multimedia/ffmpeg: switch VMAF to v2 API (built-in models)

2 files changed, 809 insertions, 13 deletions

diff --git a/multimedia/ffmpeg/Makefile b/multimedia/ffmpeg/Makefile
index ee7b7ce9179d..ba4de1fc5fcd 100644
--- a/multimedia/ffmpeg/Makefile
+++ b/multimedia/ffmpeg/Makefile
@@ -2,7 +2,7 @@
 
 PORTNAME=	ffmpeg
 PORTVERSION=	4.4.1
-PORTREVISION=	1
+PORTREVISION=	2
 PORTEPOCH=	1
 CATEGORIES=	multimedia audio net
 MASTER_SITES=	https://ffmpeg.org/releases/
diff --git a/multimedia/ffmpeg/files/patch-vmaf b/multimedia/ffmpeg/files/patch-vmaf
index b7b205fd9eb3..b503561803c1 100644
--- a/multimedia/ffmpeg/files/patch-vmaf
+++ b/multimedia/ffmpeg/files/patch-vmaf
@@ -1,32 +1,828 @@
-https://github.com/Netflix/vmaf/commit/122089fa3d23
+https://lists.ffmpeg.org/pipermail/ffmpeg-devel/2021-June/281878.html
 
+--- configure.orig	2021-10-24 20:47:11 UTC
++++ configure
+@@ -3663,7 +3663,7 @@ uspp_filter_deps="gpl avcodec"
+ vaguedenoiser_filter_deps="gpl"
+ vidstabdetect_filter_deps="libvidstab"
+ vidstabtransform_filter_deps="libvidstab"
+-libvmaf_filter_deps="libvmaf pthreads"
++libvmaf_filter_deps="libvmaf"
+ zmq_filter_deps="libzmq"
+ zoompan_filter_deps="swscale"
+ zscale_filter_deps="libzimg const_nan"
+@@ -6441,7 +6441,7 @@ enabled libtwolame        && require libtwolame twolam
+ enabled libuavs3d         && require_pkg_config libuavs3d "uavs3d >= 1.1.41" uavs3d.h uavs3d_decode
+ enabled libv4l2           && require_pkg_config libv4l2 libv4l2 libv4l2.h v4l2_ioctl
+ enabled libvidstab        && require_pkg_config libvidstab "vidstab >= 0.98" vid.stab/libvidstab.h vsMotionDetectInit
+-enabled libvmaf           && require_pkg_config libvmaf "libvmaf >= 1.5.2" libvmaf.h compute_vmaf
++enabled libvmaf           && require_pkg_config libvmaf "libvmaf >= 2.0.0" libvmaf.h vmaf_init
+ enabled libvo_amrwbenc    && require libvo_amrwbenc vo-amrwbenc/enc_if.h E_IF_init -lvo-amrwbenc
+ enabled libvorbis         && require_pkg_config libvorbis vorbis vorbis/codec.h vorbis_info_init &&
+                              require_pkg_config libvorbisenc vorbisenc vorbis/vorbisenc.h vorbis_encode_init
 --- doc/filters.texi.orig	2021-10-24 20:47:07 UTC
 +++ doc/filters.texi
-@@ -13875,14 +13875,14 @@ The obtained VMAF score is printed through the logging
+@@ -13867,66 +13867,37 @@ ffmpeg -i input.mov -vf lensfun=make=Canon:model="Cano
+ 
+ @section libvmaf
+ 
+-Obtain the VMAF (Video Multi-Method Assessment Fusion)
+-score between two input videos.
++Calulate the VMAF (Video Multi-Method Assessment Fusion) score for a
++reference/distorted pair of input videos.
+ 
+ The obtained 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}.
-+If no model path is specified it uses the default model: @code{vmaf_v0.6.1.json}.
  
  The filter has following options:
  
  @table @option
- @item model_path
- Set the model path which is to be used for SVM.
+-@item model_path
+-Set the model path which is to be used for SVM.
