FFmpeg
vf_vmafmotion.c
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2017 Ronald S. Bultje <rsbultje@gmail.com>
3  * Copyright (c) 2017 Ashish Pratap Singh <ashk43712@gmail.com>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21 
22 /**
23  * @file
24  * Calculate VMAF Motion score.
25  */
26 
27 #include "libavutil/opt.h"
28 #include "libavutil/pixdesc.h"
29 #include "avfilter.h"
30 #include "formats.h"
31 #include "internal.h"
32 #include "vmaf_motion.h"
33 
34 #define BIT_SHIFT 15
35 
36 static const float FILTER_5[5] = {
37  0.054488685,
38  0.244201342,
39  0.402619947,
40  0.244201342,
41  0.054488685
42 };
43 
44 typedef struct VMAFMotionContext {
45  const AVClass *class;
47  FILE *stats_file;
50 
51 #define OFFSET(x) offsetof(VMAFMotionContext, x)
52 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
53 
54 static const AVOption vmafmotion_options[] = {
55  {"stats_file", "Set file where to store per-frame difference information", OFFSET(stats_file_str), AV_OPT_TYPE_STRING, {.str=NULL}, 0, 0, FLAGS },
56  { NULL }
57 };
58 
59 AVFILTER_DEFINE_CLASS(vmafmotion);
60 
61 static uint64_t image_sad(const uint16_t *img1, const uint16_t *img2, int w,
62  int h, ptrdiff_t _img1_stride, ptrdiff_t _img2_stride)
63 {
64  ptrdiff_t img1_stride = _img1_stride / sizeof(*img1);
65  ptrdiff_t img2_stride = _img2_stride / sizeof(*img2);
66  uint64_t sum = 0;
67  int i, j;
68 
69  for (i = 0; i < h; i++) {
70  for (j = 0; j < w; j++) {
71  sum += abs(img1[j] - img2[j]);
72  }
73  img1 += img1_stride;
74  img2 += img2_stride;
75  }
76 
77  return sum;
78 }
79 
80 static void convolution_x(const uint16_t *filter, int filt_w, const uint16_t *src,
81  uint16_t *dst, int w, int h, ptrdiff_t _src_stride,
82  ptrdiff_t _dst_stride)
83 {
84  ptrdiff_t src_stride = _src_stride / sizeof(*src);
85  ptrdiff_t dst_stride = _dst_stride / sizeof(*dst);
86  int radius = filt_w / 2;
87  int borders_left = radius;
88  int borders_right = w - (filt_w - radius);
89  int i, j, k;
90  int sum = 0;
91 
92  for (i = 0; i < h; i++) {
93  for (j = 0; j < borders_left; j++) {
94  sum = 0;
95  for (k = 0; k < filt_w; k++) {
96  int j_tap = FFABS(j - radius + k);
97  if (j_tap >= w) {
98  j_tap = w - (j_tap - w + 1);
99  }
100  sum += filter[k] * src[i * src_stride + j_tap];
101  }
102  dst[i * dst_stride + j] = sum >> BIT_SHIFT;
103  }
104 
105  for (j = borders_left; j < borders_right; j++) {
106  int sum = 0;
107  for (k = 0; k < filt_w; k++) {
108  sum += filter[k] * src[i * src_stride + j - radius + k];
109  }
110  dst[i * dst_stride + j] = sum >> BIT_SHIFT;
111  }
112 
113  for (j = borders_right; j < w; j++) {
114  sum = 0;
115  for (k = 0; k < filt_w; k++) {
116  int j_tap = FFABS(j - radius + k);
117  if (j_tap >= w) {
118  j_tap = w - (j_tap - w + 1);
119  }
120  sum += filter[k] * src[i * src_stride + j_tap];
121  }
122  dst[i * dst_stride + j] = sum >> BIT_SHIFT;
123  }
124  }
125 }
126 
127 #define conv_y_fn(type, bits) \
128 static void convolution_y_##bits##bit(const uint16_t *filter, int filt_w, \
129  const uint8_t *_src, uint16_t *dst, \
