[FFmpeg-devel] avfilter/vf_nlmeans: add >8 bit support

Submitted by Paul B Mahol on Nov. 20, 2019, 9:54 a.m.

Details

Message ID 20191120095442.21068-1-onemda@gmail.com
State New
Headers show

Commit Message

Paul B Mahol Nov. 20, 2019, 9:54 a.m.
Signed-off-by: Paul B Mahol <onemda@gmail.com>
---
 libavfilter/aarch64/vf_nlmeans_init.c |   6 +-
 libavfilter/nlmeans_template.c        | 370 ++++++++++++++++++
 libavfilter/vf_nlmeans.c              | 539 ++++++--------------------
 libavfilter/vf_nlmeans.h              |  59 ++-
 4 files changed, 548 insertions(+), 426 deletions(-)
 create mode 100644 libavfilter/nlmeans_template.c

Comments

mypopy@gmail.com Nov. 20, 2019, 12:12 p.m.
On Wed, Nov 20, 2019 at 5:55 PM Paul B Mahol <onemda@gmail.com> wrote:
>
> Signed-off-by: Paul B Mahol <onemda@gmail.com>
> ---
>  libavfilter/aarch64/vf_nlmeans_init.c |   6 +-
>  libavfilter/nlmeans_template.c        | 370 ++++++++++++++++++
>  libavfilter/vf_nlmeans.c              | 539 ++++++--------------------
>  libavfilter/vf_nlmeans.h              |  59 ++-
>  4 files changed, 548 insertions(+), 426 deletions(-)
>  create mode 100644 libavfilter/nlmeans_template.c
>
> diff --git a/libavfilter/aarch64/vf_nlmeans_init.c b/libavfilter/aarch64/vf_nlmeans_init.c
> index a1edefb144..6ee157dfe0 100644
> --- a/libavfilter/aarch64/vf_nlmeans_init.c
> +++ b/libavfilter/aarch64/vf_nlmeans_init.c
> @@ -24,10 +24,10 @@ void ff_compute_safe_ssd_integral_image_neon(uint32_t *dst, ptrdiff_t dst_linesi
>                                               const uint8_t *s2, ptrdiff_t linesize2,
>                                               int w, int h);
>
> -av_cold void ff_nlmeans_init_aarch64(NLMeansDSPContext *dsp)
> +av_cold void ff_nlmeans_init_aarch64(NLMeansDSPContext *dsp, int depth)
>  {
>      int cpu_flags = av_get_cpu_flags();
>
> -    if (have_neon(cpu_flags))
> -        dsp->compute_safe_ssd_integral_image = ff_compute_safe_ssd_integral_image_neon;
> +    if (have_neon(cpu_flags) && depth == 8)
> +        dsp->compute_safe_ssd_integral_image32 = ff_compute_safe_ssd_integral_image_neon;
>  }
> diff --git a/libavfilter/nlmeans_template.c b/libavfilter/nlmeans_template.c
> new file mode 100644
> index 0000000000..b0d7f7482d
> --- /dev/null
> +++ b/libavfilter/nlmeans_template.c
> @@ -0,0 +1,370 @@
> +/*
> + * Copyright (c) 2016 Clément Bœsch <u pkh me>
> + *
> + * 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 "avfilter.h"
> +#include "formats.h"
> +#include "internal.h"
> +#include "vf_nlmeans.h"
> +#include "video.h"
> +
> +#undef pixel
> +#undef integral
> +#undef INTEGRAL_DEPTH
> +#if DEPTH <= 8
> +#define INTEGRAL_DEPTH 32
> +#define pixel uint8_t
> +#define integral uint32_t
> +#else
> +#define INTEGRAL_DEPTH 64
> +#define pixel uint16_t
> +#define integral uint64_t
> +#endif
> +
> +#define fn3(a,b)   a##b
> +#define fn2(a,b)   fn3(a,b)
> +#define fn(a)      fn2(a, DEPTH)
> +#define ii(a)      fn2(a, INTEGRAL_DEPTH)
> +
> +/**
> + * Compute squared difference of the safe area (the zone where s1 and s2
> + * overlap). It is likely the largest integral zone, so it is interesting to do
> + * as little checks as possible; contrary to the unsafe version of this
> + * function, we do not need any clipping here.
> + *
> + * The line above dst and the column to its left are always readable.
> + */
> +static void fn(compute_safe_ssd_integral_image_c)(integral *dst, ptrdiff_t dst_linesize_32,
> +                                                  const uint8_t *ss1, ptrdiff_t linesize1,
> +                                                  const uint8_t *ss2, ptrdiff_t linesize2,
> +                                                  int w, int h)
> +{
> +    const pixel *s1 = (const pixel *)ss1;
> +    const pixel *s2 = (const pixel *)ss2;
> +    int x, y;
> +    const integral *dst_top = dst - dst_linesize_32;
> +
> +    linesize1 /= sizeof(pixel);
> +    linesize2 /= sizeof(pixel);
> +
> +    /* SIMD-friendly assumptions allowed here */
> +    av_assert2(!(w & 0xf) && w >= 16 && h >= 1);
> +
> +    for (y = 0; y < h; y++) {
> +        for (x = 0; x < w; x += 4) {
> +            const int d0 = s1[x    ] - s2[x    ];
> +            const int d1 = s1[x + 1] - s2[x + 1];
> +            const int d2 = s1[x + 2] - s2[x + 2];
> +            const int d3 = s1[x + 3] - s2[x + 3];
> +
> +            dst[x    ] = dst_top[x    ] - dst_top[x - 1] + d0*d0;
> +            dst[x + 1] = dst_top[x + 1] - dst_top[x    ] + d1*d1;
> +            dst[x + 2] = dst_top[x + 2] - dst_top[x + 1] + d2*d2;
> +            dst[x + 3] = dst_top[x + 3] - dst_top[x + 2] + d3*d3;
> +
> +            dst[x    ] += dst[x - 1];
> +            dst[x + 1] += dst[x    ];
> +            dst[x + 2] += dst[x + 1];
> +            dst[x + 3] += dst[x + 2];
> +        }
> +        s1  += linesize1;
> +        s2  += linesize2;
> +        dst += dst_linesize_32;
> +        dst_top += dst_linesize_32;
> +    }
> +}
> +
> +/**
> + * Compute squared difference of an unsafe area (the zone nor s1 nor s2 could
> + * be readable).
> + *
> + * On the other hand, the line above dst and the column to its left are always
> + * readable.
> + *
> + * There is little point in having this function SIMDified as it is likely too
> + * complex and only handle small portions of the image.
> + *
> + * @param dst               integral image
> + * @param dst_linesize_32   integral image linesize (in 32-bit integers unit)
> + * @param startx            integral starting x position
> + * @param starty            integral starting y position
> + * @param ssrc              source plane buffer
> + * @param linesize          source plane linesize
> + * @param offx              source offsetting in x
> + * @param offy              source offsetting in y
> + * @paran r                 absolute maximum source offsetting
> + * @param sw                source width
> + * @param sh                source height
> + * @param w                 width to compute
> + * @param h                 height to compute
> + */
> +static inline void fn(compute_unsafe_ssd_integral_image)(integral *dst, ptrdiff_t dst_linesize_32,
> +                                                         int startx, int starty,
> +                                                         const uint8_t *ssrc, ptrdiff_t linesize,
> +                                                         int offx, int offy, int r, int sw, int sh,
> +                                                         int w, int h)
> +{
> +    const pixel *src = (const pixel *)ssrc;
> +    int x, y;
> +
> +    linesize /= sizeof(pixel);
> +
> +    for (y = starty; y < starty + h; y++) {
> +        integral acc = dst[y*dst_linesize_32 + startx - 1] - dst[(y-1)*dst_linesize_32 + startx - 1];
> +        const int s1y = av_clip(y -  r,         0, sh - 1);
> +        const int s2y = av_clip(y - (r + offy), 0, sh - 1);
> +
> +        for (x = startx; x < startx + w; x++) {
> +            const int s1x = av_clip(x -  r,         0, sw - 1);
> +            const int s2x = av_clip(x - (r + offx), 0, sw - 1);
> +            const pixel v1 = src[s1y*linesize + s1x];
> +            const pixel v2 = src[s2y*linesize + s2x];
> +            const int d = v1 - v2;
> +            acc += d * d;
> +            dst[y*dst_linesize_32 + x] = dst[(y-1)*dst_linesize_32 + x] + acc;
> +        }
> +    }
> +}
> +
> +/*
> + * Compute the sum of squared difference integral image
> + * http://www.ipol.im/pub/art/2014/57/
> + * Integral Images for Block Matching - Gabriele Facciolo, Nicolas Limare, Enric Meinhardt-Llopis
> + *
> + * @param ii                integral image of dimension (w+e*2) x (h+e*2) with
> + *                          an additional zeroed top line and column already
> + *                          "applied" to the pointer value
> + * @param ii_linesize_32    integral image linesize (in 32-bit integers unit)
> + * @param src               source plane buffer
> + * @param linesize          source plane linesize
> + * @param offx              x-offsetting ranging in [-e;e]
> + * @param offy              y-offsetting ranging in [-e;e]
> + * @param w                 source width
> + * @param h                 source height
> + * @param e                 research padding edge
> + */
> +static void fn(compute_ssd_integral_image)(const NLMeansDSPContext *dsp,
> +                                           integral *ii, ptrdiff_t ii_linesize_32,
> +                                           const uint8_t *src, ptrdiff_t linesize, int offx, int offy,
> +                                           int e, int w, int h)
> +{
> +    // ii has a surrounding padding of thickness "e"
> +    const int ii_w = w + e*2;
> +    const int ii_h = h + e*2;
> +
> +    // we center the first source
> +    const int s1x = e;
> +    const int s1y = e;
> +
> +    // 2nd source is the frame with offsetting
> +    const int s2x = e + offx;
> +    const int s2y = e + offy;
> +
> +    // get the dimension of the overlapping rectangle where it is always safe
> +    // to compare the 2 sources pixels
> +    const int startx_safe = FFMAX(s1x, s2x);
> +    const int starty_safe = FFMAX(s1y, s2y);
> +    const int u_endx_safe = FFMIN(s1x + w, s2x + w); // unaligned
> +    const int endy_safe   = FFMIN(s1y + h, s2y + h);
> +
> +    // deduce the safe area width and height
> +    const int safe_pw = (u_endx_safe - startx_safe) & ~0xf;
> +    const int safe_ph = endy_safe - starty_safe;
> +
> +    // adjusted end x position of the safe area after width of the safe area gets aligned
> +    const int endx_safe = startx_safe + safe_pw;
> +
> +    // top part where only one of s1 and s2 is still readable, or none at all
> +    fn(compute_unsafe_ssd_integral_image)(ii, ii_linesize_32,
> +                                          0, 0,
> +                                          src, linesize,
> +                                          offx, offy, e, w, h,
> +                                          ii_w, starty_safe);
> +
> +    // fill the left column integral required to compute the central
> +    // overlapping one
> +    fn(compute_unsafe_ssd_integral_image)(ii, ii_linesize_32,
> +                                          0, starty_safe,
> +                                          src, linesize,
> +                                          offx, offy, e, w, h,
> +                                          startx_safe, safe_ph);
> +
> +    // main and safe part of the integral
> +    av_assert1(startx_safe - s1x >= 0); av_assert1(startx_safe - s1x < w);
> +    av_assert1(starty_safe - s1y >= 0); av_assert1(starty_safe - s1y < h);
> +    av_assert1(startx_safe - s2x >= 0); av_assert1(startx_safe - s2x < w);
> +    av_assert1(starty_safe - s2y >= 0); av_assert1(starty_safe - s2y < h);
> +    if (safe_pw && safe_ph)
> +        ii(dsp->compute_safe_ssd_integral_image)(ii + starty_safe*ii_linesize_32 + startx_safe, ii_linesize_32,
> +                                                 src + (starty_safe - s1y) * linesize + (startx_safe - s1x) * sizeof(pixel), linesize,
> +                                                 src + (starty_safe - s2y) * linesize + (startx_safe - s2x) * sizeof(pixel), linesize,
> +                                                 safe_pw, safe_ph);
> +
> +    // right part of the integral
> +    fn(compute_unsafe_ssd_integral_image)(ii, ii_linesize_32,
> +                                          endx_safe, starty_safe,
> +                                          src, linesize,
> +                                          offx, offy, e, w, h,
> +                                          ii_w - endx_safe, safe_ph);
> +
> +    // bottom part where only one of s1 and s2 is still readable, or none at all
> +    fn(compute_unsafe_ssd_integral_image)(ii, ii_linesize_32,
> +                                          0, endy_safe,
> +                                          src, linesize,
> +                                          offx, offy, e, w, h,
> +                                          ii_w, ii_h - endy_safe);
> +}
> +
> +static int fn(nlmeans_slice)(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
> +{
> +    int x, y;
> +    NLMeansContext *s = ctx->priv;
> +    const struct thread_data *td = arg;
> +    const ptrdiff_t src_linesize = td->src_linesize;
> +    const int process_h = td->endy - td->starty;
> +    const int slice_start = (process_h *  jobnr   ) / nb_jobs;
