avfilter: add LIBVMAF filter

This one changes the previous vmaf patch to libvmaf to keep it separate from the
native implementation of vmaf inside ffmpeg later.

Signed-off-by: Ashish Singh <ashk43712@gmail.com>
Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>

Changelog

@@ -27,6 +27,7 @@ version <next>:
 - additional frame format support for Interplay MVE movies
 - support for decoding through D3D11VA in ffmpeg
 - limiter video filter
+- libvmaf video filter
 
 version 3.3:
 - CrystalHD decoder moved to new decode API

configure

@@ -256,6 +256,7 @@ External library support:
   --enable-libtwolame      enable MP2 encoding via libtwolame [no]
   --enable-libv4l2         enable libv4l2/v4l-utils [no]
   --enable-libvidstab      enable video stabilization using vid.stab [no]
+  --enable-libvmaf         enable vmaf filter via libvmaf [no]
   --enable-libvo-amrwbenc  enable AMR-WB encoding via libvo-amrwbenc [no]
   --enable-libvorbis       enable Vorbis en/decoding via libvorbis,
                            native implementation exists [no]
@@ -1569,6 +1570,7 @@ EXTERNAL_LIBRARY_LIST="
     libtheora
     libtwolame
     libv4l2
+    libvmaf
     libvorbis
     libvpx
     libwavpack
@@ -3186,6 +3188,7 @@ uspp_filter_deps="gpl avcodec"
 vaguedenoiser_filter_deps="gpl"
 vidstabdetect_filter_deps="libvidstab"
 vidstabtransform_filter_deps="libvidstab"
+vmaf_filter_deps="libvmaf"
 zmq_filter_deps="libzmq"
 zoompan_filter_deps="swscale"
 zscale_filter_deps="libzimg"
@@ -5902,6 +5905,8 @@ enabled libtwolame        && require libtwolame twolame.h twolame_init -ltwolame
                                die "ERROR: libtwolame must be installed and version must be >= 0.3.10"; }
 enabled libv4l2           && require_pkg_config libv4l2 libv4l2.h v4l2_ioctl
 enabled libvidstab        && require_pkg_config "vidstab >= 0.98" vid.stab/libvidstab.h vsMotionDetectInit
+enabled libvmaf           && { check_lib libvmaf "libvmaf.h" "compute_vmaf" -lvmaf -lstdc++ -lpthread -lm ||
+                               die "ERROR: libvmaf must be installed"; }
 enabled libvo_amrwbenc    && require libvo_amrwbenc vo-amrwbenc/enc_if.h E_IF_init -lvo-amrwbenc
 enabled libvorbis         && require_pkg_config vorbis vorbis/codec.h vorbis_info_init &&
                              require_pkg_config vorbisenc vorbis/vorbisenc.h vorbis_encode_init

doc/filters.texi

@@ -9639,6 +9639,70 @@ The formula that generates the correction is:
 where @var{r_0} is halve of the image diagonal and @var{r_src} and @var{r_tgt} are the
 distances from the focal point in the source and target images, respectively.
 
+@section libvmaf
+
+Obtain the average VMAF (Video Multi-Method Assessment Fusion)
+score between two input videos.
+
+This filter takes two input videos.
+
+Both video inputs must have the same resolution and pixel format for
+this filter to work correctly. Also it assumes that both inputs
+have the same number of frames, which are compared one by one.
+
+The obtained average VMAF score is printed through the logging system.
+
+It requires Netflix's vmaf library (libvmaf) as a pre-requisite.
+After installing the library it can be enabled using:
+@code{./configure --enable-libvmaf}.
+If no model path is specified it uses the default model: @code{vmaf_v0.6.1.pkl}.
+
+On the below examples the input file @file{main.mpg} being processed is
+compared with the reference file @file{ref.mpg}.
+
+The filter has following options:
+
+@table @option
+@item model_path
+Set the model path which is to be used for SVM.
+Default value: @code{"vmaf_v0.6.1.pkl"}
+
+@item log_path
+Set the file path to be used to store logs.
+
+@item log_fmt
+Set the format of the log file (xml or json).
+
+@item enable_transform
+Enables transform for computing vmaf.
+
+@item phone_model
+Invokes the phone model which will generate VMAF scores higher than in the
+regular model, which is more suitable for laptop, TV, etc. viewing conditions.
+
+@item psnr
+Enables computing psnr along with vmaf.
+
+@item ssim
+Enables computing ssim along with vmaf.
+
+@item ms_ssim
+Enables computing ms_ssim along with vmaf.
+
+@item pool
+Set the pool method to be used for computing vmaf.
+@end table
+
+For example:
+@example
+ffmpeg -i main.mpg -i ref.mpg -lavfi libvmaf -f null -
+@end example
+
+Example with options:
+@example
+ffmpeg -i main.mpg -i ref.mpg -lavfi libvmaf="psnr=1:enable-transform=1" -f null -
+@end example
+
 @section limiter
 