130  int w, int h, ptrdiff_t _src_stride, \
131  ptrdiff_t _dst_stride) \
132 { \
133  const type *src = (const type *) _src; \
134  ptrdiff_t src_stride = _src_stride / sizeof(*src); \
135  ptrdiff_t dst_stride = _dst_stride / sizeof(*dst); \
136  int radius = filt_w / 2; \
137  int borders_top = radius; \
138  int borders_bottom = h - (filt_w - radius); \
139  int i, j, k; \
140  int sum = 0; \
141  \
142  for (i = 0; i < borders_top; i++) { \
143  for (j = 0; j < w; j++) { \
144  sum = 0; \
145  for (k = 0; k < filt_w; k++) { \
146  int i_tap = FFABS(i - radius + k); \
147  if (i_tap >= h) { \
148  i_tap = h - (i_tap - h + 1); \
149  } \
150  sum += filter[k] * src[i_tap * src_stride + j]; \
151  } \
152  dst[i * dst_stride + j] = sum >> bits; \
153  } \
154  } \
155  for (i = borders_top; i < borders_bottom; i++) { \
156  for (j = 0; j < w; j++) { \
157  sum = 0; \
158  for (k = 0; k < filt_w; k++) { \
159  sum += filter[k] * src[(i - radius + k) * src_stride + j]; \
160  } \
161  dst[i * dst_stride + j] = sum >> bits; \
162  } \
163  } \
164  for (i = borders_bottom; i < h; i++) { \
165  for (j = 0; j < w; j++) { \
166  sum = 0; \
167  for (k = 0; k < filt_w; k++) { \
168  int i_tap = FFABS(i - radius + k); \
169  if (i_tap >= h) { \
170  i_tap = h - (i_tap - h + 1); \
171  } \
172  sum += filter[k] * src[i_tap * src_stride + j]; \
173  } \
174  dst[i * dst_stride + j] = sum >> bits; \
175  } \
176  } \
177 }
178 
180 conv_y_fn(uint16_t, 10)
181 
182 static void vmafmotiondsp_init(VMAFMotionDSPContext *dsp, int bpp) {
183  dsp->convolution_x = convolution_x;
184  dsp->convolution_y = bpp == 10 ? convolution_y_10bit : convolution_y_8bit;
185  dsp->sad = image_sad;
186 }
187 
189 {
190  double score;
191 
192  s->vmafdsp.convolution_y(s->filter, 5, ref->data[0], s->temp_data,
193  s->width, s->height, ref->linesize[0], s->stride);
194  s->vmafdsp.convolution_x(s->filter, 5, s->temp_data, s->blur_data[0],
195  s->width, s->height, s->stride, s->stride);
196 
197  if (!s->nb_frames) {
198  score = 0.0;
199  } else {
200  uint64_t sad = s->vmafdsp.sad(s->blur_data[1], s->blur_data[0],
201  s->width, s->height, s->stride, s->stride);
202  // the output score is always normalized to 8 bits
203  score = (double) (sad * 1.0 / (s->width * s->height << (BIT_SHIFT - 8)));
204  }
205 
206  FFSWAP(uint16_t *, s->blur_data[0], s->blur_data[1]);
207  s->nb_frames++;
208  s->motion_sum += score;
209 
210  return score;
211 }
212 
213 static void set_meta(AVDictionary **metadata, const char *key, float d)
214 {
215  char value[128];
216  snprintf(value, sizeof(value), "%0.2f", d);
217  av_dict_set(metadata, key, value, 0);
218 }
219 
221 {
222  VMAFMotionContext *s = ctx->priv;
223  double score;
224 
225  score = ff_vmafmotion_process(&s->data, ref);
226  set_meta(&ref->metadata, "lavfi.vmafmotion.score", score);
227  if (s->stats_file) {
228  fprintf(s->stats_file,
229  "n:%"PRId64" motion:%0.2lf\n", s->data.nb_frames, score);
230  }
231 }
232 
233 
235  int w, int h, enum AVPixelFormat fmt)
236 {
237  size_t data_sz;
238  int i;
240 