> +    const int slice_end   = (process_h * (jobnr+1)) / nb_jobs;
> +    const int starty = td->starty + slice_start;
> +    const int endy   = td->starty + slice_end;
> +    const int p = td->p;
> +    const integral *ii = ii(td->ii_start) + (starty - p - 1) * s->ii_lz_32 - p - 1;
> +    const int dist_b = 2*p + 1;
> +    const int dist_d = dist_b * s->ii_lz_32;
> +    const int dist_e = dist_d + dist_b;
> +
> +    for (y = starty; y < endy; y++) {
> +        const pixel *src = (const pixel *)(td->src + y*src_linesize);
> +        struct weighted_avg *wa = s->wa + y*s->wa_linesize;
> +        for (x = td->startx; x < td->endx; x++) {
> +            /*
> +             * M is a discrete map where every entry contains the sum of all the entries
> +             * in the rectangle from the top-left origin of M to its coordinate. In the
> +             * following schema, "i" contains the sum of the whole map:
> +             *
> +             * M = +----------+-----------------+----+
> +             *     |          |                 |    |
> +             *     |          |                 |    |
> +             *     |         a|                b|   c|
> +             *     +----------+-----------------+----+
> +             *     |          |                 |    |
> +             *     |          |                 |    |
> +             *     |          |        X        |    |
> +             *     |          |                 |    |
> +             *     |         d|                e|   f|
> +             *     +----------+-----------------+----+
> +             *     |          |                 |    |
> +             *     |         g|                h|   i|
> +             *     +----------+-----------------+----+
> +             *
> +             * The sum of the X box can be calculated with:
> +             *    X = e-d-b+a
> +             *
> +             * See https://en.wikipedia.org/wiki/Summed_area_table
> +             *
> +             * The compute*_ssd functions compute the integral image M where every entry
> +             * contains the sum of the squared difference of every corresponding pixels of
> +             * two input planes of the same size as M.
> +             */
> +            const integral a = ii[x];
> +            const integral b = ii[x + dist_b];
> +            const integral d = ii[x + dist_d];
> +            const integral e = ii[x + dist_e];
> +            const integral patch_diff_sq = e + a - d - b;
> +
> +            if (patch_diff_sq < s->max_meaningful_diff) {
> +                const float weight = s->weight_lut[patch_diff_sq]; // exp(-patch_diff_sq * s->pdiff_scale)
> +                wa[x].total_weight += weight;
> +                wa[x].sum += weight * src[x];
> +            }
> +        }
> +        ii += s->ii_lz_32;
> +    }
> +    return 0;
> +}
> +
> +static void fn(weight_averages)(uint8_t *ddst, ptrdiff_t dst_linesize,
> +                                const uint8_t *ssrc, ptrdiff_t src_linesize,
> +                                struct weighted_avg *wa, ptrdiff_t wa_linesize,
> +                                int w, int h)
> +{
> +    const pixel *src = (const pixel *)ssrc;
> +    pixel *dst = (pixel *)ddst;
> +    int x, y;
> +
> +    src_linesize /= sizeof(pixel);
> +    dst_linesize /= sizeof(pixel);
> +
> +    for (y = 0; y < h; y++) {
> +        for (x = 0; x < w; x++) {
> +            // Also weight the centered pixel
> +            wa[x].total_weight += 1.f;
> +            wa[x].sum += 1.f * src[x];
> +            dst[x] = av_clip_uintp2(wa[x].sum / wa[x].total_weight + 0.5f, DEPTH);
> +        }
> +        dst += dst_linesize;
> +        src += src_linesize;
> +        wa += wa_linesize;
> +    }
> +}
> +
> +static int fn(nlmeans_plane)(AVFilterContext *ctx, int w, int h, int p, int r,
> +                             uint8_t *dst, ptrdiff_t dst_linesize,
> +                             const uint8_t *src, ptrdiff_t src_linesize)
> +{
> +    int offx, offy;
> +    NLMeansContext *s = ctx->priv;
> +    /* patches center points cover the whole research window so the patches
> +     * themselves overflow the research window */
> +    const int e = r + p;
> +    /* focus an integral pointer on the centered image (s1) */
> +    const uint32_t *centered_ii32 = s->ii32 + e*s->ii_lz_32 + e;
> +    const uint64_t *centered_ii64 = s->ii64 + e*s->ii_lz_32 + e;
> +
> +    memset(s->wa, 0, s->wa_linesize * h * sizeof(*s->wa));
> +
> +    for (offy = -r; offy <= r; offy++) {
> +        for (offx = -r; offx <= r; offx++) {
> +            if (offx || offy) {
> +                struct thread_data td = {
> +                    .src          = src + offy*src_linesize + offx * (int)(sizeof(pixel)),
> +                    .src_linesize = src_linesize,
> +                    .startx       = FFMAX(0, -offx),
> +                    .starty       = FFMAX(0, -offy),
> +                    .endx         = FFMIN(w, w - offx),
> +                    .endy         = FFMIN(h, h - offy),
> +                    .ii_start32   = centered_ii32 + offy*s->ii_lz_32 + offx,
> +                    .ii_start64   = centered_ii64 + offy*s->ii_lz_32 + offx,
> +                    .p            = p,
> +                };
> +
> +                fn(compute_ssd_integral_image)(&s->dsp, ii(s->ii), s->ii_lz_32,
> +                                               src, src_linesize,
> +                                               offx, offy, e, w, h);
> +                ctx->internal->execute(ctx, fn(nlmeans_slice), &td, NULL,
> +                                       FFMIN(td.endy - td.starty, ff_filter_get_nb_threads(ctx)));
> +            }
> +        }
> +    }
> +
> +    fn(weight_averages)(dst, dst_linesize, src, src_linesize,
> +                        s->wa, s->wa_linesize, w, h);
> +
> +    return 0;
> +}
> diff --git a/libavfilter/vf_nlmeans.c b/libavfilter/vf_nlmeans.c
> index 06233b0dd4..bf7cf3aa0e 100644
> --- a/libavfilter/vf_nlmeans.c
> +++ b/libavfilter/vf_nlmeans.c
> @@ -38,31 +38,28 @@
>  #include "vf_nlmeans.h"
>  #include "video.h"
>
> -struct weighted_avg {
> -    float total_weight;
> -    float sum;
> -};
> +#define DEPTH 8
> +#include "nlmeans_template.c"
> +
> +#undef DEPTH
> +#define DEPTH 9
> +#include "nlmeans_template.c"
> +
> +#undef DEPTH
> +#define DEPTH 10
> +#include "nlmeans_template.c"
> +
> +#undef DEPTH
> +#define DEPTH 12
> +#include "nlmeans_template.c"
> +
> +#undef DEPTH
> +#define DEPTH 14
> +#include "nlmeans_template.c"
>
> -typedef struct NLMeansContext {
> -    const AVClass *class;
> -    int nb_planes;
> -    int chroma_w, chroma_h;
> -    double pdiff_scale;                         // invert of the filtering parameter (sigma*10) squared
> -    double sigma;                               // denoising strength
> -    int patch_size,    patch_hsize;             // patch size and half size
> -    int patch_size_uv, patch_hsize_uv;          // patch size and half size for chroma planes
> -    int research_size,    research_hsize;       // research size and half size
> -    int research_size_uv, research_hsize_uv;    // research size and half size for chroma planes
> -    uint32_t *ii_orig;                          // integral image
> -    uint32_t *ii;                               // integral image starting after the 0-line and 0-column
> -    int ii_w, ii_h;                             // width and height of the integral image
> -    ptrdiff_t ii_lz_32;                         // linesize in 32-bit units of the integral image
> -    struct weighted_avg *wa;                    // weighted average of every pixel
> -    ptrdiff_t wa_linesize;                      // linesize for wa in struct size unit
> -    float *weight_lut;                          // lookup table mapping (scaled) patch differences to their associated weights
> -    uint32_t max_meaningful_diff;               // maximum difference considered (if the patch difference is too high we ignore the pixel)
> -    NLMeansDSPContext dsp;
> -} NLMeansContext;
> +#undef DEPTH
> +#define DEPTH 16
> +#include "nlmeans_template.c"
>
>  #define OFFSET(x) offsetof(NLMeansContext, x)
>  #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
> @@ -87,6 +84,13 @@ static int query_formats(AVFilterContext *ctx)
>          AV_PIX_FMT_YUVJ422P, AV_PIX_FMT_YUVJ420P,
>          AV_PIX_FMT_YUVJ411P,
>          AV_PIX_FMT_GRAY8, AV_PIX_FMT_GBRP,
> +        AV_PIX_FMT_YUV420P9, AV_PIX_FMT_YUV422P9, AV_PIX_FMT_YUV444P9,
> +        AV_PIX_FMT_YUV420P10, AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV444P10,
> +        AV_PIX_FMT_YUV440P10,
> +        AV_PIX_FMT_YUV444P12, AV_PIX_FMT_YUV422P12, AV_PIX_FMT_YUV420P12,
> +        AV_PIX_FMT_YUV440P12,
> +        AV_PIX_FMT_YUV444P14, AV_PIX_FMT_YUV422P14, AV_PIX_FMT_YUV420P14,
> +        AV_PIX_FMT_YUV420P16, AV_PIX_FMT_YUV422P16, AV_PIX_FMT_YUV444P16,
>          AV_PIX_FMT_NONE
>      };
>
> @@ -96,380 +100,6 @@ static int query_formats(AVFilterContext *ctx)
>      return ff_set_common_formats(ctx, fmts_list);
>  }
>
> -/**
> - * Compute squared difference of the safe area (the zone where s1 and s2
> - * overlap). It is likely the largest integral zone, so it is interesting to do
> - * as little checks as possible; contrary to the unsafe version of this
> - * function, we do not need any clipping here.
> - *
> - * The line above dst and the column to its left are always readable.
> - */
> -static void compute_safe_ssd_integral_image_c(uint32_t *dst, ptrdiff_t dst_linesize_32,
> -                                              const uint8_t *s1, ptrdiff_t linesize1,
> -                                              const uint8_t *s2, ptrdiff_t linesize2,
> -                                              int w, int h)
> -{
> -    int x, y;
> -    const uint32_t *dst_top = dst - dst_linesize_32;
> -
> -    /* SIMD-friendly assumptions allowed here */
> -    av_assert2(!(w & 0xf) && w >= 16 && h >= 1);
> -
> -    for (y = 0; y < h; y++) {
> -        for (x = 0; x < w; x += 4) {
> -            const int d0 = s1[x    ] - s2[x    ];
> -            const int d1 = s1[x + 1] - s2[x + 1];
> -            const int d2 = s1[x + 2] - s2[x + 2];
> -            const int d3 = s1[x + 3] - s2[x + 3];
> -
> -            dst[x    ] = dst_top[x    ] - dst_top[x - 1] + d0*d0;
> -            dst[x + 1] = dst_top[x + 1] - dst_top[x    ] + d1*d1;
> -            dst[x + 2] = dst_top[x + 2] - dst_top[x + 1] + d2*d2;
> -            dst[x + 3] = dst_top[x + 3] - dst_top[x + 2] + d3*d3;
> -
> -            dst[x    ] += dst[x - 1];
> -            dst[x + 1] += dst[x    ];
> -            dst[x + 2] += dst[x + 1];
> -            dst[x + 3] += dst[x + 2];
> -        }
> -        s1  += linesize1;
> -        s2  += linesize2;
> -        dst += dst_linesize_32;
> -        dst_top += dst_linesize_32;
> -    }
> -}
> -
> -/**
> - * Compute squared difference of an unsafe area (the zone nor s1 nor s2 could
> - * be readable).
> - *
> - * On the other hand, the line above dst and the column to its left are always
> - * readable.
> - *
> - * There is little point in having this function SIMDified as it is likely too
> - * complex and only handle small portions of the image.
> - *
> - * @param dst               integral image
> - * @param dst_linesize_32   integral image linesize (in 32-bit integers unit)
> - * @param startx            integral starting x position
> - * @param starty            integral starting y position
> - * @param src               source plane buffer
> - * @param linesize          source plane linesize
> - * @param offx              source offsetting in x
> - * @param offy              source offsetting in y
> - * @paran r                 absolute maximum source offsetting
> - * @param sw                source width
> - * @param sh                source height
> - * @param w                 width to compute
> - * @param h                 height to compute
> - */
> -static inline void compute_unsafe_ssd_integral_image(uint32_t *dst, ptrdiff_t dst_linesize_32,
> -                                                     int startx, int starty,
> -                                                     const uint8_t *src, ptrdiff_t linesize,
> -                                                     int offx, int offy, int r, int sw, int sh,
> -                                                     int w, int h)
> -{
> -    int x, y;
