/*

 * 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/cpu.h"

#include "libavutil/error.h"

#include "libavutil/hwcontext.h"

#include "libavutil/imgutils.h"

#include "libavutil/macros.h"

#include "libavutil/mem.h"

#include "libavutil/opt.h"

#include "libavutil/pixdesc.h"

#include "libavutil/refstruct.h"

#include "libavutil/slicethread.h"

#include "libswscale/swscale.h"

#include "libswscale/format.h"

#include "cms.h"

#include "lut3d.h"

#include "swscale_internal.h"

#include "graph.h"

#include "ops.h"

#include "ops_dispatch.h"

#if CONFIG_VULKAN

#include "vulkan/ops.h"

#endif

int ff_sws_pass_aligned_width(const SwsPass *pass, int width)

{

    if (!pass)

        return width;

    size_t aligned_w = width;

    aligned_w = FFALIGN(aligned_w, pass->output->width_align);

    aligned_w += pass->output->width_pad;

    return aligned_w <= INT_MAX ? aligned_w : width;

}

/* Allocates (or refs) one buffer per plane */

static int frame_alloc_planes_ref(AVFrame *dst, const AVFrame *src,

                                  const int plane_copy[4])

{

    int ret = av_image_check_size2(dst->width, dst->height, INT64_MAX,

                                   dst->format, 0, NULL);

    if (ret < 0)

        return ret;

    const int align = av_cpu_max_align();

    const int aligned_w = FFALIGN(dst->width + 1, align); /* add space for over-write */

    ret = av_image_fill_linesizes(dst->linesize, dst->format, aligned_w);

    if (ret < 0)

        return ret;

    ptrdiff_t linesize1[4];

    for (int i = 0; i < 4; i++)

        linesize1[i] = dst->linesize[i] = FFALIGN(dst->linesize[i], align);

    size_t sizes[4];

    ret = av_image_fill_plane_sizes(sizes, dst->format, dst->height, linesize1);

    if (ret < 0)

        return ret;

    for (int i = 0; i < 4; i++) {

        if (!sizes[i])

            break;

        int src_idx = plane_copy[i];

        if (src_idx >= 0 && src) {

            /* Ref the source plane instead of allocating a new buffer */

            dst->buf[i] = av_buffer_ref(src->buf[src_idx]);

            if (!dst->buf[i])

                return AVERROR(ENOMEM);

            dst->data[i]     = src->data[src_idx];

            dst->linesize[i] = src->linesize[src_idx];

            continue;

        }

        AVBufferRef *buf = av_buffer_alloc(sizes[i]);

        if (!buf)

            return AVERROR(ENOMEM);

        dst->data[i] = buf->data;

        dst->buf[i] = buf;

    }

    return 0;

}

#if CONFIG_VULKAN

static int pass_alloc_output_hw(SwsPass *pass, AVFrame *avframe,

                                AVBufferRef *dev_ref)

{

    SwsPassBuffer *buffer = pass->output;

    AVBufferRef *frames_ref = av_hwframe_ctx_alloc(dev_ref);

    if (!frames_ref)

        return AVERROR(ENOMEM);

    AVHWFramesContext *hwfc = (AVHWFramesContext *)frames_ref->data;

    hwfc->format    = AV_PIX_FMT_VULKAN;

    hwfc->sw_format = pass->format;

    hwfc->width     = buffer->width;

    hwfc->height    = buffer->height;

    int ret = av_hwframe_ctx_init(frames_ref);

    if (ret >= 0) {

        avframe->format = AV_PIX_FMT_VULKAN;

        ret = av_hwframe_get_buffer(frames_ref, avframe, 0);

    }

    av_buffer_unref(&frames_ref);

    return ret;

}

#endif

static int pass_alloc_output(SwsPass *pass)

{

    if (!pass || pass->output->avframe)

        return 0;

    SwsPassBuffer *buffer = pass->output;

    AVFrame *avframe = av_frame_alloc();

    if (!avframe)

        return AVERROR(ENOMEM);

    avframe->width  = buffer->width;

    avframe->height = buffer->height;

    int ret;