 -Default value: @code{"/usr/local/share/model/vmaf_v0.6.1.pkl"}
-+Default value: @code{"/usr/local/share/model/vmaf_v0.6.1.json"}
++@item model
++A `|` delimited list of vmaf models. Each model can be configured with a number of parameters.
++Default value: @code{"version=vmaf_v0.6.1"}
  
++@item feature
++A `|` delimited list of features. Each feature can be configured with a number of parameters.
++
  @item log_path
- Set the file path to be used to store logs.
+-Set the file path to be used to store logs.
++Set the file path to be used to store log files.
+ 
+ @item log_fmt
+-Set the format of the log file (csv, json or xml).
++Set the format of the log file (xml, json, csv, or sub).
+ 
+-@item enable_transform
+-This option can enable/disable the @code{score_transform} applied to the final predicted VMAF score,
+-if you have specified score_transform option in the input parameter file passed to @code{run_vmaf_training.py}
+-Default value: @code{false}
+-
+-@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.
+-Default value: @code{false}
+-
+-@item psnr
+-Enables computing psnr along with vmaf.
+-Default value: @code{false}
+-
+-@item ssim
+-Enables computing ssim along with vmaf.
+-Default value: @code{false}
+-
+-@item ms_ssim
+-Enables computing ms_ssim along with vmaf.
+-Default value: @code{false}
+-
+-@item pool
+-Set the pool method to be used for computing vmaf.
+-Options are @code{min}, @code{harmonic_mean} or @code{mean} (default).
+-
+ @item n_threads
+-Set number of threads to be used when computing vmaf.
+-Default value: @code{0}, which makes use of all available logical processors.
++Set number of threads to be used when initializing libvmaf.
++Default value: @code{0}, no threads.
+ 
+ @item n_subsample
+-Set interval for frame subsampling used when computing vmaf.
+-Default value: @code{1}
+-
+-@item enable_conf_interval
+-Enables confidence interval.
+-Default value: @code{false}
++Set frame subsampling interval to be used.
+ @end table
+ 
+ This filter also supports the @ref{framesync} options.
+@@ -13934,23 +13905,31 @@ This filter also supports the @ref{framesync} options.
+ @subsection Examples
+ @itemize
+ @item
+-On the below examples the input file @file{main.mpg} being processed is
+-compared with the reference file @file{ref.mpg}.
++In the examples below, a distorted video @file{distorted.mpg} is
++compared with a reference file @file{reference.mpg}.
+ 
++@item
++Basic usage:
+ @example
+-ffmpeg -i main.mpg -i ref.mpg -lavfi libvmaf -f null -
++ffmpeg -i distorted.mpg -i reference.mpg -lavfi libvmaf=log_path=output.xml -f null -
+ @end example
+ 
+ @item
+-Example with options:
++Example with multiple models:
+ @example
+-ffmpeg -i main.mpg -i ref.mpg -lavfi libvmaf="psnr=1:log_fmt=json" -f null -
++ffmpeg -i distorted.mpg -i reference.mpg -lavfi libvmaf='model=version=vmaf_v0.6.1\\:name=vmaf|version=vmaf_v0.6.1neg\\:name=vmaf_neg:log_path=output.xml' -f null -
+ @end example
+ 
+ @item
++Example with multiple addtional features:
++@example
++ffmpeg -i distorted.mpg -i reference.mpg -lavfi libvmaf='feature=name=psnr|name=ciede:log_path=output.xml' -f null -
++@end example
++
++@item
+ Example with options and different containers:
+ @example
+-ffmpeg -i main.mpg -i ref.mkv -lavfi "[0:v]settb=AVTB,setpts=PTS-STARTPTS[main];[1:v]settb=AVTB,setpts=PTS-STARTPTS[ref];[main][ref]libvmaf=psnr=1:log_fmt=json" -f null -
++ffmpeg -i distorted.mpg -i reference.mkv -lavfi "[0:v]settb=AVTB,setpts=PTS-STARTPTS[main];[1:v]settb=AVTB,setpts=PTS-STARTPTS[ref];[main][ref]libvmaf=log_fmt=json:log_path=output.json" -f null -
+ @end example
+ @end itemize
+ 
 --- libavfilter/vf_libvmaf.c.orig	2021-10-24 20:47:07 UTC
 +++ libavfilter/vf_libvmaf.c
-@@ -72,7 +72,7 @@ typedef struct LIBVMAFContext {
+@@ -24,8 +24,8 @@
+  * Calculate the VMAF between two input videos.