> -
> -    for (y = starty; y < starty + h; y++) {
> -        uint32_t acc = dst[y*dst_linesize_32 + startx - 1] - dst[(y-1)*dst_linesize_32 + startx - 1];
> -        const int s1y = av_clip(y -  r,         0, sh - 1);
> -        const int s2y = av_clip(y - (r + offy), 0, sh - 1);
> -
> -        for (x = startx; x < startx + w; x++) {
> -            const int s1x = av_clip(x -  r,         0, sw - 1);
> -            const int s2x = av_clip(x - (r + offx), 0, sw - 1);
> -            const uint8_t v1 = src[s1y*linesize + s1x];
> -            const uint8_t v2 = src[s2y*linesize + s2x];
> -            const int d = v1 - v2;
> -            acc += d * d;
> -            dst[y*dst_linesize_32 + x] = dst[(y-1)*dst_linesize_32 + x] + acc;
> -        }
> -    }
> -}
> -
> -/*
> - * Compute the sum of squared difference integral image
> - * http://www.ipol.im/pub/art/2014/57/
> - * Integral Images for Block Matching - Gabriele Facciolo, Nicolas Limare, Enric Meinhardt-Llopis
> - *
> - * @param ii                integral image of dimension (w+e*2) x (h+e*2) with
> - *                          an additional zeroed top line and column already
> - *                          "applied" to the pointer value
> - * @param ii_linesize_32    integral image linesize (in 32-bit integers unit)
> - * @param src               source plane buffer
> - * @param linesize          source plane linesize
> - * @param offx              x-offsetting ranging in [-e;e]
> - * @param offy              y-offsetting ranging in [-e;e]
> - * @param w                 source width
> - * @param h                 source height
> - * @param e                 research padding edge
> - */
> -static void compute_ssd_integral_image(const NLMeansDSPContext *dsp,
> -                                       uint32_t *ii, ptrdiff_t ii_linesize_32,
> -                                       const uint8_t *src, ptrdiff_t linesize, int offx, int offy,
> -                                       int e, int w, int h)
> -{
> -    // ii has a surrounding padding of thickness "e"
> -    const int ii_w = w + e*2;
> -    const int ii_h = h + e*2;
> -
> -    // we center the first source
> -    const int s1x = e;
> -    const int s1y = e;
> -
> -    // 2nd source is the frame with offsetting
> -    const int s2x = e + offx;
> -    const int s2y = e + offy;
> -
> -    // get the dimension of the overlapping rectangle where it is always safe
> -    // to compare the 2 sources pixels
> -    const int startx_safe = FFMAX(s1x, s2x);
> -    const int starty_safe = FFMAX(s1y, s2y);
> -    const int u_endx_safe = FFMIN(s1x + w, s2x + w); // unaligned
> -    const int endy_safe   = FFMIN(s1y + h, s2y + h);
> -
> -    // deduce the safe area width and height
> -    const int safe_pw = (u_endx_safe - startx_safe) & ~0xf;
> -    const int safe_ph = endy_safe - starty_safe;
> -
> -    // adjusted end x position of the safe area after width of the safe area gets aligned
> -    const int endx_safe = startx_safe + safe_pw;
> -
> -    // top part where only one of s1 and s2 is still readable, or none at all
> -    compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
> -                                      0, 0,
> -                                      src, linesize,
> -                                      offx, offy, e, w, h,
> -                                      ii_w, starty_safe);
> -
> -    // fill the left column integral required to compute the central
> -    // overlapping one
> -    compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
> -                                      0, starty_safe,
> -                                      src, linesize,
> -                                      offx, offy, e, w, h,
> -                                      startx_safe, safe_ph);
> -
> -    // main and safe part of the integral
> -    av_assert1(startx_safe - s1x >= 0); av_assert1(startx_safe - s1x < w);
> -    av_assert1(starty_safe - s1y >= 0); av_assert1(starty_safe - s1y < h);
> -    av_assert1(startx_safe - s2x >= 0); av_assert1(startx_safe - s2x < w);
> -    av_assert1(starty_safe - s2y >= 0); av_assert1(starty_safe - s2y < h);
> -    if (safe_pw && safe_ph)
> -        dsp->compute_safe_ssd_integral_image(ii + starty_safe*ii_linesize_32 + startx_safe, ii_linesize_32,
> -                                             src + (starty_safe - s1y) * linesize + (startx_safe - s1x), linesize,
> -                                             src + (starty_safe - s2y) * linesize + (startx_safe - s2x), linesize,
> -                                             safe_pw, safe_ph);
> -
> -    // right part of the integral
> -    compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
> -                                      endx_safe, starty_safe,
> -                                      src, linesize,
> -                                      offx, offy, e, w, h,
> -                                      ii_w - endx_safe, safe_ph);
> -
> -    // bottom part where only one of s1 and s2 is still readable, or none at all
> -    compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
> -                                      0, endy_safe,
> -                                      src, linesize,
> -                                      offx, offy, e, w, h,
> -                                      ii_w, ii_h - endy_safe);
> -}
> -
> -static int config_input(AVFilterLink *inlink)
> -{
> -    AVFilterContext *ctx = inlink->dst;
> -    NLMeansContext *s = ctx->priv;
> -    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
> -    const int e = FFMAX(s->research_hsize, s->research_hsize_uv)
> -                + FFMAX(s->patch_hsize,    s->patch_hsize_uv);
> -
> -    s->chroma_w = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
> -    s->chroma_h = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
> -    s->nb_planes = av_pix_fmt_count_planes(inlink->format);
> -
> -    /* Allocate the integral image with extra edges of thickness "e"
> -     *
> -     *   +_+-------------------------------+
> -     *   |0|0000000000000000000000000000000|
> -     *   +-x-------------------------------+
> -     *   |0|\    ^                         |
> -     *   |0| ii  | e                       |
> -     *   |0|     v                         |
> -     *   |0|   +-----------------------+   |
> -     *   |0|   |                       |   |
> -     *   |0|<->|                       |   |
> -     *   |0| e |                       |   |
> -     *   |0|   |                       |   |
> -     *   |0|   +-----------------------+   |
> -     *   |0|                               |
> -     *   |0|                               |
> -     *   |0|                               |
> -     *   +-+-------------------------------+
> -     */
> -    s->ii_w = inlink->w + e*2;
> -    s->ii_h = inlink->h + e*2;
> -
> -    // align to 4 the linesize, "+1" is for the space of the left 0-column
> -    s->ii_lz_32 = FFALIGN(s->ii_w + 1, 4);
> -
> -    // "+1" is for the space of the top 0-line
> -    s->ii_orig = av_mallocz_array(s->ii_h + 1, s->ii_lz_32 * sizeof(*s->ii_orig));
> -    if (!s->ii_orig)
> -        return AVERROR(ENOMEM);
> -
> -    // skip top 0-line and left 0-column
> -    s->ii = s->ii_orig + s->ii_lz_32 + 1;
> -
> -    // allocate weighted average for every pixel
> -    s->wa_linesize = inlink->w;
> -    s->wa = av_malloc_array(s->wa_linesize, inlink->h * sizeof(*s->wa));
> -    if (!s->wa)
> -        return AVERROR(ENOMEM);
> -
> -    return 0;
> -}
> -
> -struct thread_data {
> -    const uint8_t *src;
> -    ptrdiff_t src_linesize;
> -    int startx, starty;
> -    int endx, endy;
> -    const uint32_t *ii_start;
> -    int p;
> -};
> -
> -static int nlmeans_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
> -{
> -    int x, y;
> -    NLMeansContext *s = ctx->priv;
> -    const struct thread_data *td = arg;
> -    const ptrdiff_t src_linesize = td->src_linesize;
> -    const int process_h = td->endy - td->starty;
> -    const int slice_start = (process_h *  jobnr   ) / nb_jobs;
> -    const int slice_end   = (process_h * (jobnr+1)) / nb_jobs;
> -    const int starty = td->starty + slice_start;
> -    const int endy   = td->starty + slice_end;
> -    const int p = td->p;
> -    const uint32_t *ii = td->ii_start + (starty - p - 1) * s->ii_lz_32 - p - 1;
> -    const int dist_b = 2*p + 1;
> -    const int dist_d = dist_b * s->ii_lz_32;
> -    const int dist_e = dist_d + dist_b;
> -
> -    for (y = starty; y < endy; y++) {
> -        const uint8_t *src = td->src + y*src_linesize;
> -        struct weighted_avg *wa = s->wa + y*s->wa_linesize;
> -        for (x = td->startx; x < td->endx; x++) {
> -            /*
> -             * M is a discrete map where every entry contains the sum of all the entries
> -             * in the rectangle from the top-left origin of M to its coordinate. In the
> -             * following schema, "i" contains the sum of the whole map:
> -             *
> -             * M = +----------+-----------------+----+
> -             *     |          |                 |    |
> -             *     |          |                 |    |
> -             *     |         a|                b|   c|
> -             *     +----------+-----------------+----+
> -             *     |          |                 |    |
> -             *     |          |                 |    |
> -             *     |          |        X        |    |
> -             *     |          |                 |    |
> -             *     |         d|                e|   f|
> -             *     +----------+-----------------+----+
> -             *     |          |                 |    |
> -             *     |         g|                h|   i|
> -             *     +----------+-----------------+----+
> -             *
> -             * The sum of the X box can be calculated with:
> -             *    X = e-d-b+a
> -             *
> -             * See https://en.wikipedia.org/wiki/Summed_area_table
> -             *
> -             * The compute*_ssd functions compute the integral image M where every entry
> -             * contains the sum of the squared difference of every corresponding pixels of
> -             * two input planes of the same size as M.
> -             */
> -            const uint32_t a = ii[x];
> -            const uint32_t b = ii[x + dist_b];
> -            const uint32_t d = ii[x + dist_d];
> -            const uint32_t e = ii[x + dist_e];
> -            const uint32_t patch_diff_sq = e - d - b + a;
> -
> -            if (patch_diff_sq < s->max_meaningful_diff) {
> -                const float weight = s->weight_lut[patch_diff_sq]; // exp(-patch_diff_sq * s->pdiff_scale)
> -                wa[x].total_weight += weight;
> -                wa[x].sum += weight * src[x];
> -            }
> -        }
> -        ii += s->ii_lz_32;
> -    }
> -    return 0;
> -}
> -
> -static void weight_averages(uint8_t *dst, ptrdiff_t dst_linesize,
> -                            const uint8_t *src, ptrdiff_t src_linesize,
> -                            struct weighted_avg *wa, ptrdiff_t wa_linesize,
> -                            int w, int h)
> -{
> -    int x, y;
> -
> -    for (y = 0; y < h; y++) {
> -        for (x = 0; x < w; x++) {
> -            // Also weight the centered pixel
> -            wa[x].total_weight += 1.f;
> -            wa[x].sum += 1.f * src[x];
> -            dst[x] = av_clip_uint8(wa[x].sum / wa[x].total_weight + 0.5f);
> -        }
> -        dst += dst_linesize;
> -        src += src_linesize;
> -        wa += wa_linesize;
> -    }
> -}
> -
> -static int nlmeans_plane(AVFilterContext *ctx, int w, int h, int p, int r,
> -                         uint8_t *dst, ptrdiff_t dst_linesize,
> -                         const uint8_t *src, ptrdiff_t src_linesize)
> -{
> -    int offx, offy;
> -    NLMeansContext *s = ctx->priv;
> -    /* patches center points cover the whole research window so the patches
> -     * themselves overflow the research window */
> -    const int e = r + p;
> -    /* focus an integral pointer on the centered image (s1) */
> -    const uint32_t *centered_ii = s->ii + e*s->ii_lz_32 + e;
> -
> -    memset(s->wa, 0, s->wa_linesize * h * sizeof(*s->wa));
> -
> -    for (offy = -r; offy <= r; offy++) {
> -        for (offx = -r; offx <= r; offx++) {
> -            if (offx || offy) {
> -                struct thread_data td = {
