libswscale/ops_dispatch.c

/**
 * Copyright (C) 2025 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/mathematics.h"
#include "libavutil/mem.h"
#include "libavutil/mem_internal.h"
#include "libavutil/refstruct.h"

#include "ops.h"
#include "ops_internal.h"
#include "ops_dispatch.h"
#include "swscale_internal.h"

#define RET(x)                                                                 \
    do {                                                                       \
        if ((ret = (x)) < 0)                                                   \
            goto fail;                                                         \
    } while (0)

typedef struct SwsOpPass {
    SwsCompiledOp comp;
    SwsOpExec exec_base;
    SwsOpExec exec_tail;
    size_t num_blocks;
    int tail_off_in;
    int tail_off_out;
    int tail_size_in;
    int tail_size_out;
    int planes_in;
    int planes_out;
    int pixel_bits_in;
    int pixel_bits_out;
    int idx_in[4];
    int idx_out[4];
    int palette_idx;
    int *offsets_y;
    int filter_size_h;
    bool memcpy_first;
    bool memcpy_last;
    bool memcpy_out;
    size_t tail_blocks;
    uint8_t *tail_buf; /* extra memory for fixing unpadded tails */
    unsigned int tail_buf_size;
} SwsOpPass;

static int compile_backend(SwsContext *ctx, const SwsOpBackend *backend,
                           const SwsOpList *ops, SwsCompiledOp *out)
{
    SwsOpList *copy;
    SwsCompiledOp compiled = {0};
    int ret = 0;

    copy = ff_sws_op_list_duplicate(ops);
    if (!copy)
        return AVERROR(ENOMEM);

    /* Ensure these are always set during compilation */
    ff_sws_op_list_update_comps(copy);

    ret = backend->compile(ctx, copy, &compiled);
    if (ret < 0) {
        int msg_lev = ret == AVERROR(ENOTSUP) ? AV_LOG_TRACE : AV_LOG_ERROR;
        av_log(ctx, msg_lev, "Backend '%s' failed to compile operations: %s\n",
               backend->name, av_err2str(ret));
        goto fail;
    }

    compiled.backend = backend;
    *out = compiled;

    av_log(ctx, AV_LOG_VERBOSE, "Compiled using backend '%s': "
           "block size = %d, over-read = {%d %d %d %d}, over-write = {%d %d %d %d}, "
           "cpu flags = 0x%x\n", backend->name, out->block_size,
           out->over_read[0], out->over_read[1],
           out->over_read[2], out->over_read[3],
           out->over_write[0], out->over_write[1],
           out->over_write[2], out->over_write[3],
           out->cpu_flags);

    ff_sws_op_list_print(ctx, AV_LOG_VERBOSE, AV_LOG_TRACE, ops);

fail:
    ff_sws_op_list_free(&copy);
    return ret;
}

int ff_sws_ops_compile(SwsContext *ctx, const SwsOpBackend *backend,
                       const SwsOpList *ops, SwsCompiledOp *out)
{
    if (backend)
        return compile_backend(ctx, backend, ops, out);

    const SwsBackend enabled = ff_sws_enabled_backends(ctx);
    for (int n = 0; ff_sws_op_backends[n]; n++) {
        const SwsOpBackend *backend = ff_sws_op_backends[n];
        if (ops->src.hw_format != backend->hw_format ||
            ops->dst.hw_format != backend->hw_format ||
            !(enabled & backend->flags))
            continue;
        if (compile_backend(ctx, backend, ops, out) < 0)
            continue;

        return 0;
    }

    return AVERROR(ENOTSUP);
}

void ff_sws_compiled_op_unref(SwsCompiledOp *comp)
{
    if (comp->free)
        comp->free(comp->priv);

