                out.data[i] += out.linesize[i];
+            }
+        }
+    }
+}
+
+static void run_rgb0(const SwsImg *out, const SwsImg *in, int y, int h,
+                     const SwsPass *pass)
+{
+    SwsInternal *c = pass->priv;
+    const int x0 = c->src0Alpha - 1;
+    const int w4 = 4 * pass->width;
+    const int src_stride = in->linesize[0];
+    const int dst_stride = out->linesize[0];
+    const uint8_t *src = in->data[0] + y * src_stride;
+    uint8_t *dst = out->data[0] + y * dst_stride;
+
+    for (int y = 0; y < h; y++) {
+        memcpy(dst, src, w4 * sizeof(*dst));
+        for (int x = x0; x < w4; x += 4)
+            dst[x] = 0xFF;
+
+        src += src_stride;
+        dst += dst_stride;
+    }
+}
+
+static void run_xyz2rgb(const SwsImg *out, const SwsImg *in, int y, int h,
+                        const SwsPass *pass)
+{
+    ff_xyz12Torgb48(pass->priv, out->data[0] + y * out->linesize[0], out->linesize[0],
+                    in->data[0] + y * in->linesize[0], in->linesize[0],
+                    pass->width, h);
+}
+
+static void run_rgb2xyz(const SwsImg *out, const SwsImg *in, int y, int h,
+                        const SwsPass *pass)
+{
+    ff_rgb48Toxyz12(pass->priv, out->data[0] + y * out->linesize[0], out->linesize[0],
+                    in->data[0] + y * in->linesize[0], in->linesize[0],
+                    pass->width, h);
+}
+
+static inline SwsContext *slice_ctx(const SwsPass *pass, int y)
+{
+    SwsContext *sws = pass->priv;
+    SwsInternal *parent = sws_internal(sws);
+    if (pass->num_slices == 1)
+        return sws;
+
+    av_assert1(parent->nb_slice_ctx == pass->num_slices);
+    sws = parent->slice_ctx[y / pass->slice_h];
+
+    if (usePal(sws->src_format)) {
+        SwsInternal *sub = sws_internal(sws);
+        memcpy(sub->pal_yuv, parent->pal_yuv, sizeof(sub->pal_yuv));
+        memcpy(sub->pal_rgb, parent->pal_rgb, sizeof(sub->pal_rgb));
+    }
+
+    return sws;
+}
+
+static void run_unscaled(const SwsImg *out, const SwsImg *in_base,
+                         int y, int h, const SwsPass *pass)
+{
+    SwsContext *sws = slice_ctx(pass, y);
+    SwsInternal *c = sws_internal(sws);
+    const SwsImg in = shift_img(in_base, y);
+
+    c->convert_unscaled(c, (const uint8_t *const *) in.data, in.linesize, y, h,
+                        out->data, out->linesize);
+}
+
+static void run_swscale(const SwsImg *out_base, const SwsImg *in,
+                        int y, int h, const SwsPass *pass)
+{
+    SwsContext *sws = slice_ctx(pass, y);
+    SwsInternal *c = sws_internal(sws);
+    const SwsImg out = shift_img(out_base, y);
+
+    ff_swscale(c, (const uint8_t *const *) in->data, in->linesize, 0,
+               sws->src_h, out.data, out.linesize, y, h);
+}
+
+static void get_chroma_pos(SwsGraph *graph, int *h_chr_pos, int *v_chr_pos,
+                           const SwsFormat *fmt)
+{
+    enum AVChromaLocation chroma_loc = fmt->loc;
+    const int sub_x = fmt->desc->log2_chroma_w;
+    const int sub_y = fmt->desc->log2_chroma_h;
+    int x_pos, y_pos;
+
+    /* Explicitly default to center siting for compatibility with swscale */
+    if (chroma_loc == AVCHROMA_LOC_UNSPECIFIED) {
+        chroma_loc = AVCHROMA_LOC_CENTER;
+        graph->incomplete |= sub_x || sub_y;
+    }
+
+    /* av_chroma_location_enum_to_pos() always gives us values in the range from
+     * 0 to 256, but we need to adjust this to the true value range of the
+     * subsampling grid, which may be larger for h/v_sub > 1 */
+    av_chroma_location_enum_to_pos(&x_pos, &y_pos, chroma_loc);
+    x_pos *= (1 << sub_x) - 1;
+    y_pos *= (1 << sub_y) - 1;
+
+    /* Fix vertical chroma position for interlaced frames */
+    if (sub_y && fmt->interlaced) {
+        /* When vertically subsampling, chroma samples are effectively only
+         * placed next to even rows. To access them from the odd field, we need
+         * to account for this shift by offsetting the distance of one luma row.
