[FFmpeg-devel] [PATCH] lavfi: Port fspp to FFmpeg
arwa arif
arwaarif1994 at gmail.com
Tue Dec 23 11:22:57 CET 2014
On Tue, Dec 23, 2014 at 12:40 PM, Stefano Sabatini <stefasab at gmail.com>
wrote:
> On date Monday 2014-12-22 21:51:25 +0530, arwa arif encoded:
> > On Mon, Dec 22, 2014 at 5:37 PM, Stefano Sabatini <stefasab at gmail.com>
> > wrote:
> >
> > > On date Friday 2014-12-19 16:46:40 +0530, arwa arif encoded:
> > > [...]
> > > > From c1c3255203226663fc382a0994182df3d558afe6 Mon Sep 17 00:00:00
> 2001
> > > > From: Arwa Arif <arwaarif1994 at gmail.com>
> > > > Date: Sun, 14 Dec 2014 12:03:31 +0530
> > > > Subject: [PATCH] lavfi: port mp=fspp to a native libavfilter filter
> > > >
> > > > ---
> > > > LICENSE.md | 1 +
> > > > configure | 1 +
> > > > doc/filters.texi | 30 +
> > > > libavfilter/Makefile | 1 +
> > > > libavfilter/allfilters.c | 1 +
> > > > libavfilter/libmpcodecs/vf_fspp.c | 4 +-
> > > > libavfilter/version.h | 4 +-
> > > > libavfilter/vf_fspp.c | 669 ++++++++++++++++++
> > > > libavfilter/vf_fspp.h | 96 +++
> > > > libavfilter/x86/Makefile | 1 +
> > > > libavfilter/x86/vf_fspp.c | 1405
> > > +++++++++++++++++++++++++++++++++++++
> > > > 11 files changed, 2209 insertions(+), 4 deletions(-)
> > > > create mode 100644 libavfilter/vf_fspp.c
> > > > create mode 100644 libavfilter/vf_fspp.h
> > > > create mode 100644 libavfilter/x86/vf_fspp.c
> > > >
> > > > diff --git a/LICENSE.md b/LICENSE.md
> > > > index cf9955f..188d060 100644
> > > > --- a/LICENSE.md
> > > > +++ b/LICENSE.md
> > > > @@ -31,6 +31,7 @@ Specifically, the GPL parts of FFmpeg are:
> > > > - vf_cropdetect.c
> > > > - vf_decimate.c
> > > > - vf_delogo.c
> > > > + - vf_fspp.c
> > > > - vf_geq.c
> > > > - vf_histeq.c
> > > > - vf_hqdn3d.c
> > > > diff --git a/configure b/configure
> > > > index e37285a..29f5534 100755
> > > > --- a/configure
> > > > +++ b/configure
> > > > @@ -2575,6 +2575,7 @@ ebur128_filter_deps="gpl"
> > > > flite_filter_deps="libflite"
> > > > frei0r_filter_deps="frei0r dlopen"
> > > > frei0r_src_filter_deps="frei0r dlopen"
> > > > +fspp_filter_deps="gpl"
> > > > geq_filter_deps="gpl"
> > > > histeq_filter_deps="gpl"
> > > > hqdn3d_filter_deps="gpl"
> > > > diff --git a/doc/filters.texi b/doc/filters.texi
> > > > index 882caa0..2c84ad0 100644
> > > > --- a/doc/filters.texi
> > > > +++ b/doc/filters.texi
> > > > @@ -4997,6 +4997,35 @@ frei0r=perspective:0.2/0.2|0.8/0.2
> > > > For more information, see
> > > > @url{http://frei0r.dyne.org}
> > > >
> > > > + at section fspp
> > > > +
> > > > +Apply fast and simple postprocessing. It is a faster version of the
> > > simple
> > > > +postprocessing filter - @ref{spp}.
> > > > +
> > > > +It splits (I)DCT into horizontal/vertical passes. Unlike the simple
> > > post-
> > > > +processing filter, one of them is performed once per block, not per
> > > pixel.
> > > > +This allows for much better speed.
> > >
> > > This allows for higher speed?
> > >
> > >
> > Yes, it does.
>
> I meant that "higher speed" reads more semantically correct than
> "better speed".