> -                    .src          = src + offy*src_linesize + offx,
> -                    .src_linesize = src_linesize,
> -                    .startx       = FFMAX(0, -offx),
> -                    .starty       = FFMAX(0, -offy),
> -                    .endx         = FFMIN(w, w - offx),
> -                    .endy         = FFMIN(h, h - offy),
> -                    .ii_start     = centered_ii + offy*s->ii_lz_32 + offx,
> -                    .p            = p,
> -                };
> -
> -                compute_ssd_integral_image(&s->dsp, s->ii, s->ii_lz_32,
> -                                           src, src_linesize,
> -                                           offx, offy, e, w, h);
> -                ctx->internal->execute(ctx, nlmeans_slice, &td, NULL,
> -                                       FFMIN(td.endy - td.starty, ff_filter_get_nb_threads(ctx)));
> -            }
> -        }
> -    }
> -
> -    weight_averages(dst, dst_linesize, src, src_linesize,
> -                    s->wa, s->wa_linesize, w, h);
> -
> -    return 0;
> -}
> -
>  static int filter_frame(AVFilterLink *inlink, AVFrame *in)
>  {
>      int i;
> @@ -489,15 +119,30 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in)
>          const int h = i ? s->chroma_h          : inlink->h;
>          const int p = i ? s->patch_hsize_uv    : s->patch_hsize;
>          const int r = i ? s->research_hsize_uv : s->research_hsize;
> -        nlmeans_plane(ctx, w, h, p, r,
> -                      out->data[i], out->linesize[i],
> -                      in->data[i],  in->linesize[i]);
> +        s->nlmeans_plane(ctx, w, h, p, r,
> +                         out->data[i], out->linesize[i],
> +                         in->data[i],  in->linesize[i]);
>      }
>
>      av_frame_free(&in);
>      return ff_filter_frame(outlink, out);
>  }
>
> +void ff_nlmeans_init(NLMeansDSPContext *dsp, int depth)
> +{
> +    switch (depth) {
> +    case  8: dsp->compute_safe_ssd_integral_image32 = compute_safe_ssd_integral_image_c8;  break;
> +    case  9: dsp->compute_safe_ssd_integral_image64 = compute_safe_ssd_integral_image_c9;  break;
> +    case 10: dsp->compute_safe_ssd_integral_image64 = compute_safe_ssd_integral_image_c10; break;
> +    case 12: dsp->compute_safe_ssd_integral_image64 = compute_safe_ssd_integral_image_c12; break;
> +    case 14: dsp->compute_safe_ssd_integral_image64 = compute_safe_ssd_integral_image_c14; break;
> +    case 16: dsp->compute_safe_ssd_integral_image64 = compute_safe_ssd_integral_image_c16; break;
> +    }
> +
> +    if (ARCH_AARCH64)
> +        ff_nlmeans_init_aarch64(dsp, depth);
> +}
> +
>  #define CHECK_ODD_FIELD(field, name) do {                       \
>      if (!(s->field & 1)) {                                      \
>          s->field |= 1;                                          \
> @@ -506,26 +151,22 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in)
>      }                                                           \
>  } while (0)
>
> -void ff_nlmeans_init(NLMeansDSPContext *dsp)
> -{
> -    dsp->compute_safe_ssd_integral_image = compute_safe_ssd_integral_image_c;
> -
> -    if (ARCH_AARCH64)
> -        ff_nlmeans_init_aarch64(dsp);
> -}
> -
> -static av_cold int init(AVFilterContext *ctx)
> +static int config_input(AVFilterLink *inlink)
>  {
> -    int i;
> +    AVFilterContext *ctx = inlink->dst;
>      NLMeansContext *s = ctx->priv;
> -    const double h = s->sigma * 10.;
> +    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
> +    double h;
> +    int e;
>
> +    s->depth = desc->comp[0].depth;
> +    h = s->sigma * 10. * (1 << (s->depth - 8));
>      s->pdiff_scale = 1. / (h * h);
> -    s->max_meaningful_diff = log(255.) / s->pdiff_scale;
> +    s->max_meaningful_diff = FFMIN(log(255.) / s->pdiff_scale, INT32_MAX / 8);
>      s->weight_lut = av_calloc(s->max_meaningful_diff, sizeof(*s->weight_lut));
>      if (!s->weight_lut)
>          return AVERROR(ENOMEM);
> -    for (i = 0; i < s->max_meaningful_diff; i++)
> +    for (int i = 0; i < s->max_meaningful_diff; i++)
>          s->weight_lut[i] = exp(-i * s->pdiff_scale);
>
>      CHECK_ODD_FIELD(research_size,   "Luma research window");
> @@ -542,11 +183,75 @@ static av_cold int init(AVFilterContext *ctx)
>      s->patch_hsize       = s->patch_size       / 2;
>      s->patch_hsize_uv    = s->patch_size_uv    / 2;
>
> +    e = FFMAX(s->research_hsize, s->research_hsize_uv) +
> +        FFMAX(s->patch_hsize,    s->patch_hsize_uv);
>      av_log(ctx, AV_LOG_INFO, "Research window: %dx%d / %dx%d, patch size: %dx%d / %dx%d\n",
>             s->research_size, s->research_size, s->research_size_uv, s->research_size_uv,
>             s->patch_size,    s->patch_size,    s->patch_size_uv,    s->patch_size_uv);
>
> -    ff_nlmeans_init(&s->dsp);
> +    ff_nlmeans_init(&s->dsp, s->depth);
> +
> +    s->chroma_w = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
> +    s->chroma_h = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
> +    s->nb_planes = av_pix_fmt_count_planes(inlink->format);
> +
> +    /* Allocate the integral image with extra edges of thickness "e"
> +     *
> +     *   +_+-------------------------------+
> +     *   |0|0000000000000000000000000000000|
> +     *   +-x-------------------------------+
> +     *   |0|\    ^                         |
> +     *   |0| ii  | e                       |
> +     *   |0|     v                         |
> +     *   |0|   +-----------------------+   |
> +     *   |0|   |                       |   |
> +     *   |0|<->|                       |   |
> +     *   |0| e |                       |   |
> +     *   |0|   |                       |   |
> +     *   |0|   +-----------------------+   |
> +     *   |0|                               |
> +     *   |0|                               |
> +     *   |0|                               |
> +     *   +-+-------------------------------+
> +     */
> +    s->ii_w = inlink->w + e*2;
> +    s->ii_h = inlink->h + e*2;
> +
> +    // align to 4 the linesize, "+1" is for the space of the left 0-column
> +    s->ii_lz_32 = FFALIGN(s->ii_w + 1, 4);
> +
> +    if (s->depth <= 8) {
> +        // "+1" is for the space of the top 0-line
> +        s->ii_orig32 = av_mallocz_array(s->ii_h + 1, s->ii_lz_32 * sizeof(*s->ii_orig32));
> +        if (!s->ii_orig32)
> +            return AVERROR(ENOMEM);
> +
> +        // skip top 0-line and left 0-column
> +        s->ii32 = s->ii_orig32 + s->ii_lz_32 + 1;
> +    } else {
> +        // "+1" is for the space of the top 0-line
> +        s->ii_orig64 = av_mallocz_array(s->ii_h + 1, s->ii_lz_32 * sizeof(*s->ii_orig64));
> +        if (!s->ii_orig64)
> +            return AVERROR(ENOMEM);
> +
> +        // skip top 0-line and left 0-column
> +        s->ii64 = s->ii_orig64 + s->ii_lz_32 + 1;
> +    }
> +
> +    // allocate weighted average for every pixel
> +    s->wa_linesize = inlink->w;
> +    s->wa = av_malloc_array(s->wa_linesize, inlink->h * sizeof(*s->wa));
> +    if (!s->wa)
> +        return AVERROR(ENOMEM);
> +
> +    switch (s->depth) {
> +    case  8: s->nlmeans_plane = nlmeans_plane8;  break;
> +    case  9: s->nlmeans_plane = nlmeans_plane9;  break;
> +    case 10: s->nlmeans_plane = nlmeans_plane10; break;
> +    case 12: s->nlmeans_plane = nlmeans_plane12; break;
> +    case 14: s->nlmeans_plane = nlmeans_plane14; break;
> +    case 16: s->nlmeans_plane = nlmeans_plane16; break;
> +    }
>
>      return 0;
>  }
> @@ -555,7 +260,8 @@ static av_cold void uninit(AVFilterContext *ctx)
>  {
>      NLMeansContext *s = ctx->priv;
>      av_freep(&s->weight_lut);
> -    av_freep(&s->ii_orig);
> +    av_freep(&s->ii_orig32);
> +    av_freep(&s->ii_orig64);
>      av_freep(&s->wa);
>  }
>
> @@ -581,7 +287,6 @@ AVFilter ff_vf_nlmeans = {
>      .name          = "nlmeans",
>      .description   = NULL_IF_CONFIG_SMALL("Non-local means denoiser."),
>      .priv_size     = sizeof(NLMeansContext),
> -    .init          = init,
>      .uninit        = uninit,
>      .query_formats = query_formats,
>      .inputs        = nlmeans_inputs,
> diff --git a/libavfilter/vf_nlmeans.h b/libavfilter/vf_nlmeans.h
> index 0a9aab2928..814b95e10f 100644
> --- a/libavfilter/vf_nlmeans.h
> +++ b/libavfilter/vf_nlmeans.h
> @@ -23,13 +23,60 @@
>  #include <stdint.h>
>
>  typedef struct NLMeansDSPContext {
> -    void (*compute_safe_ssd_integral_image)(uint32_t *dst, ptrdiff_t dst_linesize_32,
> -                                            const uint8_t *s1, ptrdiff_t linesize1,
> -                                            const uint8_t *s2, ptrdiff_t linesize2,
> -                                            int w, int h);
> +    void (*compute_safe_ssd_integral_image32)(uint32_t *dst, ptrdiff_t dst_linesize_32,
> +                                              const uint8_t *s1, ptrdiff_t linesize1,
> +                                              const uint8_t *s2, ptrdiff_t linesize2,
> +                                              int w, int h);
> +    void (*compute_safe_ssd_integral_image64)(uint64_t *dst, ptrdiff_t dst_linesize_32,
> +                                              const uint8_t *s1, ptrdiff_t linesize1,
> +                                              const uint8_t *s2, ptrdiff_t linesize2,
> +                                              int w, int h);
>  } NLMeansDSPContext;
>
> -void ff_nlmeans_init(NLMeansDSPContext *dsp);
> -void ff_nlmeans_init_aarch64(NLMeansDSPContext *dsp);
> +struct weighted_avg {
> +    float total_weight;
> +    float sum;
> +};
> +
> +struct thread_data {
> +    const uint8_t *src;
> +    ptrdiff_t src_linesize;
> +    int startx, starty;
> +    int endx, endy;
> +    const uint32_t *ii_start32;
> +    const uint64_t *ii_start64;
> +    int p;
> +};
> +
> +typedef struct NLMeansContext {
> +    const AVClass *class;
> +    int nb_planes;
> +    int depth;
> +    int chroma_w, chroma_h;
> +    double pdiff_scale;                         // invert of the filtering parameter (sigma*10) squared
> +    double sigma;                               // denoising strength
> +    int patch_size,    patch_hsize;             // patch size and half size
> +    int patch_size_uv, patch_hsize_uv;          // patch size and half size for chroma planes
> +    int research_size,    research_hsize;       // research size and half size
> +    int research_size_uv, research_hsize_uv;    // research size and half size for chroma planes
> +    uint32_t *ii_orig32;                        // integral image 32-bit
> +    uint32_t *ii32;                             // integral image 32-bit starting after the 0-line and 0-column
> +    uint64_t *ii_orig64;                        // integral image 64-bit
> +    uint64_t *ii64;                             // integral image 64-bit starting after the 0-line and 0-column
> +    int ii_w, ii_h;                             // width and height of the integral image
> +    ptrdiff_t ii_lz_32;                         // linesize in 32-bit units of the integral image
> +    struct weighted_avg *wa;                    // weighted average of every pixel
> +    ptrdiff_t wa_linesize;                      // linesize for wa in struct size unit
> +    float *weight_lut;                          // lookup table mapping (scaled) patch differences to their associated weights
> +    uint32_t max_meaningful_diff;               // maximum difference considered (if the patch difference is too high we ignore the pixel)
> +    NLMeansDSPContext dsp;
> +
> +    int (*nlmeans_plane)(AVFilterContext *ctx, int w, int h, int p, int r,
> +                         uint8_t *dst, ptrdiff_t dst_linesize,
> +                         const uint8_t *src, ptrdiff_t src_linesize);
> +} NLMeansContext;
> +
> +void ff_nlmeans_init(NLMeansDSPContext *dsp, int depth);
> +void ff_nlmeans_init_aarch64(NLMeansDSPContext *dsp, int depth);
>
>  #endif /* AVFILTER_NLMEANS_H */
> --
LGTM now
Michael Niedermayer Nov. 21, 2019, 3:47 p.m.
On Wed, Nov 20, 2019 at 10:54:42AM +0100, Paul B Mahol wrote:
> Signed-off-by: Paul B Mahol <onemda@gmail.com>
> ---
>  libavfilter/aarch64/vf_nlmeans_init.c |   6 +-
>  libavfilter/nlmeans_template.c        | 370 ++++++++++++++++++
>  libavfilter/vf_nlmeans.c              | 539 ++++++--------------------
>  libavfilter/vf_nlmeans.h              |  59 ++-
>  4 files changed, 548 insertions(+), 426 deletions(-)
>  create mode 100644 libavfilter/nlmeans_template.c