+  */
+ 
+-#include <pthread.h>
+ #include <libvmaf.h>
++
+ #include "libavutil/avstring.h"
+ #include "libavutil/opt.h"
+ #include "libavutil/pixdesc.h"
+@@ -39,211 +39,372 @@
+ typedef struct LIBVMAFContext {
+     const AVClass *class;
+     FFFrameSync fs;
+-    const AVPixFmtDescriptor *desc;
+-    int width;
+-    int height;
+-    double vmaf_score;
+-    int vmaf_thread_created;
+-    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;
+     int n_threads;
+     int n_subsample;
+-    int enable_conf_interval;
+-    int error;
++    char *model_cfg;
++    char *feature_cfg;
++    VmafContext *vmaf;
++    VmafModel **model;
++    unsigned model_cnt;
++    unsigned frame_cnt;
++    unsigned bpc;
+ } 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},
-+    {"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.json"}, 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 (csv, json or xml).",                       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},
+-    {"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 (csv, json or xml).",                       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},
+-    {"n_threads", "Set number of threads to be used when computing vmaf.",              OFFSET(n_threads), AV_OPT_TYPE_INT, {.i64=0}, 0, UINT_MAX, FLAGS},
+-    {"n_subsample", "Set interval for frame subsampling used when computing vmaf.",     OFFSET(n_subsample), AV_OPT_TYPE_INT, {.i64=1}, 1, UINT_MAX, FLAGS},
+-    {"enable_conf_interval",  "Enables confidence interval.",                           OFFSET(enable_conf_interval), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, FLAGS},
++    {"model", "Set the model to be used for computing vmaf.",                             OFFSET(model_cfg), AV_OPT_TYPE_STRING, {.str="version=vmaf_v0.6.1"}, 0, 1, FLAGS},
++    {"feature", "Set the feature to be used for computing vmaf.",                         OFFSET(feature_cfg), AV_OPT_TYPE_STRING, {.str=NULL}, 0, 1, FLAGS},
++    {"log_path",  "Set the file path to be used to write log.",                           OFFSET(log_path), AV_OPT_TYPE_STRING, {.str=NULL}, 0, 1, FLAGS},
++    {"log_fmt",  "Set the format of the log (csv, json, xml, or sub).",                   OFFSET(log_fmt), AV_OPT_TYPE_STRING, {.str="xml"}, 0, 1, FLAGS},
++    {"n_threads", "Set number of threads to be used when computing vmaf.",                OFFSET(n_threads), AV_OPT_TYPE_INT, {.i64=0}, 0, UINT_MAX, FLAGS},
++    {"n_subsample", "Set interval for frame subsampling used when computing vmaf.",       OFFSET(n_subsample), AV_OPT_TYPE_INT, {.i64=1}, 1, UINT_MAX, FLAGS},
+     { NULL }
+ };
+ 
+ FRAMESYNC_DEFINE_CLASS(libvmaf, LIBVMAFContext, fs);
+ 
+-#define read_frame_fn(type, bits)                                               \
+-    static int read_frame_##bits##bit(float *ref_data, float *main_data,        \
+-                                      float *temp_data, int stride, 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;                                                  \
+-        float factor = 1.f / (1 << (bits - 8));                                 \
+-        \
+-        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] = ref_ptr[j] * factor;                                   \
+-            }                                                                   \
+-            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] = main_ptr[j] * factor;                                  \
+-            }                                                                   \