>
> [...]
> > I have added one more option of filter strength. I compared the ouput of
> > this filter with mp=fspp, for different values of qp, and the ouput is
> not
> > same. I have been looking into the code, but I am not able to figure out
> > the mistake.
>
> > From d4be07bb81f3d7ef03f65eaf3403da0bbe0b5391 Mon Sep 17 00:00:00 2001
> > From: Arwa Arif <arwaarif1994 at gmail.com>
> > Date: Sun, 14 Dec 2014 12:03:31 +0530
> > Subject: [PATCH] lavfi: port mp=fspp to a native libavfilter filter
> >
> > ---
> > LICENSE.md | 1 +
> > configure | 1 +
> > doc/filters.texi | 33 +
> > libavfilter/Makefile | 1 +
> > libavfilter/allfilters.c | 1 +
> > libavfilter/libmpcodecs/vf_fspp.c | 4 +-
> > libavfilter/version.h | 4 +-
> > libavfilter/vf_fspp.c | 699 ++++++++++++++++++
> > libavfilter/vf_fspp.h | 96 +++
> > libavfilter/x86/Makefile | 1 +
> > libavfilter/x86/vf_fspp.c | 1405
> +++++++++++++++++++++++++++++++++++++
> > 11 files changed, 2242 insertions(+), 4 deletions(-)
> > create mode 100644 libavfilter/vf_fspp.c
> > create mode 100644 libavfilter/vf_fspp.h
> > create mode 100644 libavfilter/x86/vf_fspp.c
> >
> > diff --git a/LICENSE.md b/LICENSE.md
> > index cf9955f..188d060 100644
> > --- a/LICENSE.md
> > +++ b/LICENSE.md
> > @@ -31,6 +31,7 @@ Specifically, the GPL parts of FFmpeg are:
> > - vf_cropdetect.c
> > - vf_decimate.c
> > - vf_delogo.c
> > + - vf_fspp.c
> > - vf_geq.c
> > - vf_histeq.c
> > - vf_hqdn3d.c
> > diff --git a/configure b/configure
> > index e37285a..29f5534 100755
> > --- a/configure
> > +++ b/configure
> > @@ -2575,6 +2575,7 @@ ebur128_filter_deps="gpl"
> > flite_filter_deps="libflite"
> > frei0r_filter_deps="frei0r dlopen"
> > frei0r_src_filter_deps="frei0r dlopen"
> > +fspp_filter_deps="gpl"
> > geq_filter_deps="gpl"
> > histeq_filter_deps="gpl"
> > hqdn3d_filter_deps="gpl"
> > diff --git a/doc/filters.texi b/doc/filters.texi
> > index 882caa0..d23575e 100644
> > --- a/doc/filters.texi
> > +++ b/doc/filters.texi
> > @@ -4997,6 +4997,38 @@ frei0r=perspective:0.2/0.2|0.8/0.2
> > For more information, see
> > @url{http://frei0r.dyne.org}
> >
> > + at section fspp
> > +
> > +Apply fast and simple postprocessing. It is a faster version of the
> simple
> > +postprocessing filter - @ref{spp}.
>
> more direct:
> It is a faster version of the @ref{spp} filter.
>
> > +
> > +It splits (I)DCT into horizontal/vertical passes. Unlike the simple
> post-
> > +processing filter, one of them is performed once per block, not per
> pixel.
>
> > +This allows for much better speed.
>
> higher speed
>
> here and below
>
> > +
> > +The filter accepts the following options:
> > +
> > + at table @option
> > + at item quality
> > +Set quality. This option defines the number of levels for averaging. It
> accepts
> > +an integer in the range 4-5. Default value is @code{4}.
> > +
> > + at item qp
> > +Force a constant quantization parameter. It accepts an integer in range
> 0-63.
> > +If not set, the filter will use the QP from the video stream (if
> available).
> > +
> > + at item strength
> > +Set filter strength. It accepts an integer in range -15 to 32. Lower
> values mean
> > +more details but also more artifacts, while higher values make the
> image smoother
> > +but also blurrier. Default value is @code{0} − PSNR optimal.