breaks:

make testprogs 
CC	libavfilter/tests/integral.o
libavfilter/tests/integral.c: In function ‘main’:
libavfilter/tests/integral.c:66:5: error: too few arguments to function ‘ff_nlmeans_init’
     ff_nlmeans_init(&dsp);
     ^
In file included from libavfilter/tests/integral.c:19:0:
./libavfilter/vf_nlmeans.c:131:6: note: declared here
 void ff_nlmeans_init(NLMeansDSPContext *dsp, int depth)
      ^
libavfilter/tests/integral.c:72:13: error: implicit declaration of function ‘compute_ssd_integral_image’ [-Werror=implicit-function-declaration]
             compute_ssd_integral_image(&dsp, ii_start, ii_lz_32,
             ^
libavfilter/tests/integral.c:76:13: error: implicit declaration of function ‘compute_unsafe_ssd_integral_image’ [-Werror=implicit-function-declaration]
             compute_unsafe_ssd_integral_image(ii_start2, ii_lz_32,
             ^
cc1: some warnings being treated as errors
make: *** [libavfilter/tests/integral.o] Error 1
make: Target `testprogs' not remade because of errors.


[...]
Paul B Mahol Nov. 21, 2019, 6:08 p.m.
Easy to fix, do you have another comment?

On 11/21/19, Michael Niedermayer <michael@niedermayer.cc> wrote:
> On Wed, Nov 20, 2019 at 10:54:42AM +0100, Paul B Mahol wrote:
>> Signed-off-by: Paul B Mahol <onemda@gmail.com>
>> ---
>>  libavfilter/aarch64/vf_nlmeans_init.c |   6 +-
>>  libavfilter/nlmeans_template.c        | 370 ++++++++++++++++++
>>  libavfilter/vf_nlmeans.c              | 539 ++++++--------------------
>>  libavfilter/vf_nlmeans.h              |  59 ++-
>>  4 files changed, 548 insertions(+), 426 deletions(-)
>>  create mode 100644 libavfilter/nlmeans_template.c
>
> breaks:
>
> make testprogs
> CC	libavfilter/tests/integral.o
> libavfilter/tests/integral.c: In function ‘main’:
> libavfilter/tests/integral.c:66:5: error: too few arguments to function
> ‘ff_nlmeans_init’
>      ff_nlmeans_init(&dsp);
>      ^
> In file included from libavfilter/tests/integral.c:19:0:
> ./libavfilter/vf_nlmeans.c:131:6: note: declared here
>  void ff_nlmeans_init(NLMeansDSPContext *dsp, int depth)
>       ^
> libavfilter/tests/integral.c:72:13: error: implicit declaration of function
> ‘compute_ssd_integral_image’ [-Werror=implicit-function-declaration]
>              compute_ssd_integral_image(&dsp, ii_start, ii_lz_32,
>              ^
> libavfilter/tests/integral.c:76:13: error: implicit declaration of function
> ‘compute_unsafe_ssd_integral_image’ [-Werror=implicit-function-declaration]
>              compute_unsafe_ssd_integral_image(ii_start2, ii_lz_32,
>              ^
> cc1: some warnings being treated as errors
> make: *** [libavfilter/tests/integral.o] Error 1
> make: Target `testprogs' not remade because of errors.
>
>
> [...]
> --
> Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
>
> Why not whip the teacher when the pupil misbehaves? -- Diogenes of Sinope
>
Michael Niedermayer Nov. 21, 2019, 8:30 p.m.
On Thu, Nov 21, 2019 at 07:08:26PM +0100, Paul B Mahol wrote:
> Easy to fix, do you have another comment?

i didnt review the code, just noticed it broke build

Thanks

[...]
Clément Bœsch Dec. 1, 2019, 11:28 a.m.
On Wed, Nov 20, 2019 at 10:54:42AM +0100, Paul B Mahol wrote:
> Signed-off-by: Paul B Mahol <onemda@gmail.com>
> ---
>  libavfilter/aarch64/vf_nlmeans_init.c |   6 +-
>  libavfilter/nlmeans_template.c        | 370 ++++++++++++++++++
>  libavfilter/vf_nlmeans.c              | 539 ++++++--------------------
>  libavfilter/vf_nlmeans.h              |  59 ++-
>  4 files changed, 548 insertions(+), 426 deletions(-)
>  create mode 100644 libavfilter/nlmeans_template.c

On Aarch64:

In file included from src/libavfilter/aarch64/vf_nlmeans_init.c:20:
src/libavfilter/vf_nlmeans.h:52:11: error: unknown type name 'AVClass'
   52 |     const AVClass *class;
      |           ^~~~~~~
In file included from src/libavfilter/aarch64/vf_nlmeans_init.c:20:
src/libavfilter/vf_nlmeans.h:74:26: error: unknown type name 'AVFilterContext'
   74 |     int (*nlmeans_plane)(AVFilterContext *ctx, int w, int h, int p, int r,
      |                          ^~~~~~~~~~~~~~~
src/libavfilter/vf_nlmeans.h:77:1: warning: no semicolon at end of struct or union
   77 | } NLMeansContext;
      | ^
CC	libavfilter/vf_nlmeans.o

Note: not the same issue as the one reported by Michael.

[...]
> -static av_cold int init(AVFilterContext *ctx)
> +static int config_input(AVFilterLink *inlink)
>  {
> -    int i;
> +    AVFilterContext *ctx = inlink->dst;
>      NLMeansContext *s = ctx->priv;
> -    const double h = s->sigma * 10.;
> +    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
> +    double h;
> +    int e;
>  

> +    s->depth = desc->comp[0].depth;
> +    h = s->sigma * 10. * (1 << (s->depth - 8));
>      s->pdiff_scale = 1. / (h * h);
> -    s->max_meaningful_diff = log(255.) / s->pdiff_scale;
> +    s->max_meaningful_diff = FFMIN(log(255.) / s->pdiff_scale, INT32_MAX / 8);

I'm assuming log((1 << s->depth) - 1) / s->pdiff_scale wasn't doing the expected?

Can you elaborate on this FFMIN?

>      s->weight_lut = av_calloc(s->max_meaningful_diff, sizeof(*s->weight_lut));
>      if (!s->weight_lut)
>          return AVERROR(ENOMEM);
> -    for (i = 0; i < s->max_meaningful_diff; i++)
> +    for (int i = 0; i < s->max_meaningful_diff; i++)

Could be split but...

>          s->weight_lut[i] = exp(-i * s->pdiff_scale);
>  
>      CHECK_ODD_FIELD(research_size,   "Luma research window");
> @@ -542,11 +183,75 @@ static av_cold int init(AVFilterContext *ctx)
>      s->patch_hsize       = s->patch_size       / 2;
>      s->patch_hsize_uv    = s->patch_size_uv    / 2;
>  
> +    e = FFMAX(s->research_hsize, s->research_hsize_uv) +
> +        FFMAX(s->patch_hsize,    s->patch_hsize_uv);
>      av_log(ctx, AV_LOG_INFO, "Research window: %dx%d / %dx%d, patch size: %dx%d / %dx%d\n",
>             s->research_size, s->research_size, s->research_size_uv, s->research_size_uv,
>             s->patch_size,    s->patch_size,    s->patch_size_uv,    s->patch_size_uv);

You're moving a lot of code around, I believe the patch could have been split,
it's hard to follow as is.

[...]
> -void ff_nlmeans_init(NLMeansDSPContext *dsp);
> -void ff_nlmeans_init_aarch64(NLMeansDSPContext *dsp);
> +struct weighted_avg {
> +    float total_weight;
> +    float sum;
> +};

No point in switching to double for 9+ bits?

Patch hide | download patch | download mbox

diff --git a/libavfilter/aarch64/vf_nlmeans_init.c b/libavfilter/aarch64/vf_nlmeans_init.c
index a1edefb144..6ee157dfe0 100644
--- a/libavfilter/aarch64/vf_nlmeans_init.c
+++ b/libavfilter/aarch64/vf_nlmeans_init.c
@@ -24,10 +24,10 @@  void ff_compute_safe_ssd_integral_image_neon(uint32_t *dst, ptrdiff_t dst_linesi
                                              const uint8_t *s2, ptrdiff_t linesize2,
                                              int w, int h);
 
-av_cold void ff_nlmeans_init_aarch64(NLMeansDSPContext *dsp)
+av_cold void ff_nlmeans_init_aarch64(NLMeansDSPContext *dsp, int depth)
 {
     int cpu_flags = av_get_cpu_flags();
 
