On Mon, Dec 19, 2016 at 03:18:00PM +0100, Paul B Mahol wrote:
> Signed-off-by: Paul B Mahol <one...@gmail.com>
> ---
> doc/general.texi | 1 +
> libavcodec/Makefile | 1 +
> libavcodec/allcodecs.c | 1 +
> libavcodec/avcodec.h | 1 +
> libavcodec/codec_desc.c | 7 +
> libavcodec/pixlet.c | 726
> ++++++++++++++++++++++++++++++++++++++++++++++++
> libavformat/isom.c | 2 +
> 7 files changed, 739 insertions(+)
> create mode 100644 libavcodec/pixlet.c
[...]
> +static int read_low_coeffs(AVCodecContext *avctx, int16_t *dst, int size)
> +{
> + PixletContext *ctx = avctx->priv_data;
> + GetBitContext *b = &ctx->gbit;
> + unsigned value, cnt1, nbits, j, i = 0;
> + int rlen, flag = 0, escape;
> + int64_t rparam = 3;
> +
> + while (i < size) {
> + nbits = FFMIN(ff_clz((rparam >> 8) + 3) ^ 0x1F, 14);
> +
> + cnt1 = get_unary(b, 0, 8);
> + if (cnt1 < 8) {
> + value = show_bits(b, nbits);
> + if (value <= 1) {
> + skip_bits(b, nbits - 1);
> + escape = ((1 << nbits) - 1) * cnt1;
> + } else {
> + skip_bits(b, nbits);
> + escape = value + ((1 << nbits) - 1) * cnt1 - 1;
> + }
> + } else {
> + escape = get_bits(b, 16);
> + }
> +
> + rlen = -((escape + flag) & 1) | 1;
> + dst[i++] = rlen * ((escape + flag + 1) >> 1);
> + rparam += 120 * (escape + flag) - (120 * rparam >> 8);
> + flag = 0;
> +
> + if (rparam * 4 > 0xFF || i >= size)
> + continue;
> +
> + nbits = ((rparam + 8) >> 5) + (rparam ? ff_clz(rparam) : 32) - 24;
> + escape = 16383 & ((1 << nbits) - 1);
> + cnt1 = get_unary(b, 0, 8);
> + if (cnt1 > 7) {
> + rlen = get_bits(b, 16);
> + } else {
> + value = show_bits(b, nbits);
> + if (value > 1) {
> + skip_bits(b, nbits);
> + rlen = value + escape * cnt1 - 1;
> + } else {
> + if (nbits - 1 > 0)
> + skip_bits(b, nbits - 1);
> + rlen = escape * cnt1;
> + }
> + }
this and the previous such block look very similar
so do the ones in read_high_coeffs()
can they be factored or are there subtele differences that make that
annoying ?
[...]
> +static void reconstruction(AVCodecContext *avctx,
> + int16_t *dest, unsigned width, unsigned height,
> size_t stride, int nb_levels,
> + float *scaling_H, float *scaling_V, int lowres)
> +{
> + PixletContext *ctx = avctx->priv_data;
> + unsigned scaled_width, scaled_height;
> + float scale_H, scale_V;
> + int16_t *ptr, *tmp;
> + int i, j, k;
> +
> + scaled_height = height >> nb_levels;
> + scaled_width = width >> nb_levels;
> + tmp = ctx->filter[0];
> +
> + for (i = 0; i < nb_levels; i++) {
> + scaled_width <<= 1;
> + scaled_height <<= 1;
> + scale_H = scaling_H[i + lowres];
> + scale_V = scaling_V[i + lowres];
> +
> + ptr = dest;
> + for (j = 0; j < scaled_height; j++) {
> + filter(ptr, ctx->filter[1], scaled_width, scale_V);
> + ptr += stride;
> + }
> +
> + for (j = 0; j < scaled_width; j++) {
> + ptr = dest + j;
> + for (k = 0; k < scaled_height; k++) {
> + tmp[k] = *ptr;
> + ptr += stride;
> + }
> +
> + filter(tmp, ctx->filter[1], scaled_height, scale_H);
> +
> + ptr = dest + j;
> + for (k = 0; k < scaled_height; k++) {
> + *ptr = tmp[k];
> + ptr += stride;
> + }
the whole codec does quite a bit of copying and extra passes
(for example postprocess could be merged into prior steps)
not sure this matters or how important speed for this codec is?
[...]
--
Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
Opposition brings concord. Out of discord comes the fairest harmony.
-- Heraclitus
signature.asc
Description: Digital signature
_______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