    *comp = (SwsCompiledOp) {0};
}

static void op_pass_free(void *ptr)
{
    SwsOpPass *p = ptr;
    if (!p)
        return;

    ff_sws_compiled_op_unref(&p->comp);
    av_refstruct_unref(&p->offsets_y);
    av_free(p->exec_base.in_bump_y);
    av_free(p->exec_base.in_offset_x);
    av_free(p->tail_buf);
    av_free(p);
}

static inline void get_row_data(const SwsOpPass *p, const int y_dst,
                                const uint8_t *in[4], uint8_t *out[4])
{
    const SwsOpExec *base = &p->exec_base;
    const int y_src = p->offsets_y ? p->offsets_y[y_dst] : y_dst;
    for (int i = 0; i < p->planes_in; i++)
        in[i] = base->in[i] + (y_src >> base->in_sub_y[i]) * base->in_stride[i];
    for (int i = 0; i < p->planes_out; i++)
        out[i] = base->out[i] + (y_dst >> base->out_sub_y[i]) * base->out_stride[i];
}

static inline int get_lines_in(const SwsOpPass *p, const int y, const int h,
                               const int plane)
{
    const SwsOpExec *base = &p->exec_base;
    if (!p->offsets_y)
        return h >> base->in_sub_y[plane];

    const int y0 = p->offsets_y[y] >> base->in_sub_y[plane];
    const int y1 = p->offsets_y[y + h - 1] >> base->in_sub_y[plane];
    return y1 - y0 + 1;
}

static inline size_t pixel_bytes(size_t pixels, int pixel_bits,
                                 enum AVRounding rounding)
{
    const uint64_t bits = (uint64_t) pixels * pixel_bits;
    switch (rounding) {
    case AV_ROUND_ZERO:
    case AV_ROUND_DOWN:
        return bits >> 3;
    case AV_ROUND_INF:
    case AV_ROUND_UP:
        return (bits + 7) >> 3;
    default:
        av_unreachable("Invalid rounding mode");
        return (size_t) -1;
    }
}

static size_t safe_bytes_pad(int linesize, int plane_pad)
{
    av_assert1(linesize);
    int64_t safe_bytes = FFABS((int64_t) linesize) - plane_pad;
    return FFMAX(safe_bytes, 0);
}

static size_t safe_blocks_offset(size_t num_blocks, unsigned block_size,
                                 ptrdiff_t safe_offset,
                                 const int32_t *offset_bytes)
{
    size_t safe_blocks = num_blocks;
    while (safe_blocks && offset_bytes[safe_blocks * block_size - 1] > safe_offset)
        safe_blocks--;
    return safe_blocks;
}

static int op_pass_setup(const SwsFrame *out, const SwsFrame *in,
                         const SwsPass *pass)
{
    const AVPixFmtDescriptor *indesc  = av_pix_fmt_desc_get(in->format);
    const bool float_in = indesc->flags & AV_PIX_FMT_FLAG_FLOAT;
    const int width = out->width;

    SwsOpPass *p = pass->priv;
    SwsOpExec *exec = &p->exec_base;
    const SwsCompiledOp *comp = &p->comp;

    /* Set up main loop parameters */
    const unsigned block_size = comp->block_size;
    const size_t num_blocks   = (width + block_size - 1) / block_size;
    const size_t aligned_w    = num_blocks * block_size;
    if (aligned_w < width) /* overflow */
        return AVERROR(EINVAL);
    p->num_blocks   = num_blocks;
    p->memcpy_first = false;
    p->memcpy_last  = false;
    p->memcpy_out   = false;

    size_t safe_blocks = num_blocks;
    for (int i = 0; i < p->planes_in; i++) {
        const int idx = p->idx_in[i];
        size_t input_bytes = in->linesize[idx];
        if (p->filter_size_h && float_in) {
            /* Floating point inputs may contain NaN / Infinity in the padding */
            const int plane_w = AV_CEIL_RSHIFT(in->width, exec->in_sub_x[i]);
            input_bytes = pixel_bytes(plane_w, p->pixel_bits_in, AV_ROUND_UP);
        }

        size_t safe_bytes = safe_bytes_pad(input_bytes, comp->over_read[i]);
        size_t safe_blocks_in;
        if (exec->in_offset_x) {
            size_t filter_size = pixel_bytes(p->filter_size_h, p->pixel_bits_in,
                                             AV_ROUND_UP);
            safe_blocks_in = safe_blocks_offset(num_blocks, block_size,
                                                safe_bytes - filter_size,
                                                exec->in_offset_x);
        } else {
            safe_blocks_in = safe_bytes / exec->block_size_in[i];
        }