+         *
+         * For 4x vertical subsampling (v_sub == 2), they are only placed
+         * next to every *other* even row, so we need to shift by three luma
+         * rows to get to the chroma sample. */
+        if (graph->field == FIELD_BOTTOM)
+            y_pos += (256 << sub_y) - 256;
+
+        /* Luma row distance is doubled for fields, so halve offsets */
+        y_pos >>= 1;
+    }
+
+    /* Explicitly strip chroma offsets when not subsampling, because it
+     * interferes with the operation of flags like SWS_FULL_CHR_H_INP */
+    *h_chr_pos = sub_x ? x_pos : -513;
+    *v_chr_pos = sub_y ? y_pos : -513;
+}
+
+static int init_pass(SwsGraph *graph, SwsContext *sws,
+                     SwsImg input, SwsImg output)
+{
+    SwsInternal *c = sws_internal(sws);
+    const int src_w = sws->src_w, src_h = sws->src_h;
+    const int dst_w = sws->dst_w, dst_h = sws->dst_h;
+    const int unscaled = src_w == dst_w && src_h == dst_h;
+    int align = c->dst_slice_align;
+    SwsPass *pass;
+    int ret;
+
+    if (c->cascaded_context[0]) {
+        const int num_cascaded = c->cascaded_context[2] ? 3 : 2;
+        for (int i = 0; i < num_cascaded; i++) {
+            SwsContext *sub = c->cascaded_context[i];
+
+            if (i + 1 == num_cascaded) {
+                ret = init_pass(graph, sub, input, output);
+            } else {
+                /* Steal the intermediate buffers that were already allocated */
+                SwsImg tmp = { .fmt = sub->dst_format };
+                av_assert1(i < FF_ARRAY_ELEMS(c->cascaded_tmp));
+                memcpy(tmp.data, c->cascaded_tmp[i], sizeof(tmp.data));
+                memcpy(tmp.linesize, c->cascaded_tmpStride[i], sizeof(tmp.linesize));
+
+                ret = init_pass(graph, sub, input, tmp);
+                input = tmp;
+            }
+
+            if (ret < 0)
+                return ret;
+        }
+
+        return 0;
+    }
+
+    if (sws->dither == SWS_DITHER_ED && !c->convert_unscaled)
+        align = 0; /* disable slice threading */
+
+    if (c->src0Alpha && !c->dst0Alpha && isALPHA(sws->dst_format)) {
+        SwsImg tmp = {
+            .linesize = { FFALIGN(src_w * sizeof(uint8_t[4]), 16) },
+            .fmt = sws->dst_format,
+        };
+
+        SwsPass *sub = pass_add(graph, c, src_w, src_h, run_rgb0, input, tmp, 1);
+        if (!sub || pass_alloc_output(sub) < 0)
+            return AVERROR(ENOMEM);
+        input = sub->output;
+    }
+
+    if (c->srcXYZ && !(c->dstXYZ && unscaled)) {
+        SwsImg tmp = {
+            .linesize = { FFALIGN(src_w * sizeof(uint16_t[3]), 16) },
+            .fmt = sws->dst_format,
+        };
+
+        SwsPass *sub = pass_add(graph, c, src_w, src_h, run_xyz2rgb, input, tmp, 1);
+        if (!sub || pass_alloc_output(sub) < 0)
+            return AVERROR(ENOMEM);
+        input = sub->output;
+    }
+
+    pass = pass_add(graph, sws, dst_w, dst_h,
+                    c->convert_unscaled ? run_unscaled : run_swscale,
+                    input, output, align);
+    if (!pass)
+        return AVERROR(ENOMEM);
+    pass->setup = setup_swscale;
+
+    /**
+     * For slice threading, we need to create sub contexts, similar to how
+     * swscale normally handles it internally. The most important difference
+     * is that we handle cascaded contexts before threaded contexts; whereas
+     * context_init_threaded() does it the other way around.