-    if (have_neon(cpu_flags))
-        dsp->compute_safe_ssd_integral_image = ff_compute_safe_ssd_integral_image_neon;
+    if (have_neon(cpu_flags) && depth == 8)
+        dsp->compute_safe_ssd_integral_image32 = ff_compute_safe_ssd_integral_image_neon;
 }
diff --git a/libavfilter/nlmeans_template.c b/libavfilter/nlmeans_template.c
new file mode 100644
index 0000000000..b0d7f7482d
--- /dev/null
+++ b/libavfilter/nlmeans_template.c
@@ -0,0 +1,370 @@ 
+/*
+ * Copyright (c) 2016 Clément Bœsch <u pkh me>
+ *
+ * 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 "avfilter.h"
+#include "formats.h"
+#include "internal.h"
+#include "vf_nlmeans.h"
+#include "video.h"
+
+#undef pixel
+#undef integral
+#undef INTEGRAL_DEPTH
+#if DEPTH <= 8
+#define INTEGRAL_DEPTH 32
+#define pixel uint8_t
+#define integral uint32_t
+#else
+#define INTEGRAL_DEPTH 64
+#define pixel uint16_t
+#define integral uint64_t
+#endif
+
+#define fn3(a,b)   a##b
+#define fn2(a,b)   fn3(a,b)
+#define fn(a)      fn2(a, DEPTH)
+#define ii(a)      fn2(a, INTEGRAL_DEPTH)
+
+/**
+ * Compute squared difference of the safe area (the zone where s1 and s2
+ * overlap). It is likely the largest integral zone, so it is interesting to do
+ * as little checks as possible; contrary to the unsafe version of this
+ * function, we do not need any clipping here.
+ *
+ * The line above dst and the column to its left are always readable.
+ */
+static void fn(compute_safe_ssd_integral_image_c)(integral *dst, ptrdiff_t dst_linesize_32,
+                                                  const uint8_t *ss1, ptrdiff_t linesize1,
+                                                  const uint8_t *ss2, ptrdiff_t linesize2,
+                                                  int w, int h)
+{
+    const pixel *s1 = (const pixel *)ss1;
+    const pixel *s2 = (const pixel *)ss2;
+    int x, y;
+    const integral *dst_top = dst - dst_linesize_32;
+
+    linesize1 /= sizeof(pixel);
+    linesize2 /= sizeof(pixel);
+
+    /* SIMD-friendly assumptions allowed here */
+    av_assert2(!(w & 0xf) && w >= 16 && h >= 1);
+
+    for (y = 0; y < h; y++) {
+        for (x = 0; x < w; x += 4) {
+            const int d0 = s1[x    ] - s2[x    ];
+            const int d1 = s1[x + 1] - s2[x + 1];
+            const int d2 = s1[x + 2] - s2[x + 2];
+            const int d3 = s1[x + 3] - s2[x + 3];
+
+            dst[x    ] = dst_top[x    ] - dst_top[x - 1] + d0*d0;
+            dst[x + 1] = dst_top[x + 1] - dst_top[x    ] + d1*d1;
+            dst[x + 2] = dst_top[x + 2] - dst_top[x + 1] + d2*d2;
+            dst[x + 3] = dst_top[x + 3] - dst_top[x + 2] + d3*d3;
+
+            dst[x    ] += dst[x - 1];
+            dst[x + 1] += dst[x    ];
+            dst[x + 2] += dst[x + 1];
+            dst[x + 3] += dst[x + 2];
+        }
+        s1  += linesize1;
+        s2  += linesize2;
+        dst += dst_linesize_32;
+        dst_top += dst_linesize_32;
+    }
+}
+
+/**
+ * Compute squared difference of an unsafe area (the zone nor s1 nor s2 could
+ * be readable).
+ *
+ * On the other hand, the line above dst and the column to its left are always
+ * readable.
+ *
+ * There is little point in having this function SIMDified as it is likely too
+ * complex and only handle small portions of the image.
+ *
+ * @param dst               integral image
+ * @param dst_linesize_32   integral image linesize (in 32-bit integers unit)
+ * @param startx            integral starting x position
+ * @param starty            integral starting y position
+ * @param ssrc              source plane buffer
+ * @param linesize          source plane linesize
+ * @param offx              source offsetting in x
+ * @param offy              source offsetting in y
+ * @paran r                 absolute maximum source offsetting
+ * @param sw                source width
+ * @param sh                source height
+ * @param w                 width to compute
+ * @param h                 height to compute
+ */
+static inline void fn(compute_unsafe_ssd_integral_image)(integral *dst, ptrdiff_t dst_linesize_32,
+                                                         int startx, int starty,
+                                                         const uint8_t *ssrc, ptrdiff_t linesize,
+                                                         int offx, int offy, int r, int sw, int sh,
+                                                         int w, int h)
+{
+    const pixel *src = (const pixel *)ssrc;
+    int x, y;
+
+    linesize /= sizeof(pixel);
+
+    for (y = starty; y < starty + h; y++) {
+        integral acc = dst[y*dst_linesize_32 + startx - 1] - dst[(y-1)*dst_linesize_32 + startx - 1];
+        const int s1y = av_clip(y -  r,         0, sh - 1);
+        const int s2y = av_clip(y - (r + offy), 0, sh - 1);
+
+        for (x = startx; x < startx + w; x++) {
+            const int s1x = av_clip(x -  r,         0, sw - 1);
+            const int s2x = av_clip(x - (r + offx), 0, sw - 1);
+            const pixel v1 = src[s1y*linesize + s1x];
+            const pixel v2 = src[s2y*linesize + s2x];
+            const int d = v1 - v2;
+            acc += d * d;
+            dst[y*dst_linesize_32 + x] = dst[(y-1)*dst_linesize_32 + x] + acc;
+        }
+    }
+}
+
+/*
+ * Compute the sum of squared difference integral image
+ * http://www.ipol.im/pub/art/2014/57/
+ * Integral Images for Block Matching - Gabriele Facciolo, Nicolas Limare, Enric Meinhardt-Llopis
+ *
+ * @param ii                integral image of dimension (w+e*2) x (h+e*2) with
+ *                          an additional zeroed top line and column already
+ *                          "applied" to the pointer value
+ * @param ii_linesize_32    integral image linesize (in 32-bit integers unit)
+ * @param src               source plane buffer
+ * @param linesize          source plane linesize
+ * @param offx              x-offsetting ranging in [-e;e]
+ * @param offy              y-offsetting ranging in [-e;e]
+ * @param w                 source width
+ * @param h                 source height
+ * @param e                 research padding edge
+ */
+static void fn(compute_ssd_integral_image)(const NLMeansDSPContext *dsp,
+                                           integral *ii, ptrdiff_t ii_linesize_32,
+                                           const uint8_t *src, ptrdiff_t linesize, int offx, int offy,
+                                           int e, int w, int h)
+{
+    // ii has a surrounding padding of thickness "e"
+    const int ii_w = w + e*2;
+    const int ii_h = h + e*2;
+
+    // we center the first source
+    const int s1x = e;
+    const int s1y = e;
+
+    // 2nd source is the frame with offsetting
+    const int s2x = e + offx;
+    const int s2y = e + offy;
+
+    // get the dimension of the overlapping rectangle where it is always safe
+    // to compare the 2 sources pixels
+    const int startx_safe = FFMAX(s1x, s2x);
+    const int starty_safe = FFMAX(s1y, s2y);
+    const int u_endx_safe = FFMIN(s1x + w, s2x + w); // unaligned
+    const int endy_safe   = FFMIN(s1y + h, s2y + h);
+
+    // deduce the safe area width and height
+    const int safe_pw = (u_endx_safe - startx_safe) & ~0xf;
+    const int safe_ph = endy_safe - starty_safe;
+
+    // adjusted end x position of the safe area after width of the safe area gets aligned
+    const int endx_safe = startx_safe + safe_pw;
+
+    // top part where only one of s1 and s2 is still readable, or none at all
+    fn(compute_unsafe_ssd_integral_image)(ii, ii_linesize_32,
+                                          0, 0,
+                                          src, linesize,
+                                          offx, offy, e, w, h,
+                                          ii_w, starty_safe);
+
+    // fill the left column integral required to compute the central
+    // overlapping one
+    fn(compute_unsafe_ssd_integral_image)(ii, ii_linesize_32,
+                                          0, starty_safe,
+                                          src, linesize,
+                                          offx, offy, e, w, h,
+                                          startx_safe, safe_ph);
+
+    // main and safe part of the integral
+    av_assert1(startx_safe - s1x >= 0); av_assert1(startx_safe - s1x < w);
+    av_assert1(starty_safe - s1y >= 0); av_assert1(starty_safe - s1y < h);
+    av_assert1(startx_safe - s2x >= 0); av_assert1(startx_safe - s2x < w);
+    av_assert1(starty_safe - s2y >= 0); av_assert1(starty_safe - s2y < h);
+    if (safe_pw && safe_ph)
+        ii(dsp->compute_safe_ssd_integral_image)(ii + starty_safe*ii_linesize_32 + startx_safe, ii_linesize_32,
+                                                 src + (starty_safe - s1y) * linesize + (startx_safe - s1x) * sizeof(pixel), linesize,
+                                                 src + (starty_safe - s2y) * linesize + (startx_safe - s2x) * sizeof(pixel), linesize,
+                                                 safe_pw, safe_ph);
+
+    // right part of the integral
+    fn(compute_unsafe_ssd_integral_image)(ii, ii_linesize_32,
+                                          endx_safe, starty_safe,
+                                          src, linesize,
+                                          offx, offy, e, w, h,
+                                          ii_w - endx_safe, safe_ph);
+
+    // bottom part where only one of s1 and s2 is still readable, or none at all
+    fn(compute_unsafe_ssd_integral_image)(ii, ii_linesize_32,
+                                          0, endy_safe,
+                                          src, linesize,
+                                          offx, offy, e, w, h,
+                                          ii_w, ii_h - endy_safe);
+}
+
+static int fn(nlmeans_slice)(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
+{
+    int x, y;
+    NLMeansContext *s = ctx->priv;
+    const struct thread_data *td = arg;
+    const ptrdiff_t src_linesize = td->src_linesize;
+    const int process_h = td->endy - td->starty;
+    const int slice_start = (process_h *  jobnr   ) / nb_jobs;
+    const int slice_end   = (process_h * (jobnr+1)) / nb_jobs;
+    const int starty = td->starty + slice_start;
+    const int endy   = td->starty + slice_end;
+    const int p = td->p;
+    const integral *ii = ii(td->ii_start) + (starty - p - 1) * s->ii_lz_32 - p - 1;
+    const int dist_b = 2*p + 1;
+    const int dist_d = dist_b * s->ii_lz_32;
+    const int dist_e = dist_d + dist_b;
+
+    for (y = starty; y < endy; y++) {
+        const pixel *src = (const pixel *)(td->src + y*src_linesize);
+        struct weighted_avg *wa = s->wa + y*s->wa_linesize;
+        for (x = td->startx; x < td->endx; x++) {
+            /*
+             * M is a discrete map where every entry contains the sum of all the entries
+             * in the rectangle from the top-left origin of M to its coordinate. In the
+             * following schema, "i" contains the sum of the whole map:
+             *
+             * M = +----------+-----------------+----+
+             *     |          |                 |    |
+             *     |          |                 |    |
+             *     |         a|                b|   c|
+             *     +----------+-----------------+----+
+             *     |          |                 |    |
+             *     |          |                 |    |
+             *     |          |        X        |    |
+             *     |          |                 |    |
+             *     |         d|                e|   f|
+             *     +----------+-----------------+----+
+             *     |          |                 |    |
+             *     |         g|                h|   i|
+             *     +----------+-----------------+----+
+             *
+             * The sum of the X box can be calculated with:
+             *    X = e-d-b+a
+             *
+             * See https://en.wikipedia.org/wiki/Summed_area_table
+             *
+             * The compute*_ssd functions compute the integral image M where every entry
+             * contains the sum of the squared difference of every corresponding pixels of
+             * two input planes of the same size as M.
+             */
+            const integral a = ii[x];
+            const integral b = ii[x + dist_b];
+            const integral d = ii[x + dist_d];
+            const integral e = ii[x + dist_e];
+            const integral patch_diff_sq = e + a - d - b;
+
+            if (patch_diff_sq < s->max_meaningful_diff) {
+                const float weight = s->weight_lut[patch_diff_sq]; // exp(-patch_diff_sq * s->pdiff_scale)
+                wa[x].total_weight += weight;
+                wa[x].sum += weight * src[x];
+            }
+        }
+        ii += s->ii_lz_32;
+    }
+    return 0;
+}
+
+static void fn(weight_averages)(uint8_t *ddst, ptrdiff_t dst_linesize,
+                                const uint8_t *ssrc, ptrdiff_t src_linesize,
+                                struct weighted_avg *wa, ptrdiff_t wa_linesize,
+                                int w, int h)
+{
+    const pixel *src = (const pixel *)ssrc;
+    pixel *dst = (pixel *)ddst;
+    int x, y;
+
+    src_linesize /= sizeof(pixel);
+    dst_linesize /= sizeof(pixel);
+
+    for (y = 0; y < h; y++) {
+        for (x = 0; x < w; x++) {
+            // Also weight the centered pixel
+            wa[x].total_weight += 1.f;
+            wa[x].sum += 1.f * src[x];
+            dst[x] = av_clip_uintp2(wa[x].sum / wa[x].total_weight + 0.5f, DEPTH);
+        }
+        dst += dst_linesize;
+        src += src_linesize;
+        wa += wa_linesize;
+    }
+}
+
+static int fn(nlmeans_plane)(AVFilterContext *ctx, int w, int h, int p, int r,
+                             uint8_t *dst, ptrdiff_t dst_linesize,
+                             const uint8_t *src, ptrdiff_t src_linesize)
+{
+    int offx, offy;
+    NLMeansContext *s = ctx->priv;
+    /* patches center points cover the whole research window so the patches
+     * themselves overflow the research window */
+    const int e = r + p;
+    /* focus an integral pointer on the centered image (s1) */
+    const uint32_t *centered_ii32 = s->ii32 + e*s->ii_lz_32 + e;
+    const uint64_t *centered_ii64 = s->ii64 + e*s->ii_lz_32 + e;
+
+    memset(s->wa, 0, s->wa_linesize * h * sizeof(*s->wa));
+
+    for (offy = -r; offy <= r; offy++) {
+        for (offx = -r; offx <= r; offx++) {
+            if (offx || offy) {
+                struct thread_data td = {
+                    .src          = src + offy*src_linesize + offx * (int)(sizeof(pixel)),
+                    .src_linesize = src_linesize,
+                    .startx       = FFMAX(0, -offx),
+                    .starty       = FFMAX(0, -offy),
+                    .endx         = FFMIN(w, w - offx),
+                    .endy         = FFMIN(h, h - offy),
+                    .ii_start32   = centered_ii32 + offy*s->ii_lz_32 + offx,
+                    .ii_start64   = centered_ii64 + offy*s->ii_lz_32 + offx,
+                    .p            = p,
+                };
+
+                fn(compute_ssd_integral_image)(&s->dsp, ii(s->ii), s->ii_lz_32,
+                                               src, src_linesize,
+                                               offx, offy, e, w, h);
+                ctx->internal->execute(ctx, fn(nlmeans_slice), &td, NULL,
+                                       FFMIN(td.endy - td.starty, ff_filter_get_nb_threads(ctx)));
+            }
+        }
+    }
+
+    fn(weight_averages)(dst, dst_linesize, src, src_linesize,
+                        s->wa, s->wa_linesize, w, h);
+
+    return 0;
+}
diff --git a/libavfilter/vf_nlmeans.c b/libavfilter/vf_nlmeans.c
index 06233b0dd4..bf7cf3aa0e 100644
--- a/libavfilter/vf_nlmeans.c
+++ b/libavfilter/vf_nlmeans.c
@@ -38,31 +38,28 @@ 
 #include "vf_nlmeans.h"
 #include "video.h"
 
-struct weighted_avg {
-    float total_weight;
-    float sum;
-};
+#define DEPTH 8
+#include "nlmeans_template.c"
+
+#undef DEPTH
+#define DEPTH 9
+#include "nlmeans_template.c"
+
+#undef DEPTH
+#define DEPTH 10
+#include "nlmeans_template.c"
+
+#undef DEPTH
+#define DEPTH 12
+#include "nlmeans_template.c"
+
+#undef DEPTH
+#define DEPTH 14
+#include "nlmeans_template.c"
 
-typedef struct NLMeansContext {
-    const AVClass *class;
-    int nb_planes;
-    int chroma_w, chroma_h;
-    double pdiff_scale;                         // invert of the filtering parameter (sigma*10) squared
-    double sigma;                               // denoising strength
-    int patch_size,    patch_hsize;             // patch size and half size
-    int patch_size_uv, patch_hsize_uv;          // patch size and half size for chroma planes
-    int research_size,    research_hsize;       // research size and half size
-    int research_size_uv, research_hsize_uv;    // research size and half size for chroma planes
-    uint32_t *ii_orig;                          // integral image
-    uint32_t *ii;                               // integral image starting after the 0-line and 0-column
-    int ii_w, ii_h;                             // width and height of the integral image
-    ptrdiff_t ii_lz_32;                         // linesize in 32-bit units of the integral image
-    struct weighted_avg *wa;                    // weighted average of every pixel
-    ptrdiff_t wa_linesize;                      // linesize for wa in struct size unit
-    float *weight_lut;                          // lookup table mapping (scaled) patch differences to their associated weights
-    uint32_t max_meaningful_diff;               // maximum difference considered (if the patch difference is too high we ignore the pixel)
-    NLMeansDSPContext dsp;
-} NLMeansContext;
+#undef DEPTH
+#define DEPTH 16
+#include "nlmeans_template.c"
 
 #define OFFSET(x) offsetof(NLMeansContext, x)
 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
@@ -87,6 +84,13 @@  static int query_formats(AVFilterContext *ctx)
         AV_PIX_FMT_YUVJ422P, AV_PIX_FMT_YUVJ420P,
         AV_PIX_FMT_YUVJ411P,
         AV_PIX_FMT_GRAY8, AV_PIX_FMT_GBRP,
+        AV_PIX_FMT_YUV420P9, AV_PIX_FMT_YUV422P9, AV_PIX_FMT_YUV444P9,
+        AV_PIX_FMT_YUV420P10, AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV444P10,
+        AV_PIX_FMT_YUV440P10,
+        AV_PIX_FMT_YUV444P12, AV_PIX_FMT_YUV422P12, AV_PIX_FMT_YUV420P12,
+        AV_PIX_FMT_YUV440P12,
+        AV_PIX_FMT_YUV444P14, AV_PIX_FMT_YUV422P14, AV_PIX_FMT_YUV420P14,
+        AV_PIX_FMT_YUV420P16, AV_PIX_FMT_YUV422P16, AV_PIX_FMT_YUV444P16,
         AV_PIX_FMT_NONE
     };
 
@@ -96,380 +100,6 @@  static int query_formats(AVFilterContext *ctx)
     return ff_set_common_formats(ctx, fmts_list);
 }
 