+-            main_ptr += main_stride / sizeof(*main_ptr);                        \
+-            ptr += stride / sizeof(*ptr);                                       \
+-        }                                                                       \
+-    }                                                                           \
+-    \
+-    ret = !s->frame_set;                                                        \
+-    \
+-    av_frame_unref(s->gref);                                                    \
+-    av_frame_unref(s->gmain);                                                   \
+-    s->frame_set = 0;                                                           \
+-    \
+-    pthread_cond_signal(&s->cond);                                              \
+-    pthread_mutex_unlock(&s->lock);                                             \
+-    \
+-    if (ret) {                                                                  \
+-        return 2;                                                               \
+-    }                                                                           \
+-    \
+-    return 0;                                                                   \
++static enum VmafPixelFormat pix_fmt_map(enum AVPixelFormat av_pix_fmt)
++{
++    switch (av_pix_fmt) {
++    case AV_PIX_FMT_YUV420P:
++    case AV_PIX_FMT_YUV420P10LE:
++    case AV_PIX_FMT_YUV420P12LE:
++    case AV_PIX_FMT_YUV420P16LE:
++        return VMAF_PIX_FMT_YUV420P;
++    case AV_PIX_FMT_YUV422P:
++    case AV_PIX_FMT_YUV422P10LE:
++    case AV_PIX_FMT_YUV422P12LE:
++    case AV_PIX_FMT_YUV422P16LE:
++        return VMAF_PIX_FMT_YUV422P;
++    case AV_PIX_FMT_YUV444P:
++    case AV_PIX_FMT_YUV444P10LE:
++    case AV_PIX_FMT_YUV444P12LE:
++    case AV_PIX_FMT_YUV444P16LE:
++        return VMAF_PIX_FMT_YUV444P;
++    default:
++        return VMAF_PIX_FMT_UNKNOWN;
++    }
+ }
+ 
+-read_frame_fn(uint8_t, 8);
+-read_frame_fn(uint16_t, 10);
++static int copy_picture_data(AVFrame *src, VmafPicture *dst, unsigned bpc)
++{
++    int err = vmaf_picture_alloc(dst, pix_fmt_map(src->format), bpc,
++                                 src->width, src->height);
++    if (err) return AVERROR(ENOMEM);
+ 
+-static void compute_vmaf_score(LIBVMAFContext *s)
++    for (unsigned i = 0; i < 3; i++) {
++        uint8_t *src_data = src->data[i];
++        uint8_t *dst_data = dst->data[i];
++        for (unsigned j = 0; j < dst->h[i]; j++) {
++            memcpy(dst_data, src_data, sizeof(*dst_data) * dst->w[i]);
++            src_data += src->linesize[i];
++            dst_data += dst->stride[i];
++        }
++    }
++
++    return 0;
++}
++
++static int do_vmaf(FFFrameSync *fs)
+ {
+-    int (*read_frame)(float *ref_data, float *main_data, float *temp_data,
+-                      int stride, void *ctx);
+-    char *format;
++    AVFilterContext *ctx = fs->parent;
++    LIBVMAFContext *s = ctx->priv;
++    AVFrame *ref, *dist;
++    int err = 0;
+ 
+-    if (s->desc->comp[0].depth <= 8) {
+-        read_frame = read_frame_8bit;
+-    } else {
+-        read_frame = read_frame_10bit;
++    int ret = ff_framesync_dualinput_get(fs, &dist, &ref);
++    if (ret < 0)
++        return ret;
++    if (ctx->is_disabled || !ref)
++        return ff_filter_frame(ctx->outputs[0], dist);
++
++    VmafPicture pic_ref;
++    err = copy_picture_data(ref, &pic_ref, s->bpc);
++    if (err) {
++        av_log(s, AV_LOG_ERROR, "problem during vmaf_picture_alloc.\n");