> > +
> > + at item use_bframe_qp
> > +Enable the use of the QP from the B-Frames if set to @code{1}. Using
> this
> > +option may cause flicker since the B-Frames have often larger QP.
> Default is
> > + at code{0} (not enabled).
> > +
> > + at end table
> > +
> > @section geq
> >
> > The filter accepts the following options:
> > @@ -8292,6 +8324,7 @@ stereo3d=abl:sbsr
> > @end example
> > @end itemize
> >
> > + at anchor{spp}
> > @section spp
> >
> > Apply a simple postprocessing filter that compresses and decompresses
> the image
> > diff --git a/libavfilter/Makefile b/libavfilter/Makefile
> > index 6b7291e..8c523b4 100644
> > --- a/libavfilter/Makefile
> > +++ b/libavfilter/Makefile
> > @@ -125,6 +125,7 @@ OBJS-$(CONFIG_FRAMESTEP_FILTER) +=
> vf_framestep.o
> > OBJS-$(CONFIG_FPS_FILTER) += vf_fps.o
> > OBJS-$(CONFIG_FRAMEPACK_FILTER) += vf_framepack.o
> > OBJS-$(CONFIG_FREI0R_FILTER) += vf_frei0r.o
> > +OBJS-$(CONFIG_FSPP_FILTER) += vf_fspp.o
> > OBJS-$(CONFIG_GEQ_FILTER) += vf_geq.o
> > OBJS-$(CONFIG_GRADFUN_FILTER) += vf_gradfun.o
> > OBJS-$(CONFIG_HALDCLUT_FILTER) += vf_lut3d.o dualinput.o
> framesync.o
> > diff --git a/libavfilter/allfilters.c b/libavfilter/allfilters.c
> > index adb86be..4a915c7 100644
> > --- a/libavfilter/allfilters.c
> > +++ b/libavfilter/allfilters.c
> > @@ -141,6 +141,7 @@ void avfilter_register_all(void)
> > REGISTER_FILTER(FRAMEPACK, framepack, vf);
> > REGISTER_FILTER(FRAMESTEP, framestep, vf);
> > REGISTER_FILTER(FREI0R, frei0r, vf);
> > + REGISTER_FILTER(FSPP, fspp, vf);
> > REGISTER_FILTER(GEQ, geq, vf);
> > REGISTER_FILTER(GRADFUN, gradfun, vf);
> > REGISTER_FILTER(HALDCLUT, haldclut, vf);
> > diff --git a/libavfilter/libmpcodecs/vf_fspp.c
> b/libavfilter/libmpcodecs/vf_fspp.c
> > index d457859..3a80dc2 100644
> > --- a/libavfilter/libmpcodecs/vf_fspp.c
> > +++ b/libavfilter/libmpcodecs/vf_fspp.c
> > @@ -710,8 +710,8 @@ const vf_info_t ff_vf_info_fspp = {
> > #if HAVE_MMX_INLINE
> >
> > DECLARE_ASM_CONST(8, uint64_t, MM_FIX_0_382683433)=FIX64(0.382683433,
> 14);
> > -DECLARE_ALIGNED(8, uint64_t, ff_MM_FIX_0_541196100)=FIX64(0.541196100,
> 14);
> > -DECLARE_ALIGNED(8, uint64_t, ff_MM_FIX_0_707106781)=FIX64(0.707106781,
> 14);
> > +extern uint64_t ff_MM_FIX_0_707106781;
> > +extern uint64_t ff_MM_FIX_0_541196100;
> > DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_306562965)=FIX64(1.306562965,
> 14);
> >
> > DECLARE_ASM_CONST(8, uint64_t, MM_FIX_1_414213562_A)=FIX64(1.414213562,
> 14);
> > diff --git a/libavfilter/version.h b/libavfilter/version.h
> > index 4bd18f3..50fc932 100644
> > --- a/libavfilter/version.h
> > +++ b/libavfilter/version.h
> > @@ -30,8 +30,8 @@
> > #include "libavutil/version.h"
> >
>
> > #define LIBAVFILTER_VERSION_MAJOR 5
> > -#define LIBAVFILTER_VERSION_MINOR 2
> > -#define LIBAVFILTER_VERSION_MICRO 104
> > +#define LIBAVFILTER_VERSION_MINOR 3
> > +#define LIBAVFILTER_VERSION_MICRO 100
>
> Please always rebase against current master, this causes a conflict
> when applying the patch (or alternatively skip this hunk and add a
> TODO: update minor in the commit message).