#if CONFIG_VULKAN

    const SwsGraph *graph = pass->graph;

    if (graph->src.hw_format == AV_PIX_FMT_VULKAN &&

        graph->dst.hw_format == AV_PIX_FMT_VULKAN) {

        AVBufferRef *dev_ref = ff_sws_vk_device_ref(graph->ctx);

        if (dev_ref) {

            ret = pass_alloc_output_hw(pass, avframe, dev_ref);

            if (ret >= 0)

                goto done;

            av_frame_unref(avframe);

        }

    }

#endif

    const AVFrame *src = NULL;

    if (pass->input)

        src = pass->input->output->avframe;

    avframe->format = pass->format;

    ret = frame_alloc_planes_ref(avframe, src, buffer->plane_copy);

    if (ret < 0) {

        av_frame_free(&avframe);

        return ret;

    }

#if CONFIG_VULKAN

done:

#endif

    buffer->avframe = avframe;

    ff_sws_frame_from_avframe(&buffer->frame, avframe);

    return 0;

}

static void free_buffer(AVRefStructOpaque opaque, void *obj)

{

    SwsPassBuffer *buffer = obj;

    av_frame_free(&buffer->avframe);

}

static void pass_free(SwsPass *pass)

{

    if (pass->free)

        pass->free(pass->priv);

    av_refstruct_unref(&pass->output);

    av_free(pass);

}

int ff_sws_graph_add_pass(SwsGraph *graph, enum AVPixelFormat fmt,

                          int width, int height, SwsPass *input,

                          int lines, int align,

                          SwsPassFunc run, SwsPassSetup setup,

                          void *priv, void (*free_cb)(void *priv),

                          SwsPass **out_pass)

{

    int ret;

    SwsPass *pass = av_mallocz(sizeof(*pass));

    if (!pass) {

        if (free_cb)

            free_cb(priv);

        return AVERROR(ENOMEM);

    }

    if (!lines)

        lines = height;

    pass->graph  = graph;

    pass->run    = run;

    pass->setup  = setup;

    pass->priv   = priv;

    pass->free   = free_cb;

    pass->format = fmt;

    pass->lines  = lines;

    pass->input  = input;

    pass->output = av_refstruct_alloc_ext(sizeof(*pass->output), 0, NULL, free_buffer);

    if (!pass->output) {

        ret = AVERROR(ENOMEM);

        goto fail;

    }

    pass->output->height = height;

    pass->output->width  = width;

    pass->output->width_align = 1;

    memset(pass->output->plane_copy, -1, sizeof(pass->output->plane_copy));

    if (!align) {

        pass->slice_h = pass->lines;

        pass->num_slices = 1;

    } else {

        pass->slice_h = (pass->lines + graph->num_threads - 1) / graph->num_threads;

        pass->slice_h = FFALIGN(pass->slice_h, align);

        pass->num_slices = (pass->lines + pass->slice_h - 1) / pass->slice_h;

    }

    ret = av_dynarray_add_nofree(&graph->passes, &graph->num_passes, pass);

    if (ret < 0)

        goto fail;

    *out_pass = pass;

    return 0;

fail:

    pass_free(pass);

    return ret;

}

void ff_sws_pass_link_output(SwsPass *dst, const SwsPass *src)

{

    if (!dst || !src || dst == src)

        return;

    av_assert0(dst->format == src->format);

    SwsPassBuffer *keep = src->output, *drop = dst->output;

    av_assert1(keep->width  == drop->width);

    av_assert1(keep->height == drop->height);

    keep->width_align = FFMAX(keep->width_align, drop->width_align);

    keep->width_pad   = FFMAX(keep->width_pad,   drop->width_pad);

    for (int i = 0; i < FF_ARRAY_ELEMS(keep->plane_copy); i++) {

        if (keep->plane_copy[i] < 0)

            keep->plane_copy[i] = drop->plane_copy[i];

        else if (drop->plane_copy[i] >= 0)

            av_assert1(keep->plane_copy[i] == drop->plane_copy[i]);

    }

    av_refstruct_replace(&dst->output, src->output);

}

static void frame_shift(const SwsFrame *f, const int y, uint8_t *data[4])

{

    for (int i = 0; i < 4; i++) {

        if (f->data[i])

            data[i] = f->data[i] + (y >> ff_fmt_vshift(f->format, i)) * f->linesize[i];

        else

            data[i] = NULL;

    }

}

static void run_copy(const SwsFrame *out, const SwsFrame *in, int y, int h,

                     const SwsPass *pass)