-/**
- * Compute squared difference of the safe area (the zone where s1 and s2
- * overlap). It is likely the largest integral zone, so it is interesting to do
- * as little checks as possible; contrary to the unsafe version of this
- * function, we do not need any clipping here.
- *
- * The line above dst and the column to its left are always readable.
- */
-static void compute_safe_ssd_integral_image_c(uint32_t *dst, ptrdiff_t dst_linesize_32,
-                                              const uint8_t *s1, ptrdiff_t linesize1,
-                                              const uint8_t *s2, ptrdiff_t linesize2,
-                                              int w, int h)
-{
-    int x, y;
-    const uint32_t *dst_top = dst - dst_linesize_32;
-
-    /* SIMD-friendly assumptions allowed here */
-    av_assert2(!(w & 0xf) && w >= 16 && h >= 1);
-
-    for (y = 0; y < h; y++) {
-        for (x = 0; x < w; x += 4) {
-            const int d0 = s1[x    ] - s2[x    ];
-            const int d1 = s1[x + 1] - s2[x + 1];
-            const int d2 = s1[x + 2] - s2[x + 2];
-            const int d3 = s1[x + 3] - s2[x + 3];
-
-            dst[x    ] = dst_top[x    ] - dst_top[x - 1] + d0*d0;
-            dst[x + 1] = dst_top[x + 1] - dst_top[x    ] + d1*d1;
-            dst[x + 2] = dst_top[x + 2] - dst_top[x + 1] + d2*d2;
-            dst[x + 3] = dst_top[x + 3] - dst_top[x + 2] + d3*d3;
-
-            dst[x    ] += dst[x - 1];
-            dst[x + 1] += dst[x    ];
-            dst[x + 2] += dst[x + 1];
-            dst[x + 3] += dst[x + 2];
-        }
-        s1  += linesize1;
-        s2  += linesize2;
-        dst += dst_linesize_32;
-        dst_top += dst_linesize_32;
-    }
-}
-
-/**
- * Compute squared difference of an unsafe area (the zone nor s1 nor s2 could
- * be readable).
- *
- * On the other hand, the line above dst and the column to its left are always
- * readable.
- *
- * There is little point in having this function SIMDified as it is likely too
- * complex and only handle small portions of the image.
- *
- * @param dst               integral image
- * @param dst_linesize_32   integral image linesize (in 32-bit integers unit)
- * @param startx            integral starting x position
- * @param starty            integral starting y position
- * @param src               source plane buffer
- * @param linesize          source plane linesize
- * @param offx              source offsetting in x
- * @param offy              source offsetting in y
- * @paran r                 absolute maximum source offsetting
- * @param sw                source width
- * @param sh                source height
- * @param w                 width to compute
- * @param h                 height to compute
- */
-static inline void compute_unsafe_ssd_integral_image(uint32_t *dst, ptrdiff_t dst_linesize_32,
-                                                     int startx, int starty,
-                                                     const uint8_t *src, ptrdiff_t linesize,
-                                                     int offx, int offy, int r, int sw, int sh,
-                                                     int w, int h)
-{
-    int x, y;
-
-    for (y = starty; y < starty + h; y++) {
-        uint32_t acc = dst[y*dst_linesize_32 + startx - 1] - dst[(y-1)*dst_linesize_32 + startx - 1];
-        const int s1y = av_clip(y -  r,         0, sh - 1);
-        const int s2y = av_clip(y - (r + offy), 0, sh - 1);
-
-        for (x = startx; x < startx + w; x++) {
-            const int s1x = av_clip(x -  r,         0, sw - 1);
-            const int s2x = av_clip(x - (r + offx), 0, sw - 1);
-            const uint8_t v1 = src[s1y*linesize + s1x];
-            const uint8_t v2 = src[s2y*linesize + s2x];
-            const int d = v1 - v2;
-            acc += d * d;
-            dst[y*dst_linesize_32 + x] = dst[(y-1)*dst_linesize_32 + x] + acc;
-        }
-    }
-}
-
-/*
- * Compute the sum of squared difference integral image
- * http://www.ipol.im/pub/art/2014/57/
- * Integral Images for Block Matching - Gabriele Facciolo, Nicolas Limare, Enric Meinhardt-Llopis
- *
- * @param ii                integral image of dimension (w+e*2) x (h+e*2) with
- *                          an additional zeroed top line and column already
- *                          "applied" to the pointer value
- * @param ii_linesize_32    integral image linesize (in 32-bit integers unit)
- * @param src               source plane buffer
- * @param linesize          source plane linesize
- * @param offx              x-offsetting ranging in [-e;e]
- * @param offy              y-offsetting ranging in [-e;e]
- * @param w                 source width
- * @param h                 source height
- * @param e                 research padding edge
- */
-static void compute_ssd_integral_image(const NLMeansDSPContext *dsp,
-                                       uint32_t *ii, ptrdiff_t ii_linesize_32,
-                                       const uint8_t *src, ptrdiff_t linesize, int offx, int offy,
-                                       int e, int w, int h)
-{
-    // ii has a surrounding padding of thickness "e"
-    const int ii_w = w + e*2;
-    const int ii_h = h + e*2;
-
-    // we center the first source
-    const int s1x = e;
-    const int s1y = e;
-
-    // 2nd source is the frame with offsetting
-    const int s2x = e + offx;
-    const int s2y = e + offy;
-
-    // get the dimension of the overlapping rectangle where it is always safe
-    // to compare the 2 sources pixels
-    const int startx_safe = FFMAX(s1x, s2x);
-    const int starty_safe = FFMAX(s1y, s2y);
-    const int u_endx_safe = FFMIN(s1x + w, s2x + w); // unaligned
-    const int endy_safe   = FFMIN(s1y + h, s2y + h);
-
-    // deduce the safe area width and height
-    const int safe_pw = (u_endx_safe - startx_safe) & ~0xf;
-    const int safe_ph = endy_safe - starty_safe;
-
-    // adjusted end x position of the safe area after width of the safe area gets aligned
-    const int endx_safe = startx_safe + safe_pw;
-
-    // top part where only one of s1 and s2 is still readable, or none at all
-    compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
-                                      0, 0,
-                                      src, linesize,
-                                      offx, offy, e, w, h,
-                                      ii_w, starty_safe);
-
-    // fill the left column integral required to compute the central
-    // overlapping one
-    compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
-                                      0, starty_safe,
-                                      src, linesize,
-                                      offx, offy, e, w, h,
-                                      startx_safe, safe_ph);
-
-    // main and safe part of the integral
-    av_assert1(startx_safe - s1x >= 0); av_assert1(startx_safe - s1x < w);
-    av_assert1(starty_safe - s1y >= 0); av_assert1(starty_safe - s1y < h);
-    av_assert1(startx_safe - s2x >= 0); av_assert1(startx_safe - s2x < w);
-    av_assert1(starty_safe - s2y >= 0); av_assert1(starty_safe - s2y < h);
-    if (safe_pw && safe_ph)
-        dsp->compute_safe_ssd_integral_image(ii + starty_safe*ii_linesize_32 + startx_safe, ii_linesize_32,
-                                             src + (starty_safe - s1y) * linesize + (startx_safe - s1x), linesize,
-                                             src + (starty_safe - s2y) * linesize + (startx_safe - s2x), linesize,
-                                             safe_pw, safe_ph);
-
-    // right part of the integral
-    compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
-                                      endx_safe, starty_safe,
-                                      src, linesize,
-                                      offx, offy, e, w, h,
-                                      ii_w - endx_safe, safe_ph);
-
-    // bottom part where only one of s1 and s2 is still readable, or none at all
-    compute_unsafe_ssd_integral_image(ii, ii_linesize_32,
-                                      0, endy_safe,
-                                      src, linesize,
-                                      offx, offy, e, w, h,
-                                      ii_w, ii_h - endy_safe);
-}
-
-static int config_input(AVFilterLink *inlink)
-{
-    AVFilterContext *ctx = inlink->dst;
-    NLMeansContext *s = ctx->priv;
-    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
-    const int e = FFMAX(s->research_hsize, s->research_hsize_uv)
-                + FFMAX(s->patch_hsize,    s->patch_hsize_uv);
-
-    s->chroma_w = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
-    s->chroma_h = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
-    s->nb_planes = av_pix_fmt_count_planes(inlink->format);
-
-    /* Allocate the integral image with extra edges of thickness "e"
-     *
-     *   +_+-------------------------------+
-     *   |0|0000000000000000000000000000000|
-     *   +-x-------------------------------+
-     *   |0|\    ^                         |
-     *   |0| ii  | e                       |
-     *   |0|     v                         |
-     *   |0|   +-----------------------+   |
-     *   |0|   |                       |   |
-     *   |0|<->|                       |   |
-     *   |0| e |                       |   |
-     *   |0|   |                       |   |
-     *   |0|   +-----------------------+   |
-     *   |0|                               |
-     *   |0|                               |
-     *   |0|                               |
-     *   +-+-------------------------------+
-     */
-    s->ii_w = inlink->w + e*2;
-    s->ii_h = inlink->h + e*2;
-
-    // align to 4 the linesize, "+1" is for the space of the left 0-column
-    s->ii_lz_32 = FFALIGN(s->ii_w + 1, 4);
-
-    // "+1" is for the space of the top 0-line
-    s->ii_orig = av_mallocz_array(s->ii_h + 1, s->ii_lz_32 * sizeof(*s->ii_orig));
-    if (!s->ii_orig)
-        return AVERROR(ENOMEM);
-
-    // skip top 0-line and left 0-column
-    s->ii = s->ii_orig + s->ii_lz_32 + 1;
-
-    // allocate weighted average for every pixel
-    s->wa_linesize = inlink->w;
-    s->wa = av_malloc_array(s->wa_linesize, inlink->h * sizeof(*s->wa));
-    if (!s->wa)
-        return AVERROR(ENOMEM);
-
-    return 0;
-}
-
-struct thread_data {
-    const uint8_t *src;
-    ptrdiff_t src_linesize;
-    int startx, starty;
-    int endx, endy;
-    const uint32_t *ii_start;
-    int p;
-};
-
-static int nlmeans_slice(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
-{
-    int x, y;
-    NLMeansContext *s = ctx->priv;
-    const struct thread_data *td = arg;
-    const ptrdiff_t src_linesize = td->src_linesize;
-    const int process_h = td->endy - td->starty;
-    const int slice_start = (process_h *  jobnr   ) / nb_jobs;
-    const int slice_end   = (process_h * (jobnr+1)) / nb_jobs;
-    const int starty = td->starty + slice_start;
-    const int endy   = td->starty + slice_end;
-    const int p = td->p;
-    const uint32_t *ii = td->ii_start + (starty - p - 1) * s->ii_lz_32 - p - 1;
-    const int dist_b = 2*p + 1;
-    const int dist_d = dist_b * s->ii_lz_32;
-    const int dist_e = dist_d + dist_b;
-
-    for (y = starty; y < endy; y++) {
-        const uint8_t *src = td->src + y*src_linesize;
-        struct weighted_avg *wa = s->wa + y*s->wa_linesize;
-        for (x = td->startx; x < td->endx; x++) {
-            /*
-             * M is a discrete map where every entry contains the sum of all the entries
-             * in the rectangle from the top-left origin of M to its coordinate. In the
-             * following schema, "i" contains the sum of the whole map:
-             *
-             * M = +----------+-----------------+----+
-             *     |          |                 |    |
-             *     |          |                 |    |
-             *     |         a|                b|   c|
-             *     +----------+-----------------+----+
-             *     |          |                 |    |
-             *     |          |                 |    |
-             *     |          |        X        |    |
-             *     |          |                 |    |
-             *     |         d|                e|   f|
-             *     +----------+-----------------+----+
-             *     |          |                 |    |
-             *     |         g|                h|   i|
-             *     +----------+-----------------+----+
-             *
-             * The sum of the X box can be calculated with:
-             *    X = e-d-b+a
-             *
-             * See https://en.wikipedia.org/wiki/Summed_area_table
-             *
-             * The compute*_ssd functions compute the integral image M where every entry
-             * contains the sum of the squared difference of every corresponding pixels of
-             * two input planes of the same size as M.
-             */
-            const uint32_t a = ii[x];
-            const uint32_t b = ii[x + dist_b];
-            const uint32_t d = ii[x + dist_d];
-            const uint32_t e = ii[x + dist_e];
-            const uint32_t patch_diff_sq = e - d - b + a;
-
-            if (patch_diff_sq < s->max_meaningful_diff) {
-                const float weight = s->weight_lut[patch_diff_sq]; // exp(-patch_diff_sq * s->pdiff_scale)
-                wa[x].total_weight += weight;
-                wa[x].sum += weight * src[x];
-            }
-        }
-        ii += s->ii_lz_32;
-    }
-    return 0;
-}
-
-static void weight_averages(uint8_t *dst, ptrdiff_t dst_linesize,
-                            const uint8_t *src, ptrdiff_t src_linesize,
-                            struct weighted_avg *wa, ptrdiff_t wa_linesize,
-                            int w, int h)
-{
-    int x, y;
-
-    for (y = 0; y < h; y++) {
-        for (x = 0; x < w; x++) {
-            // Also weight the centered pixel
-            wa[x].total_weight += 1.f;
-            wa[x].sum += 1.f * src[x];
-            dst[x] = av_clip_uint8(wa[x].sum / wa[x].total_weight + 0.5f);
-        }
-        dst += dst_linesize;
-        src += src_linesize;
-        wa += wa_linesize;
-    }
-}
-
-static int nlmeans_plane(AVFilterContext *ctx, int w, int h, int p, int r,
-                         uint8_t *dst, ptrdiff_t dst_linesize,
-                         const uint8_t *src, ptrdiff_t src_linesize)
-{
-    int offx, offy;
-    NLMeansContext *s = ctx->priv;
-    /* patches center points cover the whole research window so the patches
-     * themselves overflow the research window */
-    const int e = r + p;
-    /* focus an integral pointer on the centered image (s1) */
-    const uint32_t *centered_ii = s->ii + e*s->ii_lz_32 + e;
-
-    memset(s->wa, 0, s->wa_linesize * h * sizeof(*s->wa));
-
-    for (offy = -r; offy <= r; offy++) {
-        for (offx = -r; offx <= r; offx++) {
-            if (offx || offy) {
-                struct thread_data td = {
-                    .src          = src + offy*src_linesize + offx,
-                    .src_linesize = src_linesize,
-                    .startx       = FFMAX(0, -offx),
-                    .starty       = FFMAX(0, -offy),
-                    .endx         = FFMIN(w, w - offx),
-                    .endy         = FFMIN(h, h - offy),
-                    .ii_start     = centered_ii + offy*s->ii_lz_32 + offx,
-                    .p            = p,
-                };
-
-                compute_ssd_integral_image(&s->dsp, s->ii, s->ii_lz_32,
-                                           src, src_linesize,
-                                           offx, offy, e, w, h);
-                ctx->internal->execute(ctx, nlmeans_slice, &td, NULL,
-                                       FFMIN(td.endy - td.starty, ff_filter_get_nb_threads(ctx)));
-            }
-        }
-    }
-
-    weight_averages(dst, dst_linesize, src, src_linesize,
-                    s->wa, s->wa_linesize, w, h);
-
-    return 0;
-}
-
 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
 {
     int i;
@@ -489,15 +119,30 @@  static int filter_frame(AVFilterLink *inlink, AVFrame *in)
         const int h = i ? s->chroma_h          : inlink->h;
         const int p = i ? s->patch_hsize_uv    : s->patch_hsize;
         const int r = i ? s->research_hsize_uv : s->research_hsize;
-        nlmeans_plane(ctx, w, h, p, r,
-                      out->data[i], out->linesize[i],
-                      in->data[i],  in->linesize[i]);
+        s->nlmeans_plane(ctx, w, h, p, r,
+                         out->data[i], out->linesize[i],
+                         in->data[i],  in->linesize[i]);
     }
 