+     */
+
+    if (pass->num_slices > 1) {
+        c->slice_ctx = av_calloc(pass->num_slices, sizeof(*c->slice_ctx));
+        if (!c->slice_ctx)
+            return AVERROR(ENOMEM);
+
+        for (int i = 0; i < pass->num_slices; i++) {
+            SwsContext *slice;
+            SwsInternal *c2;
+            slice = c->slice_ctx[i] = sws_alloc_context();
+            if (!slice)
+                return AVERROR(ENOMEM);
+            c->nb_slice_ctx++;
+
+            c2 = sws_internal(slice);
+            c2->parent = sws;
+
+            ret = av_opt_copy(slice, sws);
+            if (ret < 0)
+                return ret;
+
+            ret = sws_init_single_context(slice, NULL, NULL);
+            if (ret < 0)
+                return ret;
+
+            sws_setColorspaceDetails(slice, c->srcColorspaceTable,
+                                     slice->src_range, c->dstColorspaceTable,
+                                     slice->dst_range, c->brightness, c->contrast,
+                                     c->saturation);
+
+            for (int i = 0; i < FF_ARRAY_ELEMS(c->srcColorspaceTable); i++) {
+                c2->srcColorspaceTable[i] = c->srcColorspaceTable[i];
+                c2->dstColorspaceTable[i] = c->dstColorspaceTable[i];
+            }
+        }
+    }
+
+    if (c->dstXYZ && !(c->srcXYZ && unscaled)) {
+        SwsPass *sub = pass_add(graph, c, dst_w, dst_h, run_rgb2xyz, output, output, 1);
+        if (!sub)
+            return AVERROR(ENOMEM);
+    }
+
+    return 0;
+}
+
+/**
+ * Sentinel values to refer to the overall image input / output during
+ * filter graph construction, as the true values are not known.
+ */
+static uint8_t sws_input_sentinel, sws_output_sentinel;
+
+static const SwsImg *resolve_img(SwsGraph *graph, const SwsImg *img)
+{
+    if (img->data[0] == &sws_input_sentinel)
+        return &graph->exec.input;
+    else if (img->data[0] == &sws_output_sentinel)
+        return &graph->exec.output;
+    else
+        return img;
+}
+
+static int init_passes(SwsGraph *graph)
+{
+    SwsContext *const ctx      = graph->ctx;
+    const SwsFormat *const src = &graph->src;
+    const SwsFormat *const dst = &graph->dst;
+    const SwsDither dither     = ctx->dither;
+    SwsContext *sws;
+    int ret;
+
+    const SwsImg input  = { .data = { &sws_input_sentinel  }, .fmt = src->format };
+    const SwsImg output = { .data = { &sws_output_sentinel }, .fmt = dst->format };
+
+    graph->noop = ff_fmt_equal(dst, src);
+    if (graph->noop) {
+        /* Threaded memcpy pass */
+        SwsPass *copy = pass_add(graph, NULL, dst->width, dst->height,
+                                 run_copy, input, output, 1);
+        if (!copy)
+            return AVERROR(ENOMEM);
+        return 0;
+    }
+
+    graph->incomplete |= src->range == AVCOL_RANGE_UNSPECIFIED;
+    graph->incomplete |= dst->range == AVCOL_RANGE_UNSPECIFIED;
+
+    sws = graph->sws = sws_alloc_context();
+    if (!sws)
+        return AVERROR(ENOMEM);
+
+    sws->flags      = ctx->flags;
+    sws->dither     = dither;
+    sws->src_w      = src->width;
+    sws->src_h      = src->height;
+    sws->src_format = src->format;
+    sws->src_range  = src->range == AVCOL_RANGE_JPEG;
+
+    sws->dst_w      = dst->width;
+    sws->dst_h      = dst->height;
+    sws->dst_format = dst->format;
+    sws->dst_range  = dst->range == AVCOL_RANGE_JPEG;
+    get_chroma_pos(graph, &sws->src_h_chr_pos, &sws->src_v_chr_pos, src);
+    get_chroma_pos(graph, &sws->dst_h_chr_pos, &sws->dst_v_chr_pos, dst);
+
+    ret = sws_init_context(sws, NULL, NULL);
+    if (ret < 0)
+        return ret;
+
+    /* Set correct color matrices */
+    {
+        int in_full, out_full, brightness, contrast, saturation;
+        const int *inv_table, *table;
+        sws_getColorspaceDetails(sws, (int **)&inv_table, &in_full,
