/*

 * 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/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"

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 one buffer per plane */

static int frame_alloc_planes(AVFrame *dst)

{

    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, align);

    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;

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

        if (!buf)

            return AVERROR(ENOMEM);

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

        dst->buf[i] = buf;

    }

    return 0;

}

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->format = pass->format;

    avframe->width  = buffer->width;

    avframe->height = buffer->height;

    int ret = frame_alloc_planes(avframe);

    if (ret < 0) {

        av_frame_free(&avframe);

        return ret;

    }

    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 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);

    }

    pass->graph  = graph;

    pass->run    = run;

    pass->setup  = setup;

    pass->priv   = priv;

    pass->free   = free_cb;

    pass->format = fmt;

    pass->width  = width;

    pass->height = height;

    pass->input  = input;

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

    if (!pass->output) {

        ret = AVERROR(ENOMEM);

        goto fail;

    }

    if (!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, align);

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

    }

    /* Align output buffer to include extra slice padding */

    pass->output->height = pass->slice_h * pass->num_slices;

    pass->output->width  = pass->width;

    pass->output->width_align = 1;

    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;

}

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 * 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 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],

                    pass->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],

                    pass->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 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 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,

                                    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,

                                    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, align,

                                c->convert_unscaled ? run_legacy_unscaled : run_legacy_swscale,

                                setup_legacy_swscale, sws, free_legacy_swscale, &pass);

    if (ret < 0)

        return ret;

    /**

     * 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,

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

        if (ret < 0)

            return ret;

    }

    *output = pass;

    return 0;

}

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

                               const SwsFormat *dst, SwsPass *input,

                               SwsPass **output)

{

    int ret, warned = 0;

    SwsContext *const ctx = graph->ctx;

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

        return AVERROR(ENOTSUP);

    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;

    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);

    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);

    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);

}

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

 * Format conversion and scaling *

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

#if CONFIG_UNSTABLE

static SwsScaler get_scaler_fallback(SwsContext *ctx)

{

    if (ctx->scaler != SWS_SCALE_AUTO)

        return ctx->scaler;

    /* Backwards compatibility with legacy flags API */

    if (ctx->flags & SWS_BILINEAR) {

        return SWS_SCALE_BILINEAR;

    } else if (ctx->flags & (SWS_BICUBIC | SWS_BICUBLIN)) {

        return SWS_SCALE_BICUBIC;

    } else if (ctx->flags & SWS_POINT) {

        return SWS_SCALE_POINT;

    } else if (ctx->flags & SWS_AREA) {

        return SWS_SCALE_AREA;

    } else if (ctx->flags & SWS_GAUSS) {

        return SWS_SCALE_GAUSSIAN;

    } else if (ctx->flags & SWS_SINC) {

        return SWS_SCALE_SINC;

    } else if (ctx->flags & SWS_LANCZOS) {

        return SWS_SCALE_LANCZOS;

    } else if (ctx->flags & SWS_SPLINE) {

        return SWS_SCALE_SPLINE;

    } else {

        return SWS_SCALE_AUTO;

    }

}

static int add_filter(SwsContext *ctx, SwsPixelType type, SwsOpList *ops,

                      SwsOpType filter, int src_size, int dst_size)

{

    if (src_size == dst_size)

        return 0; /* no-op */

    SwsFilterParams params = {

        .scaler   = get_scaler_fallback(ctx),

        .src_size = src_size,

        .dst_size = dst_size,

    };

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

        params.scaler_params[i] = ctx->scaler_params[i];

    SwsFilterWeights *kernel;

    int ret = ff_sws_filter_generate(ctx, &params, &kernel);

    if (ret == AVERROR(ENOTSUP)) {

        /* Filter size exceeds limit; cascade with geometric mean size */

        int mean = sqrt((int64_t) src_size * dst_size);

        if (mean == src_size || mean == dst_size)

            return AVERROR_BUG; /* sanity, prevent infinite loop */

        ret = add_filter(ctx, type, ops, filter, src_size, mean);

        if (ret < 0)

            return ret;

        return add_filter(ctx, type, ops, filter, mean, dst_size);

    } else if (ret < 0) {

        return ret;

    }

    return ff_sws_op_list_append(ops, &(SwsOp) {

        .type = type,

        .op   = filter,

        .filter.kernel = kernel,

    });

}

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

                            const SwsFormat *dst, SwsPass *input,

                            SwsPass **output)

{

    const SwsPixelType type = SWS_PIXEL_F32;

    SwsContext *ctx = graph->ctx;

    SwsOpList *ops = NULL;

    int ret = AVERROR(ENOTSUP);

    /* Mark the entire new ops infrastructure as experimental for now */

    if (!(ctx->flags & SWS_UNSTABLE))

        goto fail;

    /* The new code does not yet support alpha blending */

    if (src->desc->flags & AV_PIX_FMT_FLAG_ALPHA &&

        ctx->alpha_blend != SWS_ALPHA_BLEND_NONE)

        goto fail;

    ops = ff_sws_op_list_alloc();

    if (!ops)

        return AVERROR(ENOMEM);

    ops->src = *src;

    ops->dst = *dst;

    ret = ff_sws_decode_pixfmt(ops, src->format);

    if (ret < 0)

        goto fail;

    ret = ff_sws_decode_colors(ctx, type, ops, src, &graph->incomplete);

    if (ret < 0)

        goto fail;

    /**

     * Always perform horizontal scaling first, since it's much more likely to

     * benefit from small integer optimizations; we should maybe flip the order

     * here if we're downscaling the vertical resolution by a lot, though.