++        return AVERROR(ENOMEM);
+     }
+ 
+-    format = (char *) s->desc->name;
++    VmafPicture pic_dist;
++    err = copy_picture_data(dist, &pic_dist, s->bpc);
++    if (err) {
++        av_log(s, AV_LOG_ERROR, "problem during vmaf_picture_alloc.\n");
++        vmaf_picture_unref(&pic_ref);
++        return AVERROR(ENOMEM);
++    }
+ 
+-    s->error = compute_vmaf(&s->vmaf_score, 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,
+-                            s->n_threads, s->n_subsample, s->enable_conf_interval);
++    err = vmaf_read_pictures(s->vmaf, &pic_ref, &pic_dist, s->frame_cnt++);
++    if (err) {
++        av_log(s, AV_LOG_ERROR, "problem during vmaf_read_pictures.\n");
++        return AVERROR(EINVAL);
++    }
++
++    return ff_filter_frame(ctx->outputs[0], dist);
+ }
+ 
+-static void *call_vmaf(void *ctx)
++static AVDictionary **delimited_dict_parse(char *str, unsigned *cnt)
+ {
+-    LIBVMAFContext *s = (LIBVMAFContext *) ctx;
+-    compute_vmaf_score(s);
+-    if (!s->error) {
+-        av_log(ctx, AV_LOG_INFO, "VMAF score: %f\n",s->vmaf_score);
+-    } else {
+-        pthread_mutex_lock(&s->lock);
+-        pthread_cond_signal(&s->cond);
+-        pthread_mutex_unlock(&s->lock);
++    int err = 0;
++    if (!str) return NULL;
++
++    *cnt = 1;
++    for (char *p = str; *p; p++) {
++        if (*p == '|')
++            (*cnt)++;
+     }
+-    pthread_exit(NULL);
++
++    AVDictionary **dict = av_calloc(*cnt, sizeof(*dict));
++    if (!dict) goto fail;
++
++    char *str_copy = av_strdup(str);
++    if (!str_copy) goto fail;
++
++    char *saveptr = NULL;
++    for (unsigned i = 0; i < *cnt; i++) {
++        char *s = av_strtok(i == 0 ? str_copy : NULL, "|", &saveptr);
++        err = av_dict_parse_string(&dict[i], s, "=", ":", 0);
++        if (err) goto fail;
++    }
++
++    av_free(str_copy);
++    return dict;
++
++fail:
++    if (dict) {
++        for (unsigned i = 0; i < *cnt; i++) {
++            if (dict[i])
++                av_dict_free(&dict[i]);
++        }
++        av_free(dict);
++    }
++    if (str_copy)
++        av_free(str_copy);
++    *cnt = 0;
+     return NULL;
+ }
+ 
+-static int do_vmaf(FFFrameSync *fs)
++static int parse_features(AVFilterContext *ctx)
+ {
+-    AVFilterContext *ctx = fs->parent;
+     LIBVMAFContext *s = ctx->priv;
+-    AVFrame *master, *ref;
+-    int ret;
++    if (!s->feature_cfg) return 0;
+ 
+-    ret = ff_framesync_dualinput_get(fs, &master, &ref);
+-    if (ret < 0)
+-        return ret;
+-    if (!ref)
+-        return ff_filter_frame(ctx->outputs[0], master);
++    int err = 0;
+ 
+-    pthread_mutex_lock(&s->lock);
++    unsigned dict_cnt = 0;
++    AVDictionary **dict = delimited_dict_parse(s->feature_cfg, &dict_cnt);
++    if (!dict) {
++        av_log(ctx, AV_LOG_ERROR,
++               "could not parse feature config: %s\n", s->feature_cfg);
++        return AVERROR(EINVAL);
++    }
+ 
+-    while (s->frame_set && !s->error) {
+-        pthread_cond_wait(&s->cond, &s->lock);
++    for (unsigned i = 0; i < dict_cnt; i++) {
++        char *feature_name = NULL;
++        VmafFeatureDictionary *feature_opts_dict = NULL;
++        AVDictionaryEntry *e = NULL;
++
++        while (e = av_dict_get(dict[i], "", e, AV_DICT_IGNORE_SUFFIX)) {
++            if (av_stristr(e->key, "name")) {
++                feature_name = e->value;
++                continue;
++            }
++
++            err = vmaf_feature_dictionary_set(&feature_opts_dict, e->key,
++                                              e->value);