+                                (int **)&table, &out_full,
+                                &brightness, &contrast, &saturation);
+
+        inv_table = sws_getCoefficients(src->csp);
+        table     = sws_getCoefficients(dst->csp);
+
+        graph->incomplete |= src->csp != dst->csp &&
+                            (src->csp == AVCOL_SPC_UNSPECIFIED ||
+                             dst->csp == AVCOL_SPC_UNSPECIFIED);
+
+        sws_setColorspaceDetails(sws, inv_table, in_full, table, out_full,
+                                brightness, contrast, saturation);
+    }
+
+    ret = init_pass(graph, sws, input, output);
+    if (ret < 0)
+        return ret;
+
+    return 0;
+}
+
+static void sws_graph_worker(void *priv, int jobnr, int threadnr, int nb_jobs,
+                             int nb_threads)
+{
+    SwsGraph *graph = priv;
+    const SwsPass *pass = graph->exec.pass;
+    const SwsImg *input  = resolve_img(graph, &pass->input);
+    const SwsImg *output = resolve_img(graph, &pass->output);
+    const int slice_y = jobnr * pass->slice_h;
+    const int slice_h = FFMIN(pass->slice_h, pass->height - slice_y);
+
+    pass->run(output, input, slice_y, slice_h, pass);
+}
+
+int sws_graph_create(SwsContext *ctx,  const SwsFormat *dst, const SwsFormat *src,
+                     int field, SwsGraph **out_graph)
+{
+    int ret;
+    SwsGraph *graph = av_mallocz(sizeof(*graph));
+    if (!graph)
+        return AVERROR(ENOMEM);
+
+    graph->ctx = ctx;
+    graph->src = *src;
+    graph->dst = *dst;
+    graph->field = field;
+    graph->opts = *ctx;
+
+    graph->exec.input.fmt  = src->format;
+    graph->exec.output.fmt = dst->format;
+
+    ret = avpriv_slicethread_create(&graph->slicethread, (void *) graph,
+                                    sws_graph_worker, NULL, ctx->threads);
+    if (ret == AVERROR(ENOSYS))
+        graph->num_threads = 1;
+    else if (ret < 0)
+        goto error;
+    else
+        graph->num_threads = ret;
+
+    ret = init_passes(graph);
+    if (ret < 0)
+        goto error;
+
+    *out_graph = graph;
+    return 0;
+
+error:
+    sws_graph_free(&graph);
+    return ret;
+}
+
+void sws_graph_free(SwsGraph **pgraph)
+{
+    SwsGraph *graph = *pgraph;
+    if (!graph)
+        return;
+
+    avpriv_slicethread_free(&graph->slicethread);
+
+    for (int i = 0; i < graph->num_passes; i++) {
+        SwsPass *pass = graph->passes[i];
+        if (pass->uninit)
+            pass->uninit(pass);
+        av_free(pass->buf);
+        av_free(pass);
+    }
+    av_free(graph->passes);
+
+    sws_freeContext(graph->sws);
+    av_free(graph);
+    *pgraph = NULL;
+}
+
+/* Tests only options relevant to SwsGraph */
+static int opts_equal(const SwsContext *c1, const SwsContext *c2)
+{
+    return c1->flags       == c2->flags       &&
+           c1->threads     == c2->threads     &&
+           c1->dither      == c2->dither      &&
+           c1->alpha_blend == c2->alpha_blend &&
+           c1->gamma_flag  == c2->gamma_flag  &&
+           !memcmp(c1->scaler_params, c2->scaler_params, sizeof(c1->scaler_params));
+
+}
+
+int sws_graph_reinit(SwsContext *ctx, const SwsFormat *dst, const SwsFormat *src,
+                     int field, SwsGraph **out_graph)
+{
+    const SwsGraph *graph = *out_graph;
+    if (graph && ff_fmt_equal(&graph->src, src) &&
+                 ff_fmt_equal(&graph->dst, dst) &&
+                 opts_equal(ctx, &graph->opts))
+        return 0;
+
+    sws_graph_free(out_graph);
+    return sws_graph_create(ctx, dst, src, field, out_graph);
+}
+
+
+void sws_graph_run(SwsGraph *graph, uint8_t *const out_data[4],
+                   const int out_linesize[4],
+                   const uint8_t *const in_data[4],
+                   const int in_linesize[4])
+{
+    SwsImg *out = &graph->exec.output;
+    SwsImg *in  = &graph->exec.input;
+    memcpy(out->data,     out_data,     sizeof(out->data));