        if (safe_blocks_in < num_blocks) {
            p->memcpy_first |= in->linesize[idx] < 0;
            p->memcpy_last  |= in->linesize[idx] > 0;
            safe_blocks = FFMIN(safe_blocks, safe_blocks_in);
        }

        size_t loop_size   = num_blocks * exec->block_size_in[i];
        exec->in[i]        = in->data[idx];
        exec->in_stride[i] = in->linesize[idx];
        exec->in_bump[i]   = in->linesize[idx] - loop_size;
    }

    for (int i = 0; i < p->planes_out; i++) {
        const int idx = p->idx_out[i];
        size_t safe_bytes = safe_bytes_pad(out->linesize[idx], comp->over_write[i]);
        size_t safe_blocks_out = safe_bytes / exec->block_size_out[i];
        if (safe_blocks_out < num_blocks) {
            p->memcpy_out = true;
            safe_blocks   = FFMIN(safe_blocks, safe_blocks_out);
        }

        size_t loop_size    = num_blocks * exec->block_size_out[i];
        exec->out[i]        = out->data[idx];
        exec->out_stride[i] = out->linesize[idx];
        exec->out_bump[i]   = out->linesize[idx] - loop_size;
    }

    if (p->palette_idx >= 0) {
        exec->in[1] = in->data[p->palette_idx];
        exec->in_stride[1] = exec->in_bump[1] = 0;
    }

    const bool memcpy_in = p->memcpy_first || p->memcpy_last;
    if (!memcpy_in && !p->memcpy_out) {
        av_assert0(safe_blocks == num_blocks);
        return 0;
    }

    /* Set-up tail section parameters and buffers */
    SwsOpExec *tail = &p->exec_tail;
    const int align = av_cpu_max_align();
    size_t alloc_size = 0;
    *tail = *exec;

    const size_t safe_width = safe_blocks * block_size;
    const size_t tail_size  = width - safe_width;
    p->tail_off_out  = pixel_bytes(safe_width, p->pixel_bits_out, AV_ROUND_DOWN);
    p->tail_size_out = pixel_bytes(tail_size,  p->pixel_bits_out, AV_ROUND_UP);
    p->tail_blocks   = num_blocks - safe_blocks;

    if (exec->in_offset_x) {
        p->tail_off_in  = exec->in_offset_x[safe_width];
        p->tail_size_in = exec->in_offset_x[width - 1] - p->tail_off_in;
        p->tail_size_in += pixel_bytes(p->filter_size_h, p->pixel_bits_in, AV_ROUND_UP);
    } else {
        p->tail_off_in  = pixel_bytes(safe_width, p->pixel_bits_in, AV_ROUND_DOWN);
        p->tail_size_in = pixel_bytes(tail_size,  p->pixel_bits_in, AV_ROUND_UP);
    }

    const size_t alloc_width = aligned_w - safe_width;
    for (int i = 0; memcpy_in && i < p->planes_in; i++) {
        size_t needed_size;
        if (exec->in_offset_x) {
            /* The input offset map is already padded to multiples of the block
             * size, and clamps the input offsets to the image boundaries; so
             * we just need to compensate for the comp->over_read */
            needed_size = p->tail_size_in;
        } else {
            needed_size = pixel_bytes(alloc_width, p->pixel_bits_in, AV_ROUND_UP);
        }
        size_t loop_size   = p->tail_blocks * exec->block_size_in[i];
        tail->in_stride[i] = FFALIGN(needed_size + comp->over_read[i], align);
        tail->in_bump[i]   = tail->in_stride[i] - loop_size;
        alloc_size += tail->in_stride[i] * in->height;
    }

    for (int i = 0; p->memcpy_out && i < p->planes_out; i++) {
        size_t needed_size  = pixel_bytes(alloc_width, p->pixel_bits_out, AV_ROUND_UP);
        size_t loop_size    = p->tail_blocks * exec->block_size_out[i];
        tail->out_stride[i] = FFALIGN(needed_size + comp->over_write[i], align);
        tail->out_bump[i]   = tail->out_stride[i] - loop_size;
        alloc_size += tail->out_stride[i] * out->height;
    }