{

    uint8_t *in_data[4], *out_data[4];

    frame_shift(in,  y, in_data);

    frame_shift(out, y, out_data);

    for (int i = 0; i < 4 && out_data[i]; i++) {

        const int lines = h >> ff_fmt_vshift(in->format, i);

        av_assert1(in_data[i]);

        if (in_data[i] == out_data[i]) {

            av_assert0(in->linesize[i] == out->linesize[i]);

        } else 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 SwsFrame *out, const SwsFrame *in, int y, int h,

                     const SwsPass *pass)

{

    SwsInternal *c = pass->priv;

    const int x0 = c->src0Alpha - 1;

    const int w4 = 4 * out->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 SwsFrame *out, const SwsFrame *in, int y, int h,

                        const SwsPass *pass)

{

    const SwsInternal *c = pass->priv;

    c->xyz12Torgb48(c, out->data[0] + y * out->linesize[0], out->linesize[0],

                    in->data[0] + y * in->linesize[0], in->linesize[0],

                    out->width, h);

}

static void run_rgb2xyz(const SwsFrame *out, const SwsFrame *in, int y, int h,

                        const SwsPass *pass)

{

    const SwsInternal *c = pass->priv;

    c->rgb48Toxyz12(c, out->data[0] + y * out->linesize[0], out->linesize[0],

                    in->data[0] + y * in->linesize[0], in->linesize[0],

                    out->width, h);

}

/***********************************************************************

 * Internal ff_swscale() wrapper. This reuses the legacy scaling API. *

 * This is considered fully deprecated, and will be replaced by a full *

 * reimplementation ASAP.                                              *

 ***********************************************************************/

static void free_legacy_swscale(void *priv)

{

    SwsContext *sws = priv;

    sws_free_context(&sws);

}

static int setup_legacy_swscale(const SwsFrame *out, const SwsFrame *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]);

    return 0;

}

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_legacy_unscaled(const SwsFrame *out, const SwsFrame *in,

                                int y, int h, const SwsPass *pass)

{

    SwsContext *sws = slice_ctx(pass, y);

    SwsInternal *c = sws_internal(sws);

    uint8_t *in_data[4];

    frame_shift(in, y, in_data);

    c->convert_unscaled(c, (const uint8_t *const *) in_data, in->linesize, y, h,

                        out->data, out->linesize);

}

static void run_legacy_swscale(const SwsFrame *out, const SwsFrame *in,

                               int y, int h, const SwsPass *pass)

{

    SwsContext *sws = slice_ctx(pass, y);

    SwsInternal *c = sws_internal(sws);

    uint8_t *out_data[4];

    frame_shift(out, y, out_data);

    ff_swscale(c, (const uint8_t *const *) in->data, in->linesize, 0,

               sws->src_h, out_data, out->linesize, y, h);

}

static void run_legacy_lut3d(const SwsFrame *out, const SwsFrame *in,

                             int y, int h, const SwsPass *pass)

{

    const SwsLut3D *lut = pass->graph->lut3d;

    uint8_t *in_data[4], *out_data[4];

    frame_shift(in,  y, in_data);

    frame_shift(out, y, out_data);

    ff_sws_lut3d_apply_rgba64(lut, in_data[0], in->linesize[0], out_data[0],

                              out->linesize[0], out->width, h);

}

static void legacy_chr_pos(SwsGraph *graph, int *chr_pos, int override, int *warned)

{

    if (override == -513 || override == *chr_pos)

        return;

    if (!*warned) {

        av_log(NULL, AV_LOG_WARNING,

               "Setting chroma position directly is deprecated, make sure "

               "the frame is tagged with the correct chroma location.\n");

        *warned = 1;

    }

    *chr_pos = override;

}

/* Takes over ownership of `sws` */

static int init_legacy_subpass(SwsGraph *graph, SwsContext *sws,

                               SwsPass *input, SwsPass **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 = NULL;

    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++) {

            const int is_last = i + 1 == num_cascaded;

            /* Steal cascaded context, so we can manage its lifetime independently */

            SwsContext *sub = c->cascaded_context[i];

            c->cascaded_context[i] = NULL;

            ret = init_legacy_subpass(graph, sub, input, is_last ? output : &input);

            if (ret < 0)

                break;

        }

        sws_free_context(&sws);

        return ret;