241  if (w < 3 || h < 3)
242  return AVERROR(EINVAL);
243 
244  s->width = w;
245  s->height = h;
246  s->stride = FFALIGN(w * sizeof(uint16_t), 32);
247 
248  data_sz = (size_t) s->stride * h;
249  if (!(s->blur_data[0] = av_malloc(data_sz)) ||
250  !(s->blur_data[1] = av_malloc(data_sz)) ||
251  !(s->temp_data = av_malloc(data_sz))) {
252  return AVERROR(ENOMEM);
253  }
254 
255  for (i = 0; i < 5; i++) {
256  s->filter[i] = lrint(FILTER_5[i] * (1 << BIT_SHIFT));
257  }
258 
259  vmafmotiondsp_init(&s->vmafdsp, desc->comp[0].depth);
260 
261  return 0;
262 }
263 
265 {
266  AVFilterFormats *fmts_list = NULL;
267  int format, ret;
268 
269  for (format = 0; av_pix_fmt_desc_get(format); format++) {
272  (desc->flags & AV_PIX_FMT_FLAG_PLANAR || desc->nb_components == 1) &&
273  (!(desc->flags & AV_PIX_FMT_FLAG_BE) == !HAVE_BIGENDIAN || desc->comp[0].depth == 8) &&
274  (desc->comp[0].depth == 8 || desc->comp[0].depth == 10) &&
275  (ret = ff_add_format(&fmts_list, format)) < 0)
276  return ret;
277  }
278 
279  return ff_set_common_formats(ctx, fmts_list);
280 }
281 
283 {
284  AVFilterContext *ctx = inlink->dst;
285  VMAFMotionContext *s = ctx->priv;
286 
287  return ff_vmafmotion_init(&s->data, ctx->inputs[0]->w,
288  ctx->inputs[0]->h, ctx->inputs[0]->format);
289 }
290 
292 {
293  av_free(s->blur_data[0]);
294  av_free(s->blur_data[1]);
295  av_free(s->temp_data);
296 
297  return s->nb_frames > 0 ? s->motion_sum / s->nb_frames : 0.0;
298 }
299 
301 {
302  AVFilterContext *ctx = inlink->dst;
304  return ff_filter_frame(ctx->outputs[0], ref);
305 }
306 
308 {
309  VMAFMotionContext *s = ctx->priv;
310 
311  if (s->stats_file_str) {
312  if (!strcmp(s->stats_file_str, "-")) {
313  s->stats_file = stdout;
314  } else {
315  s->stats_file = fopen(s->stats_file_str, "w");
316  if (!s->stats_file) {
317  int err = AVERROR(errno);
318  char buf[128];
319  av_strerror(err, buf, sizeof(buf));
320  av_log(ctx, AV_LOG_ERROR, "Could not open stats file %s: %s\n",
321  s->stats_file_str, buf);
322  return err;
323  }
324  }
325  }
326 
327  return 0;
328 }
329 
331 {
332  VMAFMotionContext *s = ctx->priv;
333  double avg_motion = ff_vmafmotion_uninit(&s->data);
334 
335  if (s->data.nb_frames > 0) {
336  av_log(ctx, AV_LOG_INFO, "VMAF Motion avg: %.3f\n", avg_motion);
337  }
338 
339  if (s->stats_file && s->stats_file != stdout)
340  fclose(s->stats_file);
341 }
342 
343 static const AVFilterPad vmafmotion_inputs[] = {
344  {
345  .name = "reference",
346  .type = AVMEDIA_TYPE_VIDEO,
347  .filter_frame = filter_frame,
348  .config_props = config_input_ref,
349  },
350  { NULL }
351 };
352 
353 static const AVFilterPad vmafmotion_outputs[] = {
354  {
355  .name = "default",
356  .type = AVMEDIA_TYPE_VIDEO,
357  },
358  { NULL }
359 };
360 
362  .name = "vmafmotion",
363  .description = NULL_IF_CONFIG_SMALL("Calculate the VMAF Motion score."),
364  .init = init,
365  .uninit = uninit,
366  .query_formats = query_formats,
367  .priv_size = sizeof(VMAFMotionContext),
368  .priv_class = &vmafmotion_class,
371 };
AVPixelFormat
AVPixelFormat
Pixel format.