++            if (err) {
++                av_log(ctx, AV_LOG_ERROR,
++                       "could not set feature option: %s.%s=%s\n",
++                       feature_name, e->key, e->value);
++                goto exit;
++            }
++        }
++
++        err = vmaf_use_feature(s->vmaf, feature_name, feature_opts_dict);
++        if (err) {
++            av_log(ctx, AV_LOG_ERROR,
++                   "problem during vmaf_use_feature: %s\n", feature_name);
++            goto exit;
++        }
+     }
+ 
+-    if (s->error) {
++exit:
++    for (unsigned i = 0; i < dict_cnt; i++) {
++        if (dict[i])
++            av_dict_free(&dict[i]);
++    }
++    av_free(dict);
++    return err;
++}
++
++static int parse_models(AVFilterContext *ctx)
++{
++    LIBVMAFContext *s = ctx->priv;
++    if (!s->model_cfg) return 0;
++
++    int err = 0;
++
++    unsigned dict_cnt = 0;
++    AVDictionary **dict = delimited_dict_parse(s->model_cfg, &dict_cnt);
++    if (!dict) {
+         av_log(ctx, AV_LOG_ERROR,
+-               "libvmaf encountered an error, check log for details\n");
+-        pthread_mutex_unlock(&s->lock);
++               "could not parse model config: %s\n", s->model_cfg);
+         return AVERROR(EINVAL);
+     }
+ 
+-    av_frame_ref(s->gref, ref);
+-    av_frame_ref(s->gmain, master);
++    s->model_cnt = dict_cnt;
++    s->model = av_calloc(s->model_cnt, sizeof(*s->model));
++    if (!s->model) return AVERROR(ENOMEM);
+ 
+-    s->frame_set = 1;
++    for (unsigned i = 0; i < dict_cnt; i++) {
++        VmafModelConfig model_cfg = {0};
++        AVDictionaryEntry *e = NULL;
++        char *version, *path;
+ 
+-    pthread_cond_signal(&s->cond);
+-    pthread_mutex_unlock(&s->lock);
++        while (e = av_dict_get(dict[i], "", e, AV_DICT_IGNORE_SUFFIX)) {
++            if (av_stristr(e->key, "disable_clip")) {
++                model_cfg.flags |= av_stristr(e->value, "true") ?
++                    VMAF_MODEL_FLAG_DISABLE_CLIP : 0;
++                continue;
++            }
+ 
+-    return ff_filter_frame(ctx->outputs[0], master);
++            if (av_stristr(e->key, "enable_transform")) {
++                model_cfg.flags |= av_stristr(e->value, "true") ?
++                    VMAF_MODEL_FLAG_ENABLE_TRANSFORM : 0;
++                continue;
++            }
++
++            if (av_stristr(e->key, "name")) {
++                model_cfg.name = e->value;
++                continue;
++            }
++
++            if (av_stristr(e->key, "version")) {
++                version = e->value;
++                continue;
++            }
++
++            if (av_stristr(e->key, "path")) {
++                path = e->value;
++                continue;
++            }
++        }
++
++        if (version) {
++            err = vmaf_model_load(&s->model[i], &model_cfg, version);
++            if (err) {
++                av_log(ctx, AV_LOG_ERROR,
++                       "could not load libvmaf model with version: %s\n",
++                       version);
++                goto exit;
++            }
++        }
++
++        if (path && !s->model[i]) {
++            err = vmaf_model_load_from_path(&s->model[i], &model_cfg, path);
++            if (err) {
++                av_log(ctx, AV_LOG_ERROR,
++                       "could not load libvmaf model with path: %s\n",
++                       e->value);
++                goto exit;
++            }
++        }
++
++        if (!s->model[i]) {
++            av_log(ctx, AV_LOG_ERROR,