    if (memcpy_in && exec->in_offset_x) {
        /* `in_offset_x` is indexed relative to the line start, not the start
         * of the section being processed; so we need to over-allocate this
         * array to the full width of the image, even though we will only
         * partially fill in the offsets relevant to the tail region */
        alloc_size += aligned_w * sizeof(*exec->in_offset_x);
    }

    av_fast_mallocz(&p->tail_buf, &p->tail_buf_size, alloc_size);
    if (!p->tail_buf)
        return AVERROR(ENOMEM);

    uint8_t *tail_buf = p->tail_buf;
    for (int i = 0; memcpy_in && i < p->planes_in; i++) {
        tail->in[i] = tail_buf;
        tail_buf += tail->in_stride[i] * in->height;
    }

    for (int i = 0; p->memcpy_out && i < p->planes_out; i++) {
        tail->out[i] = tail_buf;
        tail_buf += tail->out_stride[i] * out->height;
    }

    if (memcpy_in && exec->in_offset_x) {
        tail->in_offset_x = (int32_t *) tail_buf;
        for (int i = safe_width; i < aligned_w; i++)
            tail->in_offset_x[i] = exec->in_offset_x[i] - p->tail_off_in;
    }

    return 0;
}

static void copy_lines(uint8_t *dst, const size_t dst_stride,
                       const uint8_t *src, const size_t src_stride,
                       const int h, const size_t bytes)
{
    for (int y = 0; y < h; y++) {
        memcpy(dst, src, bytes);
        dst += dst_stride;
        src += src_stride;
    }
}

static void op_pass_run(const SwsFrame *out, const SwsFrame *in, const int y,
                        const int h, const SwsPass *pass)
{
    const SwsOpPass *p = pass->priv;
    const SwsCompiledOp *comp = &p->comp;

    /* Fill exec metadata for this slice */
    DECLARE_ALIGNED_32(SwsOpExec, exec) = p->exec_base;
    exec.slice_y = y;
    exec.slice_h = h;

    /**
     *  To ensure safety, we need to consider the following:
     *
     * 1. We can overread the input, unless this is the last line of an
     *    unpadded buffer. All defined operations can handle arbitrary pixel
     *    input, so overread of arbitrary data is fine. For flipped images,
     *    this condition is actually *inverted* to where the first line is
     *    the one at the end of the buffer.
     *
     * 2. We can overwrite the output, as long as we don't write more than the
     *    amount of pixels that fit into one linesize. So we always need to
     *    memcpy the last column on the output side if unpadded.
     */

    const bool memcpy_in  = p->memcpy_last && y + h == pass->lines ||
                            p->memcpy_first && y == 0;
    const bool memcpy_out = p->memcpy_out;
    const size_t num_blocks  = p->num_blocks;
    const size_t tail_blocks = p->tail_blocks;

    get_row_data(p, y, exec.in, exec.out);
    if (!memcpy_in && !memcpy_out) {
        /* Fast path (fully aligned/padded inputs and outputs) */
        comp->func(&exec, comp->priv, 0, y, num_blocks, y + h);
        return;
    }

    /* Non-aligned case (slow path); process main blocks as normal, and
     * a separate tail (via memcpy into an appropriately padded buffer) */
    if (num_blocks > tail_blocks) {
        for (int i = 0; i < 4; i++) {
            /* We process fewer blocks, so the in_bump needs to be increased
             * to reflect that the plane pointers are left on the last block,
             * not the end of the processed line, after each loop iteration */
            exec.in_bump[i]  += exec.block_size_in[i]  * tail_blocks;
            exec.out_bump[i] += exec.block_size_out[i] * tail_blocks;
        }

        comp->func(&exec, comp->priv, 0, y, num_blocks - tail_blocks, y + h);
    }

    DECLARE_ALIGNED_32(SwsOpExec, tail) = p->exec_tail;
    tail.slice_y = y;
    tail.slice_h = h;

