[FFmpeg-devel,2/2,v2] avcodec/gifenc: Only write frame palette entries that actually used

GIF palette entries are not compressed, and writing 256 entries,
which can be up to every frame, uses a significant amount of
space, especially in extreme cases, where palettes can be very
small.

Example, first six seconds of Tears of Steel, palette generated
with libimagequant, 320x240 resolution, and with transparency
optimization + per frame palette:

    * Before patch: 186765 bytes
    * After patch: 77895 bytes

Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
---
The global palette, if used, is no longer ever shrunk in v2 of this
patch. This is because just because any given frame doesn't reference
an index, doesn't mean a future frame won't.
---
 libavcodec/gif.c | 72 ++++++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 69 insertions(+), 3 deletions(-)

On Sat, Feb 20, 2021 at 7:58 PM Derek Buitenhuis <derek.buitenhuis@gmail.com>
wrote:

> GIF palette entries are not compressed, and writing 256 entries,
> which can be up to every frame, uses a significant amount of
> space, especially in extreme cases, where palettes can be very
> small.
>
> Example, first six seconds of Tears of Steel, palette generated
> with libimagequant, 320x240 resolution, and with transparency
> optimization + per frame palette:
>
>     * Before patch: 186765 bytes
>     * After patch: 77895 bytes
>
> Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
> ---
> The global palette, if used, is no longer ever shrunk in v2 of this
> patch. This is because just because any given frame doesn't reference
> an index, doesn't mean a future frame won't.
> ---
>  libavcodec/gif.c | 72 ++++++++++++++++++++++++++++++++++++++++++++++--
>  1 file changed, 69 insertions(+), 3 deletions(-)
>
> diff --git a/libavcodec/gif.c b/libavcodec/gif.c
> index 8c07ee2769..420a1f523e 100644
> --- a/libavcodec/gif.c
> +++ b/libavcodec/gif.c
> @@ -63,6 +63,44 @@ enum {
>      GF_TRANSDIFF  = 1<<1,
>  };
>
> +static void shrink_palette(const uint32_t *src, uint8_t *map,
> +                           uint32_t *dst, size_t *palette_count)
> +{
> +    size_t colors_seen = 0;
> +
> +    for (size_t i = 0; i < AVPALETTE_COUNT; i++) {
> +        int seen = 0;
> +        for (size_t c = 0; c < colors_seen; c++) {
> +            if (src[i] == dst[c]) {
> +                seen = 1;
> +                break;
> +            }
> +        }
> +        if (!seen) {
> +            dst[colors_seen] = src[i];
> +            map[i] = colors_seen;
> +            colors_seen++;
> +        }
> +    }
> +
> +    *palette_count = colors_seen;
> +}
> +
> +static void remap_frame_to_palette(const uint8_t *src, int src_linesize,
> +                                   uint8_t *dst, int dst_linesize,
> +                                   int w, int h, uint8_t *map)
> +{
> +    uint8_t *sptr = (uint8_t *) src;
> +    uint8_t *dptr = dst;
> +
> +    for (int i = 0; i < h; i++) {
> +        for (int j = 0; j < w; j++)
> +            dptr[j] = map[sptr[j]];
> +        dptr += dst_linesize;
> +        sptr += src_linesize;
> +    }
> +}
> +
>  static int is_image_translucent(AVCodecContext *avctx,
>                                  const uint8_t *buf, const int linesize)
>  {
> @@ -267,6 +305,18 @@ static int gif_image_write_image(AVCodecContext
> *avctx,
>      int x_start = 0, y_start = 0, trans = s->transparent_index;
>      int bcid = -1, honor_transparency = (s->flags & GF_TRANSDIFF) &&
> s->last_frame && !palette;
>      const uint8_t *ptr;
> +    uint32_t shrunk_palette[AVPALETTE_COUNT];
> +    uint8_t map[AVPALETTE_COUNT] = { 0 };
> +    size_t shrunk_palette_count = 0;
> +    uint8_t *shrunk_buf = NULL;
> +
> +    /*
> +     * We memset to 0xff instead of 0x00 so that the transparency
> detection
> +     * doesn't pick anything after the palette entries as the transparency
> +     * index, and because GIF89a requires us to always write a power-of-2
> +     * number of palette entries.
> +     */
> +    memset(shrunk_palette, 0xff, AVPALETTE_SIZE);
>
>      if (!s->image && is_image_translucent(avctx, buf, linesize)) {
>          gif_crop_translucent(avctx, buf, linesize, &width, &height,
> &x_start, &y_start);
> @@ -335,9 +385,14 @@ static int gif_image_write_image(AVCodecContext
> *avctx,
>
>      if (palette || !s->use_global_palette) {
>          const uint32_t *pal = palette ? palette : s->palette;
> +        unsigned pow2_count;
>          unsigned i;
> -        bytestream_put_byte(bytestream, 1<<7 | 0x7); /* flags */
> -        for (i = 0; i < AVPALETTE_COUNT; i++) {
> +
> +        shrink_palette(pal, map, shrunk_palette, &shrunk_palette_count);
> +        pow2_count = av_log2(shrunk_palette_count - 1);
> +
> +        bytestream_put_byte(bytestream, 1<<7 | pow2_count); /* flags */
> +        for (i = 0; i < 1 << (pow2_count + 1); i++) {
>              const uint32_t v = pal[i];
>              bytestream_put_be24(bytestream, v);
>          }
> @@ -350,7 +405,17 @@ static int gif_image_write_image(AVCodecContext
> *avctx,
>      ff_lzw_encode_init(s->lzw, s->buf, s->buf_size,
>                         12, FF_LZW_GIF, 1);
>
> -    ptr = buf + y_start*linesize + x_start;
> +    if (shrunk_palette_count) {
> +        shrunk_buf = av_malloc(avctx->height * linesize);
>