>
> > #define LIBAVFILTER_VERSION_INT
> AV_VERSION_INT(LIBAVFILTER_VERSION_MAJOR, \
> >
> LIBAVFILTER_VERSION_MINOR, \
> > diff --git a/libavfilter/vf_fspp.c b/libavfilter/vf_fspp.c
> > new file mode 100644
> > index 0000000..269d46f
> > --- /dev/null
> > +++ b/libavfilter/vf_fspp.c
> > @@ -0,0 +1,699 @@
> > +/*
> > + * Copyright (c) 2003 Michael Niedermayer <michaelni at gmx.at>
> > + * Copyright (C) 2005 Nikolaj Poroshin <porosh3 at psu.ru>
> > + * Copyright (c) 2014 Arwa Arif <arwaarif1994 at gmail.com>
> > + *
> > + * This file is part of FFmpeg.
> > + *
> > + * FFmpeg is free software; you can redistribute it and/or modify
> > + * it under the terms of the GNU General Public License as published by
> > + * the Free Software Foundation; either version 2 of the License, or
> > + * (at your option) any later version.
> > + *
> > + * FFmpeg is distributed in the hope that it will be useful,
> > + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
> > + * GNU General Public License for more details.
> > + *
> > + * You should have received a copy of the GNU General Public License
> along
> > + * with FFmpeg; if not, write to the Free Software Foundation, Inc.,
> > + * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
> > + */
> > +
> > +/**
> > + * @file
> > + * Fast Simple Post-processing filter
> > + * This implementation is based on an algorithm described in
> > + * "Aria Nosratinia Embedded Post-Processing for
> > + * Enhancement of Compressed Images (1999)"
> > + * (http://www.utdallas.edu/~aria/papers/vlsisp99.pdf)
> > + * Further, with splitting (I)DCT into horizontal/vertical passes, one
> of
> > + * them can be performed once per block, not per pixel. This allows for
> much
> > + * better speed.
> > + *
> > + * Originally written by Michael Niedermayer and Nikolaj for the MPlayer
> > + * project, and ported by Arwa Arif for FFmpeg.
> > + */
> > +
> > +#include "libavutil/avassert.h"
> > +#include "libavutil/imgutils.h"
> > +#include "libavutil/opt.h"
> > +#include "libavutil/pixdesc.h"
> > +#include "internal.h"
> > +#include "libavcodec/avcodec.h" //for reference to FF_QSCALE_TYPE
> > +#include "vf_fspp.h"
> > +
> > +#define OFFSET(x) offsetof(FSPPContext, x)
> > +#define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
> > +static const AVOption fspp_options[] = {
> > + { "quality", "set quality",
> OFFSET(log2_count), AV_OPT_TYPE_INT, {.i64 = 4}, 4, MAX_LEVEL, FLAGS },
> > + { "qp", "force a constant quantizer parameter",
> OFFSET(qp), AV_OPT_TYPE_INT, {.i64 = 0}, 0, 64, FLAGS },
> > + { "strength", "set filter strength",
> OFFSET(strength), AV_OPT_TYPE_INT, {.i64 = 0}, -15, 32, FLAGS },
> > + { "use_bframe_qp", "use B-frames' QP",
> OFFSET(use_bframe_qp), AV_OPT_TYPE_INT, {.i64 = 0}, 0, 1, FLAGS
> },
> > + { NULL }
> > +};
> > +
> > +AVFILTER_DEFINE_CLASS(fspp);
> > +
> > +DECLARE_ALIGNED(32, static const uint8_t, dither)[8][8] = {
> > + { 0, 48, 12, 60, 3, 51, 15, 63, },