Definition: pixfmt.h:64
AVERROR
Filter the word “frame” indicates either a video frame or a group of audio as stored in an AVFrame structure Format for each input and each output the list of supported formats For video that means pixel format For audio that means channel sample they are references to shared objects When the negotiation mechanism computes the intersection of the formats supported at each end of a all references to both lists are replaced with a reference to the intersection And when a single format is eventually chosen for a link amongst the remaining all references to the list are updated That means that if a filter requires that its input and output have the same format amongst a supported all it has to do is use a reference to the same list of formats query_formats can leave some formats unset and return AVERROR(EAGAIN) to cause the negotiation mechanism toagain later. That can be used by filters with complex requirements to use the format negotiated on one link to set the formats supported on another. Frame references ownership and permissions
opt.h
FFSWAP
#define FFSWAP(type, a, b)
Definition: common.h:108
ff_filter_frame
int ff_filter_frame(AVFilterLink *link, AVFrame *frame)
Send a frame of data to the next filter.
Definition: avfilter.c:1096
av_pix_fmt_desc_get
const AVPixFmtDescriptor * av_pix_fmt_desc_get(enum AVPixelFormat pix_fmt)
Definition: pixdesc.c:2573
inlink
The exact code depends on how similar the blocks are and how related they are to the and needs to apply these operations to the correct inlink or outlink if there are several Macros are available to factor that when no extra processing is inlink
Definition: filter_design.txt:212
OFFSET
#define OFFSET(x)
Definition: vf_vmafmotion.c:51
AVFrame
This structure describes decoded (raw) audio or video data.
Definition: frame.h:318
pixdesc.h
w
uint8_t w
Definition: llviddspenc.c:39
AVOption
AVOption.
Definition: opt.h:248
filter
filter_frame For filters that do not use the this method is called when a frame is pushed to the filter s input It can be called at any time except in a reentrant way If the input frame is enough to produce then the filter should push the output frames on the output link immediately As an exception to the previous rule if the input frame is enough to produce several output frames then the filter needs output only at least one per link The additional frames can be left buffered in the filter
Definition: filter_design.txt:228
AVDictionary
Definition: dict.c:30
AVFilter::name
const char * name
Filter name.
Definition: avfilter.h:149
ff_vmafmotion_process
double ff_vmafmotion_process(VMAFMotionData *s, AVFrame *ref)
Definition: vf_vmafmotion.c:188
BIT_SHIFT
#define BIT_SHIFT
Definition: vf_vmafmotion.c:34
av_malloc
#define av_malloc(s)
Definition: tableprint_vlc.h:31
AVFilterFormats
A list of supported formats for one end of a filter link.
Definition: formats.h:65
formats.h
VMAFMotionContext::stats_file
FILE * stats_file
Definition: vf_vmafmotion.c:47
filter_frame
static int filter_frame(AVFilterLink *inlink, AVFrame *ref)
Definition: vf_vmafmotion.c:300
av_strerror
int av_strerror(int errnum, char *errbuf, size_t errbuf_size)
Put a description of the AVERROR code errnum in errbuf.
Definition: error.c:105
AV_PIX_FMT_FLAG_HWACCEL
#define AV_PIX_FMT_FLAG_HWACCEL
Pixel format is an HW accelerated format.