    for (int i = 0; i < p->planes_in; i++) {
        /* Input offsets are relative to the base pointer */
        if (!exec.in_offset_x || memcpy_in)
            exec.in[i] += p->tail_off_in;
        tail.in[i] += y * tail.in_stride[i];
    }
    for (int i = 0; i < p->planes_out; i++) {
        exec.out[i] += p->tail_off_out;
        tail.out[i] += y * tail.out_stride[i];
    }

    for (int i = 0; i < p->planes_in; i++) {
        if (memcpy_in) {
            const int lines = get_lines_in(p, y, h, i);
            copy_lines((uint8_t *) tail.in[i], tail.in_stride[i],
                       exec.in[i], exec.in_stride[i], lines, p->tail_size_in);
        } else {
            /* Reuse input pointers directly */
            const size_t loop_size = tail_blocks * exec.block_size_in[i];
            tail.in[i]        = exec.in[i];
            tail.in_stride[i] = exec.in_stride[i];
            tail.in_bump[i]   = exec.in_stride[i] - loop_size;
        }
    }

    for (int i = 0; !memcpy_out && i < p->planes_out; i++) {
        /* Reuse output pointers directly */
        const size_t loop_size = tail_blocks * exec.block_size_out[i];
        tail.out[i]        = exec.out[i];
        tail.out_stride[i] = exec.out_stride[i];
        tail.out_bump[i]   = exec.out_stride[i] - loop_size;
    }

    /* Dispatch kernel over tail */
    av_assert1(tail_blocks > 0);
    comp->func(&tail, comp->priv, num_blocks - tail_blocks, y, num_blocks, y + h);

    for (int i = 0; memcpy_out && i < p->planes_out; i++) {
        const int lines = h >> tail.out_sub_y[i];
        copy_lines(exec.out[i], exec.out_stride[i],
                   tail.out[i], tail.out_stride[i], lines, p->tail_size_out);
    }
}

/* Updates the plane copy (no-op) map for this operation list */
static void op_list_get_plane_copy(const SwsOpList *ops, SwsPass *pass)
{
    const SwsOp *write = ff_sws_op_list_output(ops);
    const SwsOp *read  = ff_sws_op_list_input(ops);
    if (!write || write->rw.mode != SWS_RW_PLANAR ||
        !read  || read->rw.mode  != SWS_RW_PLANAR ||
        read->type    != write->type ||
        read->rw.frac != write->rw.frac)
        return; /* only regular planes can be directly ref'd */

    const SwsOp *prev = &ops->ops[ops->num_ops - 2];
    int *plane_copy = pass->output->plane_copy;
    for (int i = 0; i < write->rw.elems; i++) {
        SwsCompFlags flags = prev->comps.flags[i];
        if (!(flags & SWS_COMP_COPY))
            continue;

        const int out_idx = ops->plane_dst[i];
        switch (prev->comps.dep_in[i]) {
        case SWS_COMP(0): plane_copy[out_idx] = ops->plane_src[0]; break;
        case SWS_COMP(1): plane_copy[out_idx] = ops->plane_src[1]; break;
        case SWS_COMP(2): plane_copy[out_idx] = ops->plane_src[2]; break;
        case SWS_COMP(3): plane_copy[out_idx] = ops->plane_src[3]; break;
        }
    }
}

static int rw_data_planes(const SwsOp *op)
{
    /* Exclude the palette plane from the plane count, since it does not need
     * to be directly processed/adjusted by the dispatch layer */
    return op->rw.mode == SWS_RW_PALETTE ? 1 : ff_sws_rw_op_planes(op);
}

static int rw_pixel_bits(const SwsOp *op)
{
    if (op->rw.mode == SWS_RW_PALETTE)
        return 8; /* index size */

    int elems = 0;
    switch (op->rw.mode) {
    case SWS_RW_PLANAR: elems = 1; break;
    case SWS_RW_PACKED: elems = op->rw.elems; break;
    }

    const int size  = ff_sws_pixel_type_size(op->type);
    const int bits  = 8 >> op->rw.frac;
    av_assert1(bits >= 1);
    return elems * size * bits;
}

static void align_pass(SwsPass *pass, int block_size, const int *over_rw,
                       int pixel_bits)
{
    if (!pass || pixel_bits <= 0)
        return;