+    memcpy(out->linesize, out_linesize, sizeof(out->linesize));
+    memcpy(in->data,      in_data,      sizeof(in->data));
+    memcpy(in->linesize,  in_linesize,  sizeof(in->linesize));
+
+    for (int i = 0; i < graph->num_passes; i++) {
+        const SwsPass *pass = graph->passes[i];
+        graph->exec.pass = pass;
+        if (pass->setup)
+            pass->setup(out, in, pass);
+        avpriv_slicethread_execute(graph->slicethread, pass->num_slices, 0);
+    }
+}
diff --git a/libswscale/graph.h b/libswscale/graph.h
new file mode 100644
index 0000000000..8b4544973e
--- /dev/null
+++ b/libswscale/graph.h
@@ -0,0 +1,122 @@ 
+/*
+ * Copyright (C) 2024 Niklas Haas
+ *
+ * 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 SWSCALE_GRAPH_H
+#define SWSCALE_GRAPH_H
+
+#include "libavutil/slicethread.h"
+#include "swscale.h"
+#include "utils.h"
+
+/* Represents a view into a single field of frame data */
+typedef struct SwsImg {
+    enum AVPixelFormat fmt;
+    uint8_t *data[4]; /* points to y=0 */
+    int linesize[4];
+} SwsImg;
+
+typedef struct SwsPass  SwsPass;
+typedef struct SwsGraph SwsGraph;
+
+/**
+ * Output `h` lines of filtered data. `out` and `in` point to the
+ * start of the image buffer for this pass.
+ */
+typedef void (*sws_filter_run_t)(const SwsImg *out, const SwsImg *in,
+                                 int y, int h, const SwsPass *pass);
+
+struct SwsPass {
+    const SwsGraph *graph;
+    uint8_t *buf; /* temporary buffer for this pass, freed automatically */
+
+    sws_filter_run_t run;
+    int width, height; /* new output size */
+    int pixel_bytes;   /* bytes per pixel */
+    int slice_h;       /* filter granularity */
+    int num_slices;
+
+    /* Filter input/output. */
+    SwsImg input;
+    SwsImg output;
+
+    /**
+     * Called once from the main thread before running the filter. Optional.
+     * `out` and `in` always point to the main image input/output, regardless
+     * of `input` and `output` fields.
+     */
+    void (*setup)(const SwsImg *out, const SwsImg *in, const SwsPass *pass);
+
+    void (*uninit)(const SwsPass *pass); /* optional */
+    void *priv;
+};
+
+/* Filter graph, which represents a 'baked' pixel format conversion */
+typedef struct SwsGraph {
+    SwsContext *ctx;
+    AVSliceThread *slicethread;
+    int num_threads; /* resolved at init() time */
+    int incomplete;  /* set during init() if formats had to be inferred */
+    SwsContext *sws; /* wrapped legacy context */
+
+    /* Sorted sequence of filter passes to apply */
+    SwsPass **passes;
+    int num_passes;
+
+    /* Overall image parameters and flags */
+    SwsContext opts;
+    SwsFormat src, dst;
+    int field;
+    int noop; /* true if the graph is a no-op */
+
+    /* Temporary execution state inside sws_graph_run */
+    struct {
+        const SwsPass *pass; /* current filter pass */
+        SwsImg input;
+        SwsImg output;
+    } exec;
+} SwsGraph;
+
+/**
+ * Allocate and initialize the filter graph. Returns 0 or a negative error.
+ */
+int sws_graph_create(SwsContext *ctx, const SwsFormat *dst, const SwsFormat *src,
+                     int field, SwsGraph **out_graph);
+
+/**
+ * Uninitialize any state associate with this filter graph and free it.
+ */
+void sws_graph_free(SwsGraph **graph);
+
+/**
+ * Wrapper around sws_graph_create that does nothing if the format is
+ * unchanged.
+ */
+int sws_graph_reinit(SwsContext *ctx, const SwsFormat *dst, const SwsFormat *src,
+                     int field, SwsGraph **graph);
+
+/**
+ * Dispatch the filter graph on a single field. Internally threaded.
+ */
+void sws_graph_run(SwsGraph *graph, uint8_t *const out_data[4],
+                   const int out_linesize[4],
+                   const uint8_t *const in_data[4],
+                   const int in_linesize[4]);
+
+#endif /* SWSCALE_GRAPH_H */