[FFmpeg-devel] [PATCH 4/7] Adds gray floating-point pixel formats.

[Sergey Lavrushkin]

2018-08-04 0:11 GMT+03:00 Michael Niedermayer <michael at niedermayer.cc>:

> On Fri, Aug 03, 2018 at 10:33:00PM +0300, Sergey Lavrushkin wrote:
> > 2018-08-03 16:07 GMT+03:00 Michael Niedermayer <michael at niedermayer.cc>:
> >
> > > On Thu, Aug 02, 2018 at 09:52:45PM +0300, Sergey Lavrushkin wrote:
> > > > This patch adds two floating-point gray formats to use them in sr
> filter
> > > for
> > > > conversion with libswscale. I added conversion from uint gray to
> float
> > > and
> > > > backwards in swscale_unscaled.c, that is enough for sr filter. But
> for
> > > > proper format addition, should I add anything else?
> > > >
> > > > ---
> > > >  libavutil/pixdesc.c           | 22 ++++++++++++++++++
> > > >  libavutil/pixfmt.h            |  5 ++++
> > > >  libswscale/swscale_internal.h |  7 ++++++
> > > >  libswscale/swscale_unscaled.c | 54 ++++++++++++++++++++++++++++++
> > > +++++++++++--
> > > >  libswscale/utils.c            |  5 +++-
> > >
> > > please split this in a patch or libavutil and one for libswscale
> > > they also need some version.h bump
> > >
> >
> > Ok.
> >
> > also fate tests need an update, (make fate) fails otherwise, the update
> > > should
> > > be part of the patch that causes the failure otherwise
> >
> >
> > In one test for these formats I get:
> >
> > filter-pixfmts-scale
> > grayf32be           grayf32le           monob
> >  f01cb0b623357387827902d9d0963435
> >
> > I guess, it is because I only implemented conversion in swscale_unscaled.
> > What can I do to fix it? Should I implement conversion for scaling or
> maybe
> > change something in the test, so it would not check these formats (if it
> is
> > possible).
> > Anyway, I need to know what changes should I do and where.
>
> well, swscale shouldnt really have formats only half supported
> so for any supported format in and out it should work with any
> width / height in / out
>
> Theres a wide range of possibilities how to implement this.
> The correct / ideal way is of course to implement a full floating point
> path
> for scaling along side the integer code.
> a simpler aprouch would be to convert from/to float to/from  integers and
> use
> the existing code. (this of course has the disadvantage of loosing
> precission)
>

Well, I want to implement simpler approach, as I still have to finish
correcting sr filter.
But I need some explanations regarding what I should add. If I understand
correcly,
I need to add conversion from float to the ff_sws_init_input_funcs function
in libswscale/input.c
and conversion to float to the ff_sws_init_output_funcs function in
libswscale/output.c
If I am not mistaken, in the first case I need to provide c->lumToYV12 and
in the second case -
yuv2plane1 and yuv2planeX. So, in the first case, to what format should I
add
conversion, specifically what number of bits per pixel should be used? As I
look through other
conversion functions, it seems that somewhere uint8 is used and somewhere -
uint16.
Is it somehow determined later during scaling? If I am going to convert to
uint8 from
my float format, should I define it somewhere, that I am converting to
uint8?
And in the second case, I don't completely understand, what these two
functions are
doing, especially tha last one with filters. Is it also just simple
conversions or
these functions also cover something else? And in their descriptions it is
written, that:

 * @param src     scaled source data, 15 bits for 8-10-bit output,
 *                19 bits for 16-bit output (in int32_t)
 * @param dest    pointer to the output plane. For >8-bit
 *                output, this is in uint16_t

In my case, the output is 32-bit. Does this mean, that float type,
basically, is not
supported and I also have to modify something in scaling? If so, what
should I add?



> [...]
> > > +    const uint8_t *srcPtr = src[0];
> > > > +    float *dstPtr = (float *)(dst[0] + dstStride[0] * srcSliceY);
> > > > +
> > > > +    for (y = 0; y < srcSliceH; ++y){
> > > > +        for (x = 0; x < c->srcW; ++x){
> > > > +            dstPtr[x] = (float)srcPtr[x] / 255.0f;
> > >
> > > division is slow. This should either be a multiplication with the
> > > inverse or a LUT with 8bit index changing to float.
> > >
> > > The faster of them should be used
> > >
> >
> > LUT seems to be faster. Can I place it in SwsContext and initialize it in
> > sws_init_context when necessary?
>
> yes of course
>
> thanks
>
>