    /* Add at least as many pixels as needed to cover the padding requirement */
    int pad_max = 0;
    for (int i = 0; i < 4; i++) {
        const int pad = (over_rw[i] * 8 + pixel_bits - 1) / pixel_bits;
        pad_max = FFMAX(pad_max, pad);
    }

    SwsPassBuffer *buf = pass->output;
    buf->width_align = FFMAX(buf->width_align, block_size);
    buf->width_pad = FFMAX(buf->width_pad, pad_max);
}

/* Unchanging part of parameter list */
typedef struct CompileArgs {
    const SwsOpBackend *backend;
    SwsGraph *graph;
    int flags;
} CompileArgs;

static int compile_single(const CompileArgs *args, const SwsOpList *ops,
                          SwsPass *link, SwsPass *input, SwsPass **output)
{
    SwsGraph *graph = args->graph;
    SwsContext *ctx = graph->ctx;
    SwsOpPass *p = av_mallocz(sizeof(*p));
    if (!p)
        return AVERROR(ENOMEM);

    int ret = ff_sws_ops_compile(ctx, args->backend, ops, &p->comp);
    if (ret < 0)
        goto fail;
    else if (args->flags & SWS_OP_FLAG_DRY_RUN)
        goto fail; /* nothing to do, just return */

    const SwsCompiledOp *comp = &p->comp;
    const SwsFormat *src = &ops->src;
    const SwsFormat *dst = &ops->dst;
    av_assert0(!link || link->format == dst->format);
    if (p->comp.opaque) {
        SwsCompiledOp c = *comp;
        av_free(p);
        ret = ff_sws_graph_add_pass(graph, dst->format, dst->width, dst->height,
                                    input, 0, c.slice_align, c.func_opaque,
                                    NULL, c.priv, c.free, output);
        if (ret >= 0) {
            (*output)->backend = c.backend->flags;
            op_list_get_plane_copy(ops, *output);
            ff_sws_pass_link_output(*output, link);
        }
        return ret;
    }

    const AVPixFmtDescriptor *indesc  = av_pix_fmt_desc_get(src->format);
    const AVPixFmtDescriptor *outdesc = av_pix_fmt_desc_get(dst->format);
    const SwsOp *write = ff_sws_op_list_output(ops);
    p->planes_out     = rw_data_planes(write);
    p->pixel_bits_out = rw_pixel_bits(write);
    p->palette_idx    = -1;
    p->exec_base = (SwsOpExec) {
        .width  = dst->width,
        .height = dst->height,
    };

    const SwsOp *read = ff_sws_op_list_input(ops);
    if (read) {
        p->planes_in     = rw_data_planes(read);
        p->pixel_bits_in = rw_pixel_bits(read);
        if (read->rw.mode == SWS_RW_PALETTE)
            p->palette_idx = ops->plane_src[1];
    }

    const int64_t block_bits_in  = (int64_t) comp->block_size * p->pixel_bits_in;
    const int64_t block_bits_out = (int64_t) comp->block_size * p->pixel_bits_out;
    if (block_bits_in & 0x7 || block_bits_out & 0x7) {
        av_log(ctx, AV_LOG_ERROR, "Block size must be byte-aligned.\n");
        ret = AVERROR(EINVAL);
        goto fail;
    }

    for (int i = 0; i < 4; i++)
        p->idx_in[i] = p->idx_out[i] = -1;

    for (int i = 0; i < p->planes_in; i++) {
        const int idx = ops->plane_src[i];
        const int chroma = idx == 1 || idx == 2;
        const int sub_x = chroma ? indesc->log2_chroma_w : 0;
        const int sub_y = chroma ? indesc->log2_chroma_h : 0;
        p->exec_base.in_sub_x[i] = sub_x;
        p->exec_base.in_sub_y[i] = sub_y;
        p->exec_base.block_size_in[i] = block_bits_in >> 3;
        p->idx_in[i] = idx;
    }