Definition: pixdesc.h:140
convolution_x
static void convolution_x(const uint16_t *filter, int filt_w, const uint16_t *src, uint16_t *dst, int w, int h, ptrdiff_t _src_stride, ptrdiff_t _dst_stride)
Definition: vf_vmafmotion.c:80
config_input_ref
static int config_input_ref(AVFilterLink *inlink)
Definition: vf_vmafmotion.c:282
AVFilterPad
A filter pad used for either input or output.
Definition: internal.h:54
lrint
#define lrint
Definition: tablegen.h:53
AV_LOG_ERROR
#define AV_LOG_ERROR
Something went wrong and cannot losslessly be recovered.
Definition: log.h:194
av_cold
#define av_cold
Definition: attributes.h:90
ff_set_common_formats
int ff_set_common_formats(AVFilterContext *ctx, AVFilterFormats *formats)
A helper for query_formats() which sets all links to the same list of formats.
Definition: formats.c:587
VMAFMotionData
Definition: vmaf_motion.h:42
s
#define s(width, name)
Definition: cbs_vp9.c:257
format
Filter the word “frame” indicates either a video frame or a group of audio as stored in an AVFrame structure Format for each input and each output the list of supported formats For video that means pixel format For audio that means channel sample format(the sample packing is implied by the sample format) and sample rate. The lists are not just lists
ff_vmafmotion_init
int ff_vmafmotion_init(VMAFMotionData *s, int w, int h, enum AVPixelFormat fmt)
Definition: vf_vmafmotion.c:234
outputs
static const AVFilterPad outputs[]
Definition: af_acontrast.c:203
ctx
AVFormatContext * ctx
Definition: movenc.c:48
key
const char * key
Definition: hwcontext_opencl.c:168
FFABS
#define FFABS(a)
Absolute value, Note, INT_MIN / INT64_MIN result in undefined behavior as they are not representable ...
Definition: common.h:72
if
if(ret)
Definition: filter_design.txt:179
AVClass
Describe the class of an AVClass context structure.
Definition: log.h:67
NULL
#define NULL
Definition: coverity.c:32
AVFILTER_DEFINE_CLASS
AVFILTER_DEFINE_CLASS(vmafmotion)
vmafmotion_outputs
static const AVFilterPad vmafmotion_outputs[]
Definition: vf_vmafmotion.c:353
src
#define src
Definition: vp8dsp.c:255
ff_add_format
int ff_add_format(AVFilterFormats **avff, int64_t fmt)
Add fmt to the list of media formats contained in *avff.
Definition: formats.c:332
abs
#define abs(x)
Definition: cuda_runtime.h:35
inputs
these buffered frames must be flushed immediately if a new input produces new the filter must not call request_frame to get more It must just process the frame or queue it The task of requesting more frames is left to the filter s request_frame method or the application If a filter has several inputs
Definition: filter_design.txt:243
img1
static uint8_t img1[WIDTH *HEIGHT]
Definition: motion.c:42
ff_vmafmotion_uninit
double ff_vmafmotion_uninit(VMAFMotionData *s)
Definition: vf_vmafmotion.c:291
NULL_IF_CONFIG_SMALL
#define NULL_IF_CONFIG_SMALL(x)
Return NULL if CONFIG_SMALL is true, otherwise the argument without modification.
Definition: internal.h:117
AV_PIX_FMT_FLAG_RGB
#define AV_PIX_FMT_FLAG_RGB
The pixel format contains RGB-like data (as opposed to YUV/grayscale).
Definition: pixdesc.h:148
img2
static uint8_t img2[WIDTH *HEIGHT]
Definition: motion.c:43
AV_PIX_FMT_FLAG_BITSTREAM
#define AV_PIX_FMT_FLAG_BITSTREAM
All values of a component are bit-wise packed end to end.
Definition: pixdesc.h:136
init
static av_cold int init(AVFilterContext *ctx)
Definition: vf_vmafmotion.c:307
AV_LOG_INFO
#define AV_LOG_INFO
Standard information.