 Limits the pixel components values to the specified range [min, max].

libavfilter/Makefile

@@ -216,6 +216,7 @@ OBJS-$(CONFIG_INTERLACE_FILTER)              += vf_interlace.o
 OBJS-$(CONFIG_INTERLEAVE_FILTER)             += f_interleave.o
 OBJS-$(CONFIG_KERNDEINT_FILTER)              += vf_kerndeint.o
 OBJS-$(CONFIG_LENSCORRECTION_FILTER)         += vf_lenscorrection.o
+OBJS-$(CONFIG_LIBVMAF_FILTER)                += vf_libvmaf.o dualinput.o framesync.o
 OBJS-$(CONFIG_LIMITER_FILTER)                += vf_limiter.o
 OBJS-$(CONFIG_LOOP_FILTER)                   += f_loop.o
 OBJS-$(CONFIG_LUMAKEY_FILTER)                += vf_lumakey.o

libavfilter/allfilters.c

@@ -228,6 +228,7 @@ static void register_all(void)
     REGISTER_FILTER(INTERLEAVE,     interleave,     vf);
     REGISTER_FILTER(KERNDEINT,      kerndeint,      vf);
     REGISTER_FILTER(LENSCORRECTION, lenscorrection, vf);
+    REGISTER_FILTER(LIBVMAF,        libvmaf,        vf);
     REGISTER_FILTER(LIMITER,        limiter,        vf);
     REGISTER_FILTER(LOOP,           loop,           vf);
     REGISTER_FILTER(LUMAKEY,        lumakey,        vf);