I would prefer it this allocation is done only once per encoder instance.


> +        if (!shrunk_buf) {
> +            av_log(avctx, AV_LOG_ERROR, "Could not allocated remapped
> frame buffer.\n");
> +            return AVERROR(ENOMEM);
> +        }
> +        remap_frame_to_palette(buf, linesize, shrunk_buf, linesize,
> avctx->width, avctx->height, map);
> +        ptr = shrunk_buf + y_start*linesize + x_start;
> +    } else {
> +        ptr = buf + y_start*linesize + x_start;
> +    }
>      if (honor_transparency) {
>          const int ref_linesize = s->last_frame->linesize[0];
>          const uint8_t *ref = s->last_frame->data[0] +
> y_start*ref_linesize + x_start;
> @@ -383,6 +448,7 @@ static int gif_image_write_image(AVCodecContext *avctx,
>          len -= size;
>      }
>      bytestream_put_byte(bytestream, 0x00); /* end of image block */
> +    av_freep(&shrunk_buf);
>      return 0;
>  }
>
> --
> 2.30.0
>
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".

Derek Buitenhuis:
> GIF palette entries are not compressed, and writing 256 entries,
> which can be up to every frame, uses a significant amount of
> space, especially in extreme cases, where palettes can be very
> small.
> 
> Example, first six seconds of Tears of Steel, palette generated
> with libimagequant, 320x240 resolution, and with transparency
> optimization + per frame palette:
> 
>     * Before patch: 186765 bytes
>     * After patch: 77895 bytes
> 
> Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
> ---
> The global palette, if used, is no longer ever shrunk in v2 of this
> patch. This is because just because any given frame doesn't reference
> an index, doesn't mean a future frame won't.
> ---
>  libavcodec/gif.c | 72 ++++++++++++++++++++++++++++++++++++++++++++++--
>  1 file changed, 69 insertions(+), 3 deletions(-)
> 
> diff --git a/libavcodec/gif.c b/libavcodec/gif.c
> index 8c07ee2769..420a1f523e 100644
> --- a/libavcodec/gif.c
> +++ b/libavcodec/gif.c
> @@ -63,6 +63,44 @@ enum {
>      GF_TRANSDIFF  = 1<<1,
>  };
>  
> +static void shrink_palette(const uint32_t *src, uint8_t *map,
> +                           uint32_t *dst, size_t *palette_count)
> +{
> +    size_t colors_seen = 0;
> +
> +    for (size_t i = 0; i < AVPALETTE_COUNT; i++) {
> +        int seen = 0;
> +        for (size_t c = 0; c < colors_seen; c++) {
> +            if (src[i] == dst[c]) {
> +                seen = 1;
> +                break;
> +            }
> +        }
> +        if (!seen) {
> +            dst[colors_seen] = src[i];
> +            map[i] = colors_seen;
> +            colors_seen++;
> +        }
> +    }
> +
> +    *palette_count = colors_seen;
> +}
> +
> +static void remap_frame_to_palette(const uint8_t *src, int src_linesize,
> +                                   uint8_t *dst, int dst_linesize,
> +                                   int w, int h, uint8_t *map)
> +{
> +    uint8_t *sptr = (uint8_t *) src;

Casting const away here is completely unnecessary as sptr is not used to
modify what it points to; and actually sptr and dptr are completely
unnecessary, too.