    for (int i = 0; i < p->planes_out; i++) {
        const int idx = ops->plane_dst[i];
        const int chroma = idx == 1 || idx == 2;
        const int sub_x = chroma ? outdesc->log2_chroma_w : 0;
        const int sub_y = chroma ? outdesc->log2_chroma_h : 0;
        p->exec_base.out_sub_x[i] = sub_x;
        p->exec_base.out_sub_y[i] = sub_y;
        p->exec_base.block_size_out[i] = block_bits_out >> 3;
        p->idx_out[i] = idx;
    }

    const SwsFilterWeights *filter = read ? read->rw.filter.kernel : NULL;
    if (read && read->rw.filter.op == SWS_OP_FILTER_V) {
        p->offsets_y = av_refstruct_ref(filter->offsets);

        /* Compute relative pointer bumps for each output line */
        int32_t *bump = av_malloc_array(filter->dst_size, sizeof(*bump));
        if (!bump) {
            ret = AVERROR(ENOMEM);
            goto fail;
        }

        int line = filter->offsets[0];
        for (int y = 0; y < filter->dst_size - 1; y++) {
            int next = filter->offsets[y + 1];
            bump[y] = next - line - 1;
            line = next;
        }
        bump[filter->dst_size - 1] = 0;
        p->exec_base.in_bump_y = bump;
    } else if (read && read->rw.filter.op == SWS_OP_FILTER_H) {
        /* Compute pixel offset map for each output line */
        const int pixels = FFALIGN(filter->dst_size, p->comp.block_size);
        int32_t *offset = av_malloc_array(pixels, sizeof(*offset));
        if (!offset) {
            ret = AVERROR(ENOMEM);
            goto fail;
        }
        p->exec_base.in_offset_x = offset;

        for (int x = 0; x < filter->dst_size; x++) {
            /* Sanity check; if the tap would land on a half-pixel, we cannot
             * reasonably expect the implementation to know about this. Just
             * error out in such (theoretical) cases. */
            int64_t bits = (int64_t) filter->offsets[x] * p->pixel_bits_in;
            if ((bits & 0x7) || (bits >> 3) > INT32_MAX) {
                ret = AVERROR(EINVAL);
                goto fail;
            }
            offset[x] = bits >> 3;
        }
        for (int x = filter->dst_size; x < pixels; x++)
            offset[x] = offset[filter->dst_size - 1];
        for (int i = 0; i < 4; i++)
            p->exec_base.block_size_in[i] = 0; /* ptr does not advance */
        p->filter_size_h = filter->filter_size;
    }

    ret = ff_sws_graph_add_pass(graph, dst->format, dst->width, dst->height,
                                input, 0, comp->slice_align, op_pass_run,
                                op_pass_setup, p, op_pass_free, output);
    if (ret < 0)
        return ret;

    (*output)->backend = comp->backend->flags;
    op_list_get_plane_copy(ops, *output);
    ff_sws_pass_link_output(*output, link);
    align_pass(*output, comp->block_size, comp->over_write, p->pixel_bits_out);
    if (read)
        align_pass(input, comp->block_size, comp->over_read,  p->pixel_bits_in);
    return 0;

fail:
    op_pass_free(p);
    return ret;
}

/* Return a mask of all planes matching any flag in `flags` */
static SwsCompMask plane_mask_flags(const SwsOp *op, SwsCompFlags flags)
{
    SwsCompMask planes = 0;
    for (int c = 0; c < 4; c++) {
        if (op->comps.flags[c] & flags)
            planes |= SWS_COMP(c);
    }

    return planes;
}

/* Takes over ownership of *pops, even on failure */
static int compile_subpass(const CompileArgs *args, SwsOpList **pops,
                           SwsPass *link, SwsPass *input, SwsPass **output)
{
    int ret;
    SwsContext *ctx = args->graph->ctx;
    SwsOpList *ops  = *pops;
    SwsOpList *rest = NULL;
    SwsPass *tmp = NULL;
    *pops = NULL;