Definition: log.h:205
internal.h
i
int i
Definition: input.c:407
image_sad
static uint64_t image_sad(const uint16_t *img1, const uint16_t *img2, int w, int h, ptrdiff_t _img1_stride, ptrdiff_t _img2_stride)
Definition: vf_vmafmotion.c:61
VMAFMotionContext
Definition: vf_vmafmotion.c:44
value
it s the only field you need to keep assuming you have a context There is some magic you don t need to care about around this just let it vf default value
Definition: writing_filters.txt:86
AV_PIX_FMT_FLAG_BE
#define AV_PIX_FMT_FLAG_BE
Pixel format is big-endian.
Definition: pixdesc.h:128
uint8_t
uint8_t
Definition: audio_convert.c:194
AVFilterPad::name
const char * name
Pad name.
Definition: internal.h:60
AVFilter
Filter definition.
Definition: avfilter.h:145
ret
ret
Definition: filter_design.txt:187
ff_vf_vmafmotion
AVFilter ff_vf_vmafmotion
Definition: vf_vmafmotion.c:361
vmaf_motion.h
VMAFMotionContext::data
VMAFMotionData data
Definition: vf_vmafmotion.c:46
query_formats
static int query_formats(AVFilterContext *ctx)
Definition: vf_vmafmotion.c:264
FILTER_5
static const float FILTER_5[5]
Definition: vf_vmafmotion.c:36
avfilter.h
VMAFMotionDSPContext
Definition: vmaf_motion.h:29
ref
static int ref[MAX_W *MAX_W]
Definition: jpeg2000dwt.c:107
AV_PIX_FMT_FLAG_PLANAR
#define AV_PIX_FMT_FLAG_PLANAR
At least one pixel component is not in the first data plane.
Definition: pixdesc.h:144
do_vmafmotion
static void do_vmafmotion(AVFilterContext *ctx, AVFrame *ref)
Definition: vf_vmafmotion.c:220
vmafmotion_inputs
static const AVFilterPad vmafmotion_inputs[]
Definition: vf_vmafmotion.c:343
AVFilterContext
An instance of a filter.
Definition: avfilter.h:341
FLAGS
#define FLAGS
Definition: vf_vmafmotion.c:52
desc
const char * desc
Definition: libsvtav1.c:79
AVMEDIA_TYPE_VIDEO
@ AVMEDIA_TYPE_VIDEO
Definition: avutil.h:201
conv_y_fn
#define conv_y_fn(type, bits)
Definition: vf_vmafmotion.c:127
vmafmotion_options
static const AVOption vmafmotion_options[]
Definition: vf_vmafmotion.c:54
AVPixFmtDescriptor
Descriptor that unambiguously describes how the bits of a pixel are stored in the up to 4 data planes...
Definition: pixdesc.h:81
av_free
#define av_free(p)
Definition: tableprint_vlc.h:34
FFALIGN
#define FFALIGN(x, a)
Definition: macros.h:48
av_dict_set
int av_dict_set(AVDictionary **pm, const char *key, const char *value, int flags)
Set the given entry in *pm, overwriting an existing entry.
Definition: dict.c:70
uninit
static av_cold void uninit(AVFilterContext *ctx)
Definition: vf_vmafmotion.c:330
VMAFMotionContext::stats_file_str
char * stats_file_str
Definition: vf_vmafmotion.c:48
av_log
#define av_log(a,...)
Definition: tableprint_vlc.h:28
h
h
Definition: vp9dsp_template.c:2038
AV_OPT_TYPE_STRING
@ AV_OPT_TYPE_STRING
Definition: opt.h:229
AV_PIX_FMT_FLAG_PAL
#define AV_PIX_FMT_FLAG_PAL
Pixel format has a palette in data[1], values are indexes in this palette.
Definition: pixdesc.h:132
snprintf
#define snprintf
Definition: snprintf.h:34
set_meta
static void set_meta(AVDictionary **metadata, const char *key, float d)
Definition: vf_vmafmotion.c:213