[FFmpeg-devel] GSoC Weekly report (libswscale)

[Michael Niedermayer]

On Tue, Jul 28, 2015 at 06:06:04PM -0300, Pedro Arthur wrote:
> All these weird bugs were caused by the mmx code expecting the chroma U
> and V buffer to be contiguous in memory. As I was allocating the slice
> lines U
> and V separately the mmx code was using some random memory.
> In the attached patch there is a fix for it.
> I also added some documentation to the new structs.

very good, this works much better
one remaining issue i could find is
./ffplay -i matrixbench_mpeg2.mpg -vf scale=96:96,format=nv12,scale=128:128:flags=1
shows a green line at the right border
it happens also with some other formats but seems specific to flags=1
-cpuflags 0 doesnt help

do you think this patch would be ready to push to main ffmpeg once
this (and any other remaining) issues are fixed
or is there still some speed loss ?
[...]

> +#define FREE_FILTERS_ON_ERROR(err, ctx) if ((err) < 0) {            \
> +                                            ff_free_filters((ctx)); \
> +                                            return (err);           \
> +                                        }

you can avoid the macro by using a goto
yes i know goto sucks but such error cleanup is one of the very few
cases where using a goto is IMO a good idea

[...]

> diff --git a/libswscale/swscale_internal.h b/libswscale/swscale_internal.h
> index 2299aa5..af4b635 100644
> --- a/libswscale/swscale_internal.h
> +++ b/libswscale/swscale_internal.h
> @@ -269,6 +269,9 @@ typedef void (*yuv2anyX_fn)(struct SwsContext *c, const int16_t *lumFilter,
>                              const int16_t **alpSrc, uint8_t **dest,
>                              int dstW, int y);
>  
> +struct SwsSlice;
> +struct SwsFilterDescriptor;
> +
>  /* This struct should be aligned on at least a 32-byte boundary. */
>  typedef struct SwsContext {
>      /**
> @@ -319,6 +322,12 @@ typedef struct SwsContext {
>      uint16_t *gamma;
>      uint16_t *inv_gamma;
>  
> +    int numDesc;
> +    int descIndex[2];
> +    int numSlice;
> +    struct SwsSlice *slice;
> +    struct SwsFilterDescriptor *desc;
> +
>      uint32_t pal_yuv[256];
>      uint32_t pal_rgb[256];
>  
> @@ -908,4 +917,75 @@ static inline void fillPlane16(uint8_t *plane, int stride, int width, int height
>      }
>  }
>  
> +#define MAX_SLICE_PLANES 4
> +
> +/// Slice plane
> +typedef struct SwsPlane
> +{
> +    int available_lines;    ///< max number of lines that can be hold by this plane
> +    int sliceY;             ///< index of first line
> +    int sliceH;             ///< number of lines
> +    uint8_t **line;         ///< line buffer
> +    uint8_t **tmp;          ///< Tmp line buffer used by mmx code
> +} SwsPlane;
> +
> +/**
> + * Struct which defines a slice of an image to be scaled or a output for
> + * a scaled slice.
> + * A slice can also be used as intermediate ring buffer for scaling steps.
> + */
> +typedef struct SwsSlice 
> +{
> +    int width;              ///< Slice line width
> +    int h_chr_sub_sample;   ///< horizontal chroma subsampling factor
> +    int v_chr_sub_sample;   ///< vertical chroma subsampling factor
> +    int is_ring;            ///< flag to identify if this slice is a ring buffer
> +    int should_free_lines;  ///< flag to identify if there are dynamic allocated lines
> +    enum AVPixelFormat fmt; ///< planes pixel format
> +    SwsPlane plane[MAX_SLICE_PLANES];   ///< color planes
> +} SwsSlice;
> +

> +/**
> + * Struct which holds all necessary data for processing a slice.
> + * A processing step can be a color conversion or horizontal/vertical scaling.
> + */
> +typedef struct SwsFilterDescriptor
> +{
> +    SwsSlice *src;  ///< Source slice
> +    SwsSlice *dst;  ///< Output slice
> +
> +    int alpha;      ///< Flag for processing alpha channel
> +    void *instance; ///< Filter instance data
> +
> +    /// Function for processing input slice sliceH lines starting from line sliceY
> +    int (*process)(SwsContext *c, struct SwsFilterDescriptor *desc, int sliceY, int sliceH);
> +} SwsFilterDescriptor;
> +
> +/// Color conversion instance data
> +typedef struct ColorContext
> +{
> +    uint32_t *pal;
> +} ColorContext;
> +
> +/// Scaler instance data
> +typedef struct FilterContext
> +{
> +    uint16_t *filter;
> +    int *filter_pos;
> +    int filter_size;
> +    int xInc;
> +} FilterContext;

This looks fine in what is in there but i think it should be
restructured a bit.

There should be a struct that contains only constant fields which
describes a type/class of filter, like a palette->yuv filter or whatever
(constant not in the sense of const but that they are the same for
 all instances)

then a generic context for filter instances is needed which would
contain all generic fields a instance needs

and then a "private" context for a filter instance which would contain
type/class specific fields

ColorContext / FilterContext would match that private
contexts already i think
though FilterContext sounds a bit too generic for a private context
in a "shared" header

SwsFilterDescriptor though mixes "class" and instance parts
the process() pointer i assume would be the same for all instances

you could also look at the AVFilter struct for a general idea of such
a "const" descriptor


> +
> +// warp input lines in the form (src + width*i + j) to slice format (line[i][j])
> +int ff_init_slice_from_src(SwsSlice * s, uint8_t *src[4], int stride[4], int srcW, int lumY, int lumH, int chrY, int chrH);
> +
> +// Initialize scaler filter descriptor chain
> +int ff_init_filters(SwsContext *c);
> +
> +// Free all filter data
> +int ff_free_filters(SwsContext *c);
> +

> +// function for applying ring buffer logic into slice s

this should be a bit more verbose, "applying ring buffer logic" is
not clear in what it does


> +int ff_rotate_slice(SwsSlice *s, int lum, int chr);


[...]

Thanks

--
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Asymptotically faster algorithms should always be preferred if you have
asymptotical amounts of data
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 181 bytes
Desc: Digital signature
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20150729/1859a286/attachment.sig>