    }

    if (sws->dither == SWS_DITHER_ED && !c->convert_unscaled)

        align = 0; /* disable slice threading */

    if (c->src0Alpha && !c->dst0Alpha && isALPHA(sws->dst_format)) {

        ret = ff_sws_graph_add_pass(graph, AV_PIX_FMT_RGBA, src_w, src_h, input,

                                    0, 1, run_rgb0, NULL, c, NULL, &input);

        if (ret < 0) {

            sws_free_context(&sws);

            return ret;

        }

    }

    if (c->srcXYZ && !(c->dstXYZ && unscaled)) {

        ret = ff_sws_graph_add_pass(graph, AV_PIX_FMT_RGB48, src_w, src_h, input,

                                    0, 1, run_xyz2rgb, NULL, c, NULL, &input);

        if (ret < 0) {

            sws_free_context(&sws);

            return ret;

        }

    }

    ret = ff_sws_graph_add_pass(graph, sws->dst_format, dst_w, dst_h, input, 0, align,

                                c->convert_unscaled ? run_legacy_unscaled : run_legacy_swscale,

                                setup_legacy_swscale, sws, free_legacy_swscale, &pass);

    if (ret < 0)

        return ret;

    pass->backend = SWS_BACKEND_LEGACY;

    /**

     * 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 = ff_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)) {

        ret = ff_sws_graph_add_pass(graph, AV_PIX_FMT_RGB48, dst_w, dst_h, pass,

                                    0, 1, run_rgb2xyz, NULL, c, NULL, &pass);

        if (ret < 0)

            return ret;

    }

    *output = pass;

    return 0;

}

static int add_legacy_3dlut_pass(SwsGraph *graph, const SwsFormat *src,

                                 SwsPass *input, SwsPass **output);

static int add_legacy_sws_pass(SwsGraph *graph, const SwsFormat *src,

                               const SwsFormat *dst, const SwsLut3D *lut3d,

                               SwsPass *input, SwsPass **output)

{

    int ret, warned = 0;

    SwsContext *const ctx = graph->ctx;

    const SwsBackend backend = ff_sws_enabled_backends(ctx);

    if (!(backend & SWS_BACKEND_LEGACY))

        return AVERROR(ENOTSUP);

    if (src->hw_format != AV_PIX_FMT_NONE || dst->hw_format != AV_PIX_FMT_NONE)

        return AVERROR(ENOTSUP);

    /* Re-check this here because this might not be excluded if the caller was

     * testing against multiple backends */

    if (!sws_isSupportedInput(src->format) || !sws_isSupportedOutput(dst->format))

        return AVERROR(ENOTSUP);

    /* If we need to apply a 3D LUT, add it as an explicit input prepass */

    if (lut3d) {

        ret = add_legacy_3dlut_pass(graph, src, input, &input);

        if (ret < 0)

            return ret;

        SwsFormat tmp = *src;

        tmp.format = input->format;

        tmp.color  = lut3d->map.dst;

        return add_legacy_sws_pass(graph, &tmp, dst, NULL, input, output);

    }

    SwsContext *sws = sws_alloc_context();

    if (!sws)

        return AVERROR(ENOMEM);

    sws->flags       = ctx->flags;

    sws->dither      = ctx->dither;

    sws->alpha_blend = ctx->alpha_blend;

    sws->gamma_flag  = ctx->gamma_flag;

    sws->scaler      = ctx->scaler;

    sws->scaler_sub  = ctx->scaler_sub;

    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;

    ff_sws_chroma_pos(src, &graph->incomplete, &sws->src_h_chr_pos, &sws->src_v_chr_pos);

    ff_sws_chroma_pos(dst, &graph->incomplete, &sws->dst_h_chr_pos, &sws->dst_v_chr_pos);

    graph->incomplete |= src->range == AVCOL_RANGE_UNSPECIFIED;

    graph->incomplete |= dst->range == AVCOL_RANGE_UNSPECIFIED;

    /* Allow overriding chroma position with the legacy API */

    legacy_chr_pos(graph, &sws->src_h_chr_pos, ctx->src_h_chr_pos, &warned);

    legacy_chr_pos(graph, &sws->src_v_chr_pos, ctx->src_v_chr_pos, &warned);

    legacy_chr_pos(graph, &sws->dst_h_chr_pos, ctx->dst_h_chr_pos, &warned);

    legacy_chr_pos(graph, &sws->dst_v_chr_pos, ctx->dst_v_chr_pos, &warned);