    if (args->flags & SWS_OP_FLAG_SPLIT_MEMCPY) {
        /* Split off copied and constant planes into a separate subpass,
         * since these are likely to be handled by the memcpy backend */
        av_assert0(ops->num_ops >= 2);
        const SwsOp *prev = &ops->ops[ops->num_ops - 2];
        SwsCompMask planes = plane_mask_flags(prev, SWS_COMP_COPY | SWS_COMP_CONST);
        RET(ff_sws_op_list_split_planes(ops, &rest, planes));
        if (rest) {
            /* Parallel split: share input and link all outputs together */
            av_log(ctx, AV_LOG_DEBUG, "Splitting const/memcpy planes: %s\n",
                   ff_sws_comp_mask_str(planes));
            RET(compile_subpass(args, &ops,  link, input, &tmp));
            RET(compile_subpass(args, &rest, tmp, input, output));
            return 0;
        }
    }

    ret = compile_single(args, ops, link, input, output);
    if (ret != AVERROR(ENOTSUP))
        goto fail; /* either success or a hard error */

    /* Find any unresolved filter */
    for (int idx = 1; idx < ops->num_ops - 1; idx++) {
        const SwsOp *op = &ops->ops[idx];
        if (op->op == SWS_OP_FILTER_H || op->op == SWS_OP_FILTER_V) {
            RET(ff_sws_op_list_split_at(ops, &rest, idx));
            if (ff_sws_op_list_is_noop(ops)) {
                /* Prevent infinite recursion by avoiding splitting in a way
                 * that does not meaningfully reduce the number of operations
                 * performed in the second part. */
                FFSWAP(SwsOpList *, ops, rest);
                break;
            }
            /* Serial split: feed first pass into second */
            RET(compile_subpass(args, &ops,  NULL, input, &tmp));
            RET(compile_subpass(args, &rest, link, tmp, output));
            return 0;
        }
    }

    /* If we didn't find any more operations to eliminate, then this ops list
     * is simply unsupported by any of the available backends */
    av_log(ctx, AV_LOG_WARNING, "No backend found for operations:\n");
    ff_sws_op_list_print(ctx, AV_LOG_WARNING, AV_LOG_TRACE, ops);
    ret = AVERROR(ENOTSUP);

fail:
    ff_sws_op_list_free(&ops);
    ff_sws_op_list_free(&rest);
    return ret;
}

int ff_sws_compile_pass(SwsGraph *graph, const SwsOpBackend *backend,
                        SwsOpList **pops, int flags, SwsPass *input,
                        SwsPass **output)
{
    const int passes_orig = graph->num_passes;
    SwsContext *ctx = graph->ctx;
    SwsOpList *ops = *pops;
    int ret = 0;

    const SwsOp *write = ff_sws_op_list_output(ops);
    if (!write) {
        av_log(ctx, AV_LOG_ERROR, "Last operation must be SWS_OP_WRITE.\n");
        ret = AVERROR(EINVAL);
        goto out;
    }

    if (flags & SWS_OP_FLAG_OPTIMIZE) {
        ret = ff_sws_op_list_optimize(ops);
        if (ret < 0)
            goto out;
        av_log(ctx, AV_LOG_DEBUG, "Operation list after optimizing:\n");
        ff_sws_op_list_print(ctx, AV_LOG_DEBUG, AV_LOG_TRACE, ops);
    }

    /* Check if the whole operation graph is an end-to-end no-op */
    if (ff_sws_op_list_is_noop(ops)) {
        if (output)
            *output = input;
        goto out;
    }

    const CompileArgs args = {
        .backend = backend,
        .graph   = graph,
        .flags   = flags,
    };

    ret = compile_subpass(&args, &ops, NULL, input, output);
    if (ret < 0)
        goto out;

    const int num_passes = graph->num_passes - passes_orig;
    if (num_passes > 1)
        av_log(ctx, AV_LOG_VERBOSE, "Using %d separate passes.\n", num_passes);

out:
    if (ret < 0)
        ff_sws_graph_rollback(graph, passes_orig);
    ff_sws_op_list_free(&ops);
    *pops = NULL;
    return ret;
}