     av_frame_free(&in);
     return ff_filter_frame(outlink, out);
 }
 
+void ff_nlmeans_init(NLMeansDSPContext *dsp, int depth)
+{
+    switch (depth) {
+    case  8: dsp->compute_safe_ssd_integral_image32 = compute_safe_ssd_integral_image_c8;  break;
+    case  9: dsp->compute_safe_ssd_integral_image64 = compute_safe_ssd_integral_image_c9;  break;
+    case 10: dsp->compute_safe_ssd_integral_image64 = compute_safe_ssd_integral_image_c10; break;
+    case 12: dsp->compute_safe_ssd_integral_image64 = compute_safe_ssd_integral_image_c12; break;
+    case 14: dsp->compute_safe_ssd_integral_image64 = compute_safe_ssd_integral_image_c14; break;
+    case 16: dsp->compute_safe_ssd_integral_image64 = compute_safe_ssd_integral_image_c16; break;
+    }
+
+    if (ARCH_AARCH64)
+        ff_nlmeans_init_aarch64(dsp, depth);
+}
+
 #define CHECK_ODD_FIELD(field, name) do {                       \
     if (!(s->field & 1)) {                                      \
         s->field |= 1;                                          \
@@ -506,26 +151,22 @@  static int filter_frame(AVFilterLink *inlink, AVFrame *in)
     }                                                           \
 } while (0)
 
-void ff_nlmeans_init(NLMeansDSPContext *dsp)
-{
-    dsp->compute_safe_ssd_integral_image = compute_safe_ssd_integral_image_c;
-
-    if (ARCH_AARCH64)
-        ff_nlmeans_init_aarch64(dsp);
-}
-
-static av_cold int init(AVFilterContext *ctx)
+static int config_input(AVFilterLink *inlink)
 {
-    int i;
+    AVFilterContext *ctx = inlink->dst;
     NLMeansContext *s = ctx->priv;
-    const double h = s->sigma * 10.;
+    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
+    double h;
+    int e;
 
+    s->depth = desc->comp[0].depth;
+    h = s->sigma * 10. * (1 << (s->depth - 8));
     s->pdiff_scale = 1. / (h * h);
-    s->max_meaningful_diff = log(255.) / s->pdiff_scale;
+    s->max_meaningful_diff = FFMIN(log(255.) / s->pdiff_scale, INT32_MAX / 8);
     s->weight_lut = av_calloc(s->max_meaningful_diff, sizeof(*s->weight_lut));
     if (!s->weight_lut)
         return AVERROR(ENOMEM);
-    for (i = 0; i < s->max_meaningful_diff; i++)
+    for (int i = 0; i < s->max_meaningful_diff; i++)
         s->weight_lut[i] = exp(-i * s->pdiff_scale);
 
     CHECK_ODD_FIELD(research_size,   "Luma research window");
@@ -542,11 +183,75 @@  static av_cold int init(AVFilterContext *ctx)
     s->patch_hsize       = s->patch_size       / 2;
     s->patch_hsize_uv    = s->patch_size_uv    / 2;
 
+    e = FFMAX(s->research_hsize, s->research_hsize_uv) +
+        FFMAX(s->patch_hsize,    s->patch_hsize_uv);
     av_log(ctx, AV_LOG_INFO, "Research window: %dx%d / %dx%d, patch size: %dx%d / %dx%d\n",
            s->research_size, s->research_size, s->research_size_uv, s->research_size_uv,
            s->patch_size,    s->patch_size,    s->patch_size_uv,    s->patch_size_uv);
 
-    ff_nlmeans_init(&s->dsp);
+    ff_nlmeans_init(&s->dsp, s->depth);
+
+    s->chroma_w = AV_CEIL_RSHIFT(inlink->w, desc->log2_chroma_w);
+    s->chroma_h = AV_CEIL_RSHIFT(inlink->h, desc->log2_chroma_h);
+    s->nb_planes = av_pix_fmt_count_planes(inlink->format);
+
+    /* Allocate the integral image with extra edges of thickness "e"
+     *
+     *   +_+-------------------------------+
+     *   |0|0000000000000000000000000000000|
+     *   +-x-------------------------------+
+     *   |0|\    ^                         |
+     *   |0| ii  | e                       |
+     *   |0|     v                         |
+     *   |0|   +-----------------------+   |
+     *   |0|   |                       |   |
+     *   |0|<->|                       |   |
+     *   |0| e |                       |   |
+     *   |0|   |                       |   |
+     *   |0|   +-----------------------+   |
+     *   |0|                               |
+     *   |0|                               |
+     *   |0|                               |
+     *   +-+-------------------------------+
+     */
+    s->ii_w = inlink->w + e*2;
+    s->ii_h = inlink->h + e*2;
+
+    // align to 4 the linesize, "+1" is for the space of the left 0-column
+    s->ii_lz_32 = FFALIGN(s->ii_w + 1, 4);
+
+    if (s->depth <= 8) {
+        // "+1" is for the space of the top 0-line
+        s->ii_orig32 = av_mallocz_array(s->ii_h + 1, s->ii_lz_32 * sizeof(*s->ii_orig32));
+        if (!s->ii_orig32)
+            return AVERROR(ENOMEM);
+
+        // skip top 0-line and left 0-column
+        s->ii32 = s->ii_orig32 + s->ii_lz_32 + 1;
+    } else {
+        // "+1" is for the space of the top 0-line
+        s->ii_orig64 = av_mallocz_array(s->ii_h + 1, s->ii_lz_32 * sizeof(*s->ii_orig64));
+        if (!s->ii_orig64)
+            return AVERROR(ENOMEM);
+
+        // skip top 0-line and left 0-column
+        s->ii64 = s->ii_orig64 + s->ii_lz_32 + 1;
+    }
+
+    // allocate weighted average for every pixel
+    s->wa_linesize = inlink->w;
+    s->wa = av_malloc_array(s->wa_linesize, inlink->h * sizeof(*s->wa));
+    if (!s->wa)
+        return AVERROR(ENOMEM);
+
+    switch (s->depth) {
+    case  8: s->nlmeans_plane = nlmeans_plane8;  break;
+    case  9: s->nlmeans_plane = nlmeans_plane9;  break;
+    case 10: s->nlmeans_plane = nlmeans_plane10; break;
+    case 12: s->nlmeans_plane = nlmeans_plane12; break;
+    case 14: s->nlmeans_plane = nlmeans_plane14; break;
+    case 16: s->nlmeans_plane = nlmeans_plane16; break;
+    }
 
     return 0;
 }
@@ -555,7 +260,8 @@  static av_cold void uninit(AVFilterContext *ctx)
 {
     NLMeansContext *s = ctx->priv;
     av_freep(&s->weight_lut);
-    av_freep(&s->ii_orig);
+    av_freep(&s->ii_orig32);
+    av_freep(&s->ii_orig64);
     av_freep(&s->wa);
 }
 
@@ -581,7 +287,6 @@  AVFilter ff_vf_nlmeans = {
     .name          = "nlmeans",
     .description   = NULL_IF_CONFIG_SMALL("Non-local means denoiser."),
     .priv_size     = sizeof(NLMeansContext),
-    .init          = init,
     .uninit        = uninit,
     .query_formats = query_formats,
     .inputs        = nlmeans_inputs,
diff --git a/libavfilter/vf_nlmeans.h b/libavfilter/vf_nlmeans.h
index 0a9aab2928..814b95e10f 100644
--- a/libavfilter/vf_nlmeans.h
+++ b/libavfilter/vf_nlmeans.h
@@ -23,13 +23,60 @@ 
 #include <stdint.h>
 
 typedef struct NLMeansDSPContext {
-    void (*compute_safe_ssd_integral_image)(uint32_t *dst, ptrdiff_t dst_linesize_32,
-                                            const uint8_t *s1, ptrdiff_t linesize1,
-                                            const uint8_t *s2, ptrdiff_t linesize2,
-                                            int w, int h);
+    void (*compute_safe_ssd_integral_image32)(uint32_t *dst, ptrdiff_t dst_linesize_32,
+                                              const uint8_t *s1, ptrdiff_t linesize1,
+                                              const uint8_t *s2, ptrdiff_t linesize2,
+                                              int w, int h);
+    void (*compute_safe_ssd_integral_image64)(uint64_t *dst, ptrdiff_t dst_linesize_32,
+                                              const uint8_t *s1, ptrdiff_t linesize1,
+                                              const uint8_t *s2, ptrdiff_t linesize2,
+                                              int w, int h);
 } NLMeansDSPContext;
 
-void ff_nlmeans_init(NLMeansDSPContext *dsp);
-void ff_nlmeans_init_aarch64(NLMeansDSPContext *dsp);
+struct weighted_avg {
+    float total_weight;
+    float sum;
+};
+
+struct thread_data {
+    const uint8_t *src;
+    ptrdiff_t src_linesize;
+    int startx, starty;
+    int endx, endy;
+    const uint32_t *ii_start32;
+    const uint64_t *ii_start64;
+    int p;
+};
+
+typedef struct NLMeansContext {
+    const AVClass *class;
+    int nb_planes;
+    int depth;
+    int chroma_w, chroma_h;
+    double pdiff_scale;                         // invert of the filtering parameter (sigma*10) squared
+    double sigma;                               // denoising strength
+    int patch_size,    patch_hsize;             // patch size and half size
+    int patch_size_uv, patch_hsize_uv;          // patch size and half size for chroma planes
+    int research_size,    research_hsize;       // research size and half size
+    int research_size_uv, research_hsize_uv;    // research size and half size for chroma planes
+    uint32_t *ii_orig32;                        // integral image 32-bit
+    uint32_t *ii32;                             // integral image 32-bit starting after the 0-line and 0-column
+    uint64_t *ii_orig64;                        // integral image 64-bit
+    uint64_t *ii64;                             // integral image 64-bit starting after the 0-line and 0-column
+    int ii_w, ii_h;                             // width and height of the integral image
+    ptrdiff_t ii_lz_32;                         // linesize in 32-bit units of the integral image
+    struct weighted_avg *wa;                    // weighted average of every pixel
+    ptrdiff_t wa_linesize;                      // linesize for wa in struct size unit
+    float *weight_lut;                          // lookup table mapping (scaled) patch differences to their associated weights
+    uint32_t max_meaningful_diff;               // maximum difference considered (if the patch difference is too high we ignore the pixel)
+    NLMeansDSPContext dsp;
+
+    int (*nlmeans_plane)(AVFilterContext *ctx, int w, int h, int p, int r,
+                         uint8_t *dst, ptrdiff_t dst_linesize,
+                         const uint8_t *src, ptrdiff_t src_linesize);
+} NLMeansContext;
+
+void ff_nlmeans_init(NLMeansDSPContext *dsp, int depth);
+void ff_nlmeans_init_aarch64(NLMeansDSPContext *dsp, int depth);
 
 #endif /* AVFILTER_NLMEANS_H */