    /* Explicitly strip chroma offsets when not subsampling, because it

     * interferes with the operation of flags like SWS_FULL_CHR_H_INP */

    if (!src->desc->log2_chroma_w)

        sws->src_h_chr_pos = -513;

    if (!src->desc->log2_chroma_h)

        sws->src_v_chr_pos = -513;

    if (!dst->desc->log2_chroma_w)

        sws->dst_h_chr_pos = -513;

    if (!dst->desc->log2_chroma_h)

        sws->dst_v_chr_pos = -513;

    for (int i = 0; i < SWS_NUM_SCALER_PARAMS; i++)

        sws->scaler_params[i] = ctx->scaler_params[i];

    ret = sws_init_context(sws, NULL, NULL);

    if (ret < 0) {

        sws_free_context(&sws);

        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);

    }

    return init_legacy_subpass(graph, sws, input, output);

}

static int add_legacy_3dlut_pass(SwsGraph *graph, const SwsFormat *src,

                                 SwsPass *input, SwsPass **output)

{

    int ret;

    const SwsLut3D *lut3d = graph->lut3d;

    if (!lut3d)

        return 0;

    const enum AVPixelFormat fmt = AV_PIX_FMT_RGBA64;

    if (src->format != fmt) {

        SwsFormat tmp = *src;

        tmp.format = fmt;

        ret = add_legacy_sws_pass(graph, src, &tmp, NULL, input, &input);

        if (ret < 0)

            return ret;

    }

    ret = ff_sws_graph_add_pass(graph, fmt, src->width, src->height,

                                input, 0, 1, run_legacy_lut3d, NULL, NULL, NULL,

                                output);

    if (ret < 0)

        return ret;

    return 0;

}

/*********************************

 * Format conversion and scaling *

 *********************************/

static int add_ops_convert_pass(SwsGraph *graph, const SwsFormat *src,

                                const SwsFormat *dst, const SwsLut3D *lut3d,

                                SwsPass *input, SwsPass **output)

{

#if CONFIG_UNSTABLE

    SwsContext *ctx = graph->ctx;

    /* Preemptively skip the ops list generation if the backend was

     * constrained to the legacy implementation only. This would

     * normally also fail in ff_sws_compile_pass() with the same

     * error, but this way saves a bit of unnecessary overhead */

    const SwsBackend backends = ff_sws_enabled_backends(ctx);

    if (backends == SWS_BACKEND_LEGACY)

        return AVERROR(ENOTSUP);

    SwsOpList *ops;

    int ret = ff_sws_op_list_generate(ctx, src, dst, lut3d, &ops, &graph->incomplete);

    if (ret < 0)

        return ret;

    av_log(ctx, AV_LOG_VERBOSE, "Conversion pass for %s -> %s:\n",

           av_get_pix_fmt_name(src->format), av_get_pix_fmt_name(dst->format));

    av_log(ctx, AV_LOG_DEBUG, "Unoptimized operation list:\n");

    ff_sws_op_list_print(ctx, AV_LOG_DEBUG, AV_LOG_TRACE, ops);

    const int flags = SWS_OP_FLAG_OPTIMIZE | SWS_OP_FLAG_SPLIT_MEMCPY;

    return ff_sws_compile_pass(graph, NULL, &ops, flags, input, output);

#else

    return AVERROR(ENOTSUP);

#endif

}

static bool prefer_ops_backend(SwsContext *ctx, const SwsFormat *src, const SwsFormat *dst)

{

    if (ctx->flags & SWS_UNSTABLE)

        return true;

    if (isFloat(src->format) || isFloat(dst->format))

        return true; /* ops backend has better support for float formats */

    return false; /* default to legacy for stability reasons */

}

static int add_convert_pass(SwsGraph *graph, const SwsFormat *src,

                            const SwsFormat *dst, const SwsLut3D *lut3d,

                            SwsPass *input, SwsPass **output)

{

    SwsContext *ctx = graph->ctx;

    int ret;

    if (prefer_ops_backend(ctx, src, dst)) {

        ret = add_ops_convert_pass(graph, src, dst, lut3d, input, output);

        if (ret == AVERROR(ENOTSUP))

            ret = add_legacy_sws_pass(graph, src, dst, lut3d, input, output);