> > + { 32, 16, 44, 28, 35, 19, 47, 31, },
> > + { 8, 56, 4, 52, 11, 59, 7, 55, },
> > + { 40, 24, 36, 20, 43, 27, 39, 23, },
> > + { 2, 50, 14, 62, 1, 49, 13, 61, },
> > + { 34, 18, 46, 30, 33, 17, 45, 29, },
> > + { 10, 58, 6, 54, 9, 57, 5, 53, },
> > + { 42, 26, 38, 22, 41, 25, 37, 21, },
> > +};
> > +
> > +static const short custom_threshold[64] = {
> > +// values (296) can't be too high
> > +// -it causes too big quant dependence
> > +// or maybe overflow(check), which results in some flashing
> > + 71, 296, 295, 237, 71, 40, 38, 19,
> > + 245, 193, 185, 121, 102, 73, 53, 27,
> > + 158, 129, 141, 107, 97, 73, 50, 26,
> > + 102, 116, 109, 98, 82, 66, 45, 23,
> > + 71, 94, 95, 81, 70, 56, 38, 20,
> > + 56, 77, 74, 66, 56, 44, 30, 15,
> > + 38, 53, 50, 45, 38, 30, 21, 11,
> > + 20, 27, 26, 23, 20, 15, 11, 5
> > +};
> > +
> > +static inline int norm_qscale(int qscale, int type)
> > +{
> > + switch (type) {
> > + case FF_QSCALE_TYPE_MPEG1: return qscale;
> > + case FF_QSCALE_TYPE_MPEG2: return qscale >> 1;
> > + case FF_QSCALE_TYPE_H264: return qscale >> 2;
> > + case FF_QSCALE_TYPE_VP56: return (63 - qscale + 2) >> 2;
> > + }
> > + return qscale;
> > +}
> > +
> > +//This func reads from 1 slice, 1 and clears 0 & 1
> > +static void store_slice_c(uint8_t *dst, int16_t *src,
> > + ptrdiff_t dst_stride, ptrdiff_t src_stride,
> > + ptrdiff_t width, ptrdiff_t height, ptrdiff_t
> log2_scale)
> > +{
> > + int y, x;
> > +#define STORE(pos)
> \
> > + temp = (src[x + pos] + (d[pos] >> log2_scale)) >> (6 -
> log2_scale); \
> > + src[x + pos] = src[x + pos - 8 * src_stride] = 0;
> \
> > + if (temp & 0x100) temp = ~(temp >> 31);
> \
> > + dst[x + pos] = temp;
> > +
> > + for (y = 0; y < height; y++) {
> > + const uint8_t *d = dither[y];
> > + for (x = 0; x < width; x += 8) {
> > + int temp;
> > + STORE(0);
> > + STORE(1);
> > + STORE(2);
> > + STORE(3);
> > + STORE(4);
> > + STORE(5);
> > + STORE(6);
> > + STORE(7);
> > + }
> > + src += src_stride;
> > + dst += dst_stride;
> > + }
> > +}
> > +
> > +//This func reads from 2 slices, 0 & 2 and clears 2-nd
> > +static void store_slice2_c(uint8_t *dst, int16_t *src,
> > + ptrdiff_t dst_stride, ptrdiff_t src_stride,
> > + ptrdiff_t width, ptrdiff_t height, ptrdiff_t
> log2_scale)
> > +{
> > + int y, x;
> > +#define STORE2(pos)
> \
> > + temp = (src[x + pos] + src[x + pos + 16 * src_stride] + (d[pos] >>
> log2_scale)) >> (6 - log2_scale); \
> > + src[x + pos + 16 * src_stride] = 0;
> \
> > + if (temp & 0x100) temp = ~(temp >> 31);
> \
> > + dst[x + pos] = temp;
> > +
> > + for (y = 0; y < height; y++) {
> > + const uint8_t *d = dither[y];
> > + for (x = 0; x < width; x += 8) {
> > + int temp;
> > + STORE2(0);
> > + STORE2(1);
> > + STORE2(2);
> > + STORE2(3);
> > + STORE2(4);
> > + STORE2(5);
> > + STORE2(6);
> > + STORE2(7);
> > + }
> > + src += src_stride;
> > + dst += dst_stride;
> > + }
> > +}
> > +
> > +static void mul_thrmat_c(FSPPContext *p, int q)
> > +{
> > + int a;