++                   "could not load libvmaf model with config: %s\n",
++                   s->model_cfg);
++            goto exit;
++        }
++
++        e = NULL;
++        while (e = av_dict_get(dict[i], "", e, AV_DICT_IGNORE_SUFFIX)) {
++            char *feature_opt = NULL;
++            char *feature_name = av_strtok(e->key, ".", &feature_opt);
++            if (!feature_opt) continue;
++
++            VmafFeatureDictionary *feature_opts_dict = NULL;
++            err = vmaf_feature_dictionary_set(&feature_opts_dict,
++                                              feature_opt, e->value);
++            if (err) {
++                av_log(ctx, AV_LOG_ERROR,
++                       "could not set feature option: %s.%s=%s\n",
++                       feature_name, feature_opt, e->value);
++                err = AVERROR(EINVAL);
++                goto exit;
++            }
++
++            err = vmaf_model_feature_overload(s->model[i], feature_name,
++                                              feature_opts_dict);
++            if (err) {
++                av_log(ctx, AV_LOG_ERROR,
++                       "could not overload feature: %s\n", feature_name);
++                err = AVERROR(EINVAL);
++                goto exit;
++            }
++        }
++    }
++
++    for (unsigned i = 0; i < s->model_cnt; i++) {
++        err = vmaf_use_features_from_model(s->vmaf, s->model[i]);
++        if (err) {
++            av_log(ctx, AV_LOG_ERROR,
++                   "problem during vmaf_use_features_from_model\n");
++            err = AVERROR(EINVAL);
++            goto exit;
++        }
++    }
++
++exit:
++    for (unsigned i = 0; i < dict_cnt; i++) {
++        if (dict[i])
++            av_dict_free(&dict[i]);
++    }
++    av_free(dict);
++    return err;
+ }
+ 
++static enum VmafLogLevel log_level_map(int log_level)
++{
++    switch (log_level) {
++    case AV_LOG_QUIET:
++        return VMAF_LOG_LEVEL_NONE;
++    case AV_LOG_ERROR:
++        return VMAF_LOG_LEVEL_ERROR;
++    case AV_LOG_WARNING:
++        return VMAF_LOG_LEVEL_WARNING;
++    case AV_LOG_INFO:
++        return VMAF_LOG_LEVEL_INFO;
++    case AV_LOG_DEBUG:
++        return VMAF_LOG_LEVEL_DEBUG;
++    default:
++        return VMAF_LOG_LEVEL_INFO;
++    }
++}
++
+ static av_cold int init(AVFilterContext *ctx)
+ {
+     LIBVMAFContext *s = ctx->priv;
++    int err = 0;
+ 
+-    s->gref = av_frame_alloc();
+-    s->gmain = av_frame_alloc();
+-    if (!s->gref || !s->gmain)
+-        return AVERROR(ENOMEM);
++    VmafConfiguration cfg = {
++        .log_level = log_level_map(av_log_get_level()),
++        .n_subsample = s->n_subsample,
++        .n_threads = s->n_threads,
++    };
+ 
+-    s->error = 0;
++    err = vmaf_init(&s->vmaf, cfg);
++    if (err) return AVERROR(EINVAL);
+ 
+-    s->vmaf_thread_created = 0;
+-    pthread_mutex_init(&s->lock, NULL);
+-    pthread_cond_init (&s->cond, NULL);
++    err = parse_models(ctx);
++    if (err) return err;
+ 
++    err = parse_features(ctx);
++    if (err) return err;
++
+     s->fs.on_event = do_vmaf;
+     return 0;
+ }
+@@ -251,8 +412,9 @@ static av_cold int init(AVFilterContext *ctx)
+ 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_YUV420P, AV_PIX_FMT_YUV420P16LE,
++        AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV422P16LE,
++        AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV444P16LE,
+         AV_PIX_FMT_NONE
+     };
+ 
+@@ -262,33 +424,32 @@ static int query_formats(AVFilterContext *ctx)
+     return ff_set_common_formats(ctx, fmts_list);
