[FFmpeg-devel] [GSOC] [PATCH] SRCNN filter

[Sergey Lavrushkin]

> [...]
> > +#define OFFSET(x) offsetof(SRCNNContext, x)
> > +#define FLAGS AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM
> > +static const AVOption srcnn_options[] = {
> > +    { "config_file", "path to configuration file with network
> parameters", OFFSET(config_file_path), AV_OPT_TYPE_STRING, {.str=NULL}, 0,
> 0, FLAGS },
> > +    { NULL }
> > +};
> > +
> > +AVFILTER_DEFINE_CLASS(srcnn);
> > +
> > +#define CHECK_FILE(file)    if (ferror(file) || feof(file)){ \
> > +                                av_log(context, AV_LOG_ERROR, "error
> reading configuration file\n");\
> > +                                fclose(file); \
> > +                                return AVERROR(EIO); \
> > +                            }
> > +
> > +#define CHECK_ALLOCATION(conv, file)    if
> (allocate_and_read_convolution_data(&conv, file)){ \
> > +                                            av_log(context,
> AV_LOG_ERROR, "could not allocate memory for convolutions\n"); \
> > +                                            fclose(file); \
> > +                                            return AVERROR(ENOMEM); \
> > +                                        }
> > +
>
> > +static int allocate_and_read_convolution_data(Convolution* conv, FILE*
> config_file)
> > +{
> > +    int32_t kernel_size = conv->output_channels * conv->size *
> conv->size * conv->input_channels;
> > +    conv->kernel = av_malloc(kernel_size * sizeof(double));
> > +    if (!conv->kernel){
> > +        return 1;
>
> this should return an AVERROR code for consistency with the rest of
> the codebase
>

Ok.


> > +    }
>
> > +    fread(conv->kernel, sizeof(double), kernel_size, config_file);
>
> directly reading data types is not portable, it would for example be
> endian specific
> and using avio for reading may be better, though fread is as far as iam
> concerned also ok
>

Ok, I understand the problem, but I have not really worked with it before,
so I need an advice of how to properly fix it. If I understand correctly,
for
int32_t I need to check endianness and reverse bytes if necessary. But with
doubles it is more complicated. Should I write a IEEE 754 converter from
binary
to double or maybe I can somehow check IEEE 754 doubles support and
depending
on it either stick to the default network weights, or just read bytes and
check
endianness, if IEEE 754 doubles are supported? Or maybe avio provide some
utility to deal with this problem?


> [...]
> > +/**
> > + * @file
> > + * Default cnn weights for x2 upsampling with srcnn filter.
> > + */
> > +
> > +/// First convolution kernel
>
> > +static double conv1_kernel[] = {
>
> static data should be also const, otherwise it may be changed and could
> cause
> thread saftey issues
>

Ok, I just wanted to not allocate additional memory in case of using
default weights.