     */

    ret = add_filter(ctx, type, ops, SWS_OP_FILTER_H, src->width, dst->width);

    if (ret < 0)

        goto fail;

    ret = add_filter(ctx, type, ops, SWS_OP_FILTER_V, src->height, dst->height);

    if (ret < 0)

        goto fail;

    ret = ff_sws_encode_colors(ctx, type, ops, src, dst, &graph->incomplete);

    if (ret < 0)

        goto fail;

    ret = ff_sws_encode_pixfmt(ops, dst->format);

    if (ret < 0)

        goto fail;

    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);

    ret = ff_sws_compile_pass(graph, &ops, SWS_OP_FLAG_OPTIMIZE, input, output);

    if (ret < 0)

        goto fail;

    ret = 0;

    /* fall through */

fail:

    ff_sws_op_list_free(&ops);

    if (ret == AVERROR(ENOTSUP))

        return add_legacy_sws_pass(graph, src, dst, input, output);

    return ret;

}

#else

#define add_convert_pass add_legacy_sws_pass

#endif

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

 * Gamut and tone mapping *

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

static void free_lut3d(void *priv)

{

    SwsLut3D *lut = priv;

    ff_sws_lut3d_free(&lut);

}

static int setup_lut3d(const SwsFrame *out, const SwsFrame *in, const SwsPass *pass)

{

    SwsLut3D *lut = pass->priv;

    /* Update dynamic frame metadata from the original source frame */

    ff_sws_lut3d_update(lut, &pass->graph->src.color);

    return 0;

}

static void run_lut3d(const SwsFrame *out, const SwsFrame *in, int y, int h,

                      const SwsPass *pass)

{

    SwsLut3D *lut = pass->priv;

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

    frame_shift(in,  y, in_data);

    frame_shift(out, y, out_data);

    ff_sws_lut3d_apply(lut, in_data[0], in->linesize[0], out_data[0],

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

}

static int adapt_colors(SwsGraph *graph, SwsFormat src, SwsFormat dst,

                        SwsPass *input, SwsPass **output)

{

    enum AVPixelFormat fmt_in, fmt_out;

    SwsColorMap map = {0};

    SwsLut3D *lut;

    int ret;

    /**

     * 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);

    lut = ff_sws_lut3d_alloc();

    if (!lut)

        return AVERROR(ENOMEM);

    fmt_in  = ff_sws_lut3d_pick_pixfmt(src, 0);

    fmt_out = ff_sws_lut3d_pick_pixfmt(dst, 1);

    if (fmt_in != src.format) {

        SwsFormat tmp = src;

        tmp.format = fmt_in;

        ret = add_convert_pass(graph, &src, &tmp, input, &input);

        if (ret < 0) {

            ff_sws_lut3d_free(&lut);

            return ret;

        }

    }

    ret = ff_sws_lut3d_generate(lut, fmt_in, fmt_out, &map);

    if (ret < 0) {

        ff_sws_lut3d_free(&lut);

        return ret;

    }

    return ff_sws_graph_add_pass(graph, fmt_out, src.width, src.height,

                                 input, 1, run_lut3d, setup_lut3d, lut,

                                 free_lut3d, output);

}

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

 * 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 = adapt_colors(graph, src, dst, pass, &pass);

    if (ret < 0)

        return ret;

    src.format = pass ? pass->format : src.format;

    src.color  = dst.color;

    if (!ff_fmt_equal(&src, &dst)) {

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

        if (ret < 0)

            return ret;

    }

    if (pass)

        return 0;

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

    graph->noop = 1;

    /* Add threaded memcpy pass */

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

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

}

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 int slice_y = jobnr * pass->slice_h;

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

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

}

int ff_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_copy = *ctx;

    if (ctx->threads == 1) {

        graph->num_threads = 1;

    } else {

        ret = avpriv_slicethread_create(&graph->slicethread, (void *) graph,

                                        sws_graph_worker, NULL, ctx->threads);

        if (ret == AVERROR(ENOSYS)) {

            /* Fall back to single threaded operation */

            graph->num_threads = 1;

        } else if (ret < 0) {

            goto error;

        } else {

            graph->num_threads = ret;

        }

    }

    ret = init_passes(graph);