+ }
+ 
+-
+ static int config_input_ref(AVFilterLink *inlink)
+ {
+-    AVFilterContext *ctx  = inlink->dst;
++    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);
++    int err = 0;
++
++    if (ctx->inputs[0]->w != ctx->inputs[1]->w) {
++        av_log(ctx, AV_LOG_ERROR, "input width must match.\n");
++        err |= AVERROR(EINVAL);
+     }
++
++    if (ctx->inputs[0]->h != ctx->inputs[1]->h) {
++        av_log(ctx, AV_LOG_ERROR, "input height must match.\n");
++        err |= 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);
++        av_log(ctx, AV_LOG_ERROR, "input pix_fmt must match.\n");
++        err |= AVERROR(EINVAL);
+     }
+ 
+-    s->desc = av_pix_fmt_desc_get(inlink->format);
+-    s->width = ctx->inputs[0]->w;
+-    s->height = ctx->inputs[0]->h;
++    if (err) return err;
+ 
+-    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);
+-    }
+-    s->vmaf_thread_created = 1;
++    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
++    s->bpc = desc->comp[0].depth;
+ 
+     return 0;
+ }
+@@ -320,35 +481,72 @@ static int activate(AVFilterContext *ctx)
+     return ff_framesync_activate(&s->fs);
+ }
+ 
++static enum VmafOutputFormat log_fmt_map(const char *log_fmt)
++{
++    if (av_stristr(log_fmt, "xml"))
++        return VMAF_OUTPUT_FORMAT_XML;
++    if (av_stristr(log_fmt, "json"))
++        return VMAF_OUTPUT_FORMAT_JSON;
++    if (av_stristr(log_fmt, "csv"))
++        return VMAF_OUTPUT_FORMAT_CSV;
++    if (av_stristr(log_fmt, "sub"))
++        return VMAF_OUTPUT_FORMAT_SUB;
++
++    return VMAF_OUTPUT_FORMAT_XML;
++}
++
+ static av_cold void uninit(AVFilterContext *ctx)
+ {
+     LIBVMAFContext *s = ctx->priv;
+ 
++    int err = 0;
++
+     ff_framesync_uninit(&s->fs);
+ 
+-    pthread_mutex_lock(&s->lock);
+-    s->eof = 1;
+-    pthread_cond_signal(&s->cond);
+-    pthread_mutex_unlock(&s->lock);
++    if (!s->frame_cnt) goto clean_up;
+ 
+-    if (s->vmaf_thread_created)
+-    {
+-        pthread_join(s->vmaf_thread, NULL);
+-        s->vmaf_thread_created = 0;
++    err = vmaf_read_pictures(s->vmaf, NULL, NULL, 0);
++    if (err) {
++        av_log(ctx, AV_LOG_ERROR,
++               "problem flushing libvmaf context.\n");
+     }
+ 
+-    av_frame_free(&s->gref);
+-    av_frame_free(&s->gmain);
++    for (unsigned i = 0; i < s->model_cnt; i++) {
++        double vmaf_score;
++        err = vmaf_score_pooled(s->vmaf, s->model[i], VMAF_POOL_METHOD_MEAN,
++                                &vmaf_score, 0, s->frame_cnt - 1);
++        if (err) {
++            av_log(ctx, AV_LOG_ERROR,
++                   "problem getting pooled vmaf score.\n");
++        }
+ 
+-    pthread_mutex_destroy(&s->lock);
+-    pthread_cond_destroy(&s->cond);
++        av_log(ctx, AV_LOG_INFO, "VMAF score: %f\n", vmaf_score);
++    }
++
++    if (s->vmaf) {
++        if (s->log_path && !err)
++            vmaf_write_output(s->vmaf, s->log_path, log_fmt_map(s->log_fmt));
++    }
++
++clean_up:
++    if (s->model) {
++        for (unsigned i = 0; i < s->model_cnt; i++) {
++            if (s->model[i])
++                vmaf_model_destroy(s->model[i]);
++        }
++        av_free(s->model);
++    }
++
++    if (s->vmaf)
++        vmaf_close(s->vmaf);
+ }
+ 
+ static const AVFilterPad libvmaf_inputs[] = {
+     {
+         .name         = "main",
+         .type         = AVMEDIA_TYPE_VIDEO,
+-    },{
++    },
++    {
+         .name         = "reference",
+         .type         = AVMEDIA_TYPE_VIDEO,
+         .config_props = config_input_ref,