libavfilter/vf_libvmaf.c

@@ -0,0 +1,339 @@
+/*
+ * Copyright (c) 2017 Ronald S. Bultje <rsbultje@gmail.com>
+ * Copyright (c) 2017 Ashish Pratap Singh <ashk43712@gmail.com>
+ *
+ * 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
+ */
+
+/**
+ * @file
+ * Calculate the VMAF between two input videos.
+ */
+
+#include <inttypes.h>
+#include <pthread.h>
+#include <libvmaf.h>
+#include "libavutil/avstring.h"
+#include "libavutil/opt.h"
+#include "libavutil/pixdesc.h"
+#include "avfilter.h"
+#include "dualinput.h"
+#include "drawutils.h"
+#include "formats.h"
+#include "internal.h"
+#include "video.h"
+
+typedef struct LIBVMAFContext {
+    const AVClass *class;
+    FFDualInputContext dinput;
+    const AVPixFmtDescriptor *desc;
+    char *format;
+    int width;
+    int height;
+    double vmaf_score;
+    pthread_t vmaf_thread;
+    pthread_mutex_t lock;
+    pthread_cond_t cond;
+    int eof;
+    AVFrame *gmain;
+    AVFrame *gref;
+    int frame_set;
+    char *model_path;
+    char *log_path;
+    char *log_fmt;
+    int disable_clip;
+    int disable_avx;
+    int enable_transform;
+    int phone_model;
+    int psnr;
+    int ssim;
+    int ms_ssim;
+    char *pool;
+} LIBVMAFContext;
+
+#define OFFSET(x) offsetof(LIBVMAFContext, x)
+#define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
+
+static const AVOption libvmaf_options[] = {
+    {"model_path",  "Set the model to be used for computing vmaf.",                     OFFSET(model_path), AV_OPT_TYPE_STRING, {.str="/usr/local/share/model/vmaf_v0.6.1.pkl"}, 0, 1, FLAGS},
+    {"log_path",  "Set the file path to be used to store logs.",                        OFFSET(log_path), AV_OPT_TYPE_STRING, {.str=NULL}, 0, 1, FLAGS},
+    {"log_fmt",  "Set the format of the log (xml or json).",                            OFFSET(log_fmt), AV_OPT_TYPE_STRING, {.str=NULL}, 0, 1, FLAGS},
+    {"enable_transform",  "Enables transform for computing vmaf.",                      OFFSET(enable_transform), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, FLAGS},
+    {"phone_model",  "Invokes the phone model that will generate higher VMAF scores.",  OFFSET(phone_model), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, FLAGS},
+    {"psnr",  "Enables computing psnr along with vmaf.",                                OFFSET(psnr), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, FLAGS},
+    {"ssim",  "Enables computing ssim along with vmaf.",                                OFFSET(ssim), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, FLAGS},
+    {"ms_ssim",  "Enables computing ms-ssim along with vmaf.",                          OFFSET(ms_ssim), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, FLAGS},
+    {"pool",  "Set the pool method to be used for computing vmaf.",                     OFFSET(pool), AV_OPT_TYPE_STRING, {.str=NULL}, 0, 1, FLAGS},
+    { NULL }
+};
+
+AVFILTER_DEFINE_CLASS(libvmaf);
+
+#define read_frame_fn(type, bits)                                               \
+    static int read_frame_##bits##bit(float *ref_data, float *main_data,            \
+                                      float *temp_data, int stride,             \
+                                      double *score, void *ctx)                 \
+{                                                                               \
+    LIBVMAFContext *s = (LIBVMAFContext *) ctx;                                       \
+    int ret;                                                                    \
+    \
+    pthread_mutex_lock(&s->lock);                                               \
+    \
+    while (!s->frame_set && !s->eof) {                                          \
+        pthread_cond_wait(&s->cond, &s->lock);                                  \
+    }                                                                           \
+    \
+    if (s->frame_set) {                                                         \
+        int ref_stride = s->gref->linesize[0];                                  \
+        int main_stride = s->gmain->linesize[0];                                \
+        \
+        const type *ref_ptr = (const type *) s->gref->data[0];                  \
+        const type *main_ptr = (const type *) s->gmain->data[0];                \
+        \
+        float *ptr = ref_data;                                                  \
+        \
+        int h = s->height;                                                      \
+        int w = s->width;                                                       \
+        \
+        int i,j;                                                                \
+        \
+        for (i = 0; i < h; i++) {                                               \
+            for ( j = 0; j < w; j++) {                                          \
+                ptr[j] = (float)ref_ptr[j];                                     \
+            }                                                                   \
+            ref_ptr += ref_stride / sizeof(*ref_ptr);                           \
+            ptr += stride / sizeof(*ptr);                                       \
+        }                                                                       \
+        \
+        ptr = main_data;                                                        \
+        \
+        for (i = 0; i < h; i++) {                                               \
+            for (j = 0; j < w; j++) {                                           \
+                ptr[j] = (float)main_ptr[j];                                    \
+            }                                                                   \
+            main_ptr += main_stride / sizeof(*main_ptr);                        \
+            ptr += stride / sizeof(*ptr);                                       \
+        }                                                                       \
+    }                                                                           \
+    \