    } else {

        ret = add_legacy_sws_pass(graph, src, dst, lut3d, input, output);

        if (ret == AVERROR(ENOTSUP))

            ret = add_ops_convert_pass(graph, src, dst, lut3d, input, output);

    }

    return ret;

}

/**************************

 * Gamut and tone mapping *

 **************************/

static int generate_3dlut(SwsGraph *graph, SwsFormat *src, SwsFormat *dst)

{

    SwsColorMap map = {0};

    /**

     * Grayspace does not really have primaries, so just force the use of

     * the equivalent other primary set to avoid a conversion. Technically,

     * this does affect the weights used for the Grayscale conversion, but

     * in practise, that should give the expected results more often than not.

     */

    if (isGray(dst->format)) {

        dst->color = src->color;

    } else if (isGray(src->format)) {

        src->color = dst->color;

    }

    /* Fully infer color spaces before color mapping logic */

    graph->incomplete |= ff_infer_colors(&src->color, &dst->color);

    map.intent = graph->ctx->intent;

    map.src    = src->color;

    map.dst    = dst->color;

    if (ff_sws_color_map_noop(&map))

        return 0;

    if (src->hw_format != AV_PIX_FMT_NONE || dst->hw_format != AV_PIX_FMT_NONE)

        return AVERROR(ENOTSUP);

    graph->lut3d = ff_sws_lut3d_alloc();

    if (!graph->lut3d)

        return AVERROR(ENOMEM);

    return ff_sws_lut3d_generate(graph->lut3d, &map);

}

/***************************************

 * Main filter graph construction code *

 ***************************************/

static int init_passes(SwsGraph *graph)

{

    SwsFormat src = graph->src;

    SwsFormat dst = graph->dst;

    SwsPass *pass = NULL; /* read from main input image */

    int ret;

    ret = generate_3dlut(graph, &src, &dst);

    if (ret < 0)

        return ret;

    if (!ff_fmt_equal(&src, &dst) || graph->lut3d) {

        ret = add_convert_pass(graph, &src, &dst, graph->lut3d, pass, &pass);

        if (ret < 0)

            return ret;

    }

    if (!pass) {

        /* No passes were added, so no operations were necessary */

        graph->noop = 1;

        const int nb_planes = av_pix_fmt_count_planes(dst.format);

        for (int i = 0; i < nb_planes; i++)

            graph->plane_copy[i] = i;

        /* Add threaded memcpy pass */

        return ff_sws_graph_add_pass(graph, dst.format, dst.width, dst.height,

                                    pass, 0, 1, run_copy, NULL, NULL, NULL, &pass);

    }

    /* Compute end-to-end plane copy map */

    for (int n = 0; n < graph->num_passes; n++) {

        const SwsPass *pass = graph->passes[n];

        /* This pass writes to an output buffer other than the image

         * output, or copies from the output of a different pass */

        if (pass->output->avframe || pass->input)

            continue;

        for (int i = 0; i < FF_ARRAY_ELEMS(graph->plane_copy); i++) {

            const int idx = pass->output->plane_copy[i];

            if (idx < 0)

                continue;

            if (graph->plane_copy[i] < 0) {

                graph->plane_copy[i] = idx;

                av_log(graph->ctx, AV_LOG_DEBUG, "Plane %d passthrough from "

                       "plane %d\n", i, idx);

            } else {

                av_assert0(graph->plane_copy[i] == idx);

            }

        }

    }

    return 0;

}

static int 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 int slice_y = jobnr * pass->slice_h;

    const int slice_h = FFMIN(pass->slice_h, pass->lines - slice_y);

    pass->run(graph->exec.output, graph->exec.input, slice_y, slice_h, pass);

    return 0;

}

SwsGraph *ff_sws_graph_alloc(void)

{

    return av_mallocz(sizeof(SwsGraph));

}

static void graph_uninit(SwsGraph *graph)

{

    avpriv_slicethread_free(&graph->slicethread);

    for (int i = 0; i < graph->num_passes; i++)

        pass_free(graph->passes[i]);

    av_free(graph->passes);

    av_refstruct_unref(&graph->lut3d);

    memset(graph, 0, sizeof(*graph));

}

int ff_sws_graph_init(SwsGraph *graph, SwsContext *ctx, const SwsFormat *dst,

                      const SwsFormat *src)