> +    uint8_t *dptr = dst;
> +
> +    for (int i = 0; i < h; i++) {
> +        for (int j = 0; j < w; j++)
> +            dptr[j] = map[sptr[j]];
> +        dptr += dst_linesize;
> +        sptr += src_linesize;
> +    }
> +}
> +
>  static int is_image_translucent(AVCodecContext *avctx,
>                                  const uint8_t *buf, const int linesize)
>  {
> @@ -267,6 +305,18 @@ static int gif_image_write_image(AVCodecContext *avctx,
>      int x_start = 0, y_start = 0, trans = s->transparent_index;
>      int bcid = -1, honor_transparency = (s->flags & GF_TRANSDIFF) && s->last_frame && !palette;
>      const uint8_t *ptr;
> +    uint32_t shrunk_palette[AVPALETTE_COUNT];
> +    uint8_t map[AVPALETTE_COUNT] = { 0 };
> +    size_t shrunk_palette_count = 0;
> +    uint8_t *shrunk_buf = NULL;
> +
> +    /*
> +     * We memset to 0xff instead of 0x00 so that the transparency detection
> +     * doesn't pick anything after the palette entries as the transparency
> +     * index, and because GIF89a requires us to always write a power-of-2
> +     * number of palette entries.
> +     */
> +    memset(shrunk_palette, 0xff, AVPALETTE_SIZE);
>  
>      if (!s->image && is_image_translucent(avctx, buf, linesize)) {
>          gif_crop_translucent(avctx, buf, linesize, &width, &height, &x_start, &y_start);
> @@ -335,9 +385,14 @@ static int gif_image_write_image(AVCodecContext *avctx,
>  
>      if (palette || !s->use_global_palette) {
>          const uint32_t *pal = palette ? palette : s->palette;
> +        unsigned pow2_count;
>          unsigned i;
> -        bytestream_put_byte(bytestream, 1<<7 | 0x7); /* flags */
> -        for (i = 0; i < AVPALETTE_COUNT; i++) {
> +
> +        shrink_palette(pal, map, shrunk_palette, &shrunk_palette_count);
> +        pow2_count = av_log2(shrunk_palette_count - 1);
> +
> +        bytestream_put_byte(bytestream, 1<<7 | pow2_count); /* flags */
> +        for (i = 0; i < 1 << (pow2_count + 1); i++) {
>              const uint32_t v = pal[i];
>              bytestream_put_be24(bytestream, v);
>          }
> @@ -350,7 +405,17 @@ static int gif_image_write_image(AVCodecContext *avctx,
>      ff_lzw_encode_init(s->lzw, s->buf, s->buf_size,
>                         12, FF_LZW_GIF, 1);
>  
> -    ptr = buf + y_start*linesize + x_start;
> +    if (shrunk_palette_count) {
> +        shrunk_buf = av_malloc(avctx->height * linesize);
> +        if (!shrunk_buf) {
> +            av_log(avctx, AV_LOG_ERROR, "Could not allocated remapped frame buffer.\n");
> +            return AVERROR(ENOMEM);
> +        }
> +        remap_frame_to_palette(buf, linesize, shrunk_buf, linesize, avctx->width, avctx->height, map);
> +        ptr = shrunk_buf + y_start*linesize + x_start;
> +    } else {
> +        ptr = buf + y_start*linesize + x_start;
> +    }
>      if (honor_transparency) {
>          const int ref_linesize = s->last_frame->linesize[0];
>          const uint8_t *ref = s->last_frame->data[0] + y_start*ref_linesize + x_start;
> @@ -383,6 +448,7 @@ static int gif_image_write_image(AVCodecContext *avctx,
>          len -= size;
>      }
>      bytestream_put_byte(bytestream, 0x00); /* end of image block */
> +    av_freep(&shrunk_buf);
>      return 0;
>  }
>  
>