+    ret = !s->frame_set;                                                        \
+    \
+    s->frame_set = 0;                                                           \
+    \
+    pthread_cond_signal(&s->cond);                                              \
+    pthread_mutex_unlock(&s->lock);                                             \
+    \
+    if (ret) {                                                                  \
+        return 2;                                                               \
+    }                                                                           \
+    \
+    return 0;                                                                   \
+}
+
+read_frame_fn(uint8_t, 8);
+read_frame_fn(uint16_t, 10);
+
+static void compute_vmaf_score(LIBVMAFContext *s)
+{
+    int (*read_frame)(float *ref_data, float *main_data, float *temp_data,
+                      int stride, double *score, void *ctx);
+
+    if (s->desc->comp[0].depth <= 8) {
+        read_frame = read_frame_8bit;
+    } else {
+        read_frame = read_frame_10bit;
+    }
+
+    s->vmaf_score = compute_vmaf(s->format, s->width, s->height, read_frame, s,
+                                 s->model_path, s->log_path, s->log_fmt, 0, 0,
+                                 s->enable_transform, s->phone_model, s->psnr,
+                                 s->ssim, s->ms_ssim, s->pool);
+}
+
+static void *call_vmaf(void *ctx)
+{
+    LIBVMAFContext *s = (LIBVMAFContext *) ctx;
+    compute_vmaf_score(s);
+    av_log(ctx, AV_LOG_INFO, "VMAF score: %f\n",s->vmaf_score);
+    pthread_exit(NULL);
+}
+
+static AVFrame *do_vmaf(AVFilterContext *ctx, AVFrame *main, const AVFrame *ref)
+{
+    LIBVMAFContext *s = ctx->priv;
+
+    pthread_mutex_lock(&s->lock);
+
+    while (s->frame_set != 0) {
+        pthread_cond_wait(&s->cond, &s->lock);
+    }
+
+    av_frame_ref(s->gref, ref);
+    av_frame_ref(s->gmain, main);
+
+    s->frame_set = 1;
+
+    pthread_cond_signal(&s->cond);
+    pthread_mutex_unlock(&s->lock);
+
+    return main;
+}
+
+static av_cold int init(AVFilterContext *ctx)
+{
+    LIBVMAFContext *s = ctx->priv;
+
+    s->gref = av_frame_alloc();
+    s->gmain = av_frame_alloc();
+
+    pthread_mutex_init(&s->lock, NULL);
+    pthread_cond_init (&s->cond, NULL);
+
+    s->dinput.process = do_vmaf;
+    return 0;
+}
+
+static int query_formats(AVFilterContext *ctx)
+{
+    static const enum AVPixelFormat pix_fmts[] = {
+        AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV420P,
+        AV_PIX_FMT_YUV444P10LE, AV_PIX_FMT_YUV422P10LE, AV_PIX_FMT_YUV420P10LE,
+        AV_PIX_FMT_NONE
+    };
+
+    AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
+    if (!fmts_list)
+        return AVERROR(ENOMEM);
+    return ff_set_common_formats(ctx, fmts_list);
+}
+
+
+static int config_input_ref(AVFilterLink *inlink)
+{
+    AVFilterContext *ctx  = inlink->dst;
+    LIBVMAFContext *s = ctx->priv;
+    int th;
+
+    if (ctx->inputs[0]->w != ctx->inputs[1]->w ||
+        ctx->inputs[0]->h != ctx->inputs[1]->h) {
+        av_log(ctx, AV_LOG_ERROR, "Width and height of input videos must be same.\n");
+        return AVERROR(EINVAL);
+    }
+    if (ctx->inputs[0]->format != ctx->inputs[1]->format) {
+        av_log(ctx, AV_LOG_ERROR, "Inputs must be of same pixel format.\n");
+        return AVERROR(EINVAL);
+    }
+
+    s->desc = av_pix_fmt_desc_get(inlink->format);
+    s->width = ctx->inputs[0]->w;
+    s->height = ctx->inputs[0]->h;
+
+    th = pthread_create(&s->vmaf_thread, NULL, call_vmaf, (void *) s);
+    if (th) {
+        av_log(ctx, AV_LOG_ERROR, "Thread creation failed.\n");
+        return AVERROR(EINVAL);
+    }
+
+    return 0;
+}
+
+
+static int config_output(AVFilterLink *outlink)
+{
+    AVFilterContext *ctx = outlink->src;
+    LIBVMAFContext *s = ctx->priv;
+    AVFilterLink *mainlink = ctx->inputs[0];
+    int ret;
+
+    outlink->w = mainlink->w;
+    outlink->h = mainlink->h;
+    outlink->time_base = mainlink->time_base;
+    outlink->sample_aspect_ratio = mainlink->sample_aspect_ratio;
+    outlink->frame_rate = mainlink->frame_rate;
+    if ((ret = ff_dualinput_init(ctx, &s->dinput)) < 0)
+        return ret;
+
+    return 0;
+}
+
+static int filter_frame(AVFilterLink *inlink, AVFrame *inpicref)
+{
+    LIBVMAFContext *s = inlink->dst->priv;
+    return ff_dualinput_filter_frame(&s->dinput, inlink, inpicref);
+}
+
+static int request_frame(AVFilterLink *outlink)
+{
+    LIBVMAFContext *s = outlink->src->priv;
+    return ff_dualinput_request_frame(&s->dinput, outlink);
+}
+
+static av_cold void uninit(AVFilterContext *ctx)
+{
+    LIBVMAFContext *s = ctx->priv;
+
+    ff_dualinput_uninit(&s->dinput);
+
+    pthread_mutex_lock(&s->lock);
+    s->eof = 1;
+    pthread_cond_signal(&s->cond);
+    pthread_mutex_unlock(&s->lock);
+
+    pthread_join(s->vmaf_thread, NULL);
+
+    av_frame_free(&s->gref);
+    av_frame_free(&s->gmain);
+
+    pthread_mutex_destroy(&s->lock);
+    pthread_cond_destroy(&s->cond);
+}
+
+static const AVFilterPad libvmaf_inputs[] = {
+    {
+        .name         = "main",
+        .type         = AVMEDIA_TYPE_VIDEO,
+        .filter_frame = filter_frame,
+    },{
+        .name         = "reference",
+        .type         = AVMEDIA_TYPE_VIDEO,
+        .filter_frame = filter_frame,
+        .config_props = config_input_ref,
+    },
+    { NULL }
+};
+
+static const AVFilterPad libvmaf_outputs[] = {
+    {
+        .name          = "default",
+        .type          = AVMEDIA_TYPE_VIDEO,
+        .config_props  = config_output,
+        .request_frame = request_frame,
+    },
+    { NULL }
+};
+
+AVFilter ff_vf_libvmaf = {
+    .name          = "libvmaf",
+    .description   = NULL_IF_CONFIG_SMALL("Calculate the VMAF between two video streams."),
+    .init          = init,
+    .uninit        = uninit,
+    .query_formats = query_formats,
+    .priv_size     = sizeof(LIBVMAFContext),
+    .priv_class    = &libvmaf_class,
+    .inputs        = libvmaf_inputs,
+    .outputs       = libvmaf_outputs,
+};