> > + for (a = 0; a < 64; a++)
> > + ((int8_t*)p->threshold_mtx)[a] = q *
> ((int8_t*)p->threshold_mtx_noq)[a];//ints faster in C
>
> nit: (int8_t*) -> (int8_t *)
>
> > +}
> > +
> > +static void filter(FSPPContext *p, uint8_t *dst, uint8_t *src,
> > + int dst_stride, int src_stride,
> > + int width, int height,
> > + uint8_t *qp_store, int qp_stride, int is_luma)
> > +{
> > + int x, x0, y, es, qy, t;
> > +
> > + const int stride = is_luma ? p->temp_stride : (width + 16);
> > + const int step = 6 - p->log2_count;
> > + const int qpsh = 4 - p->hsub * !is_luma;
> > + const int qpsv = 4 - p->vsub * !is_luma;
> > +
> > + DECLARE_ALIGNED(32, int32_t, block_align)[4 * 8 * BLOCKSZ + 4 * 8 *
> BLOCKSZ];
> > + int16_t *block = (int16_t *)block_align;
> > + int16_t *block3 = (int16_t *)(block_align + 4 * 8 * BLOCKSZ);
> > +
> > + memset(block3, 0, 4 * 8 * BLOCKSZ);
> > +
> > + if (!src || !dst) return;
> > +
> > + for (y = 0; y < height; y++) {
> > + int index = 8 + 8 * stride + y * stride;
> > + memcpy(p->src + index, src + y * src_stride, width);
> > + for (x = 0; x < 8; x++) {
> > + p->src[index - x - 1] = p->src[index + x
> ];
> > + p->src[index + width + x ] = p->src[index + width - x -
> 1];
> > + }
> > + }
> > +
> > + for (y = 0; y < 8; y++) {
> > + memcpy(p->src + ( 7 - y ) * stride, p->src + ( y +
> 8 ) * stride, stride);
> > + memcpy(p->src + (height + 8 + y) * stride, p->src + (height - y
> + 7) * stride, stride);
> > + }
> > + //FIXME (try edge emu)
> > +
> > + for (y = 8; y < 24; y++)
> > + memset(p->temp + 8 + y * stride, 0, width * sizeof(int16_t));
> > +
> > + for (y = step; y < height + 8; y += step) { //step= 1,2
> > + const int y1 = y - 8 + step; //l5-7 l4-6;
> > + qy = y - 4;
> > +
> > + if (qy > height - 1) qy = height - 1;
> > + if (qy < 0) qy = 0;
> > +
> > + qy = (qy >> qpsv) * qp_stride;
> > + p->row_fdct(block, p->src + y * stride + 2 - (y&1), stride, 2);
> > +
> > + for (x0 = 0; x0 < width + 8 - 8 * (BLOCKSZ - 1); x0 += 8 *
> (BLOCKSZ - 1)) {
> > + p->row_fdct(block + 8 * 8, p->src + y * stride + 8 + x0 + 2
> - (y&1), stride, 2 * (BLOCKSZ - 1));
> > +
> > + if (p->qp)
> > + p->column_fidct((int16_t*)(&p->threshold_mtx[0]), block
> + 0 * 8, block3 + 0 * 8, 8 * (BLOCKSZ - 1)); //yes, this is a HOTSPOT
> > + else
> > + for (x = 0; x < 8 * (BLOCKSZ - 1); x += 8) {
> > + t = x + x0 - 2; //correct
> t=x+x0-2-(y&1), but its the same
> > +
> > + if (t < 0) t = 0; //t always <
> width-2
> > +
> > + t = qp_store[qy + (t >> qpsh)];
> > + t = norm_qscale(t, p->qscale_type);
> > +
> > + if (t != p->prev_q) p->prev_q = t, p->mul_thrmat(p,
> t);
> > + p->column_fidct((int16_t*)(&p->threshold_mtx[0]),
> block + x * 8, block3 + x * 8, 8); //yes, this is a HOTSPOT
> > + }
> > + p->row_idct(block3 + 0 * 8, p->temp + (y & 15) * stride +
> x0 + 2 - (y & 1), stride, 2 * (BLOCKSZ - 1));
> > + memmove(block, block + (BLOCKSZ - 1) * 64, 8 * 8 *
> sizeof(int16_t)); //cycling
> > + memmove(block3, block3 + (BLOCKSZ - 1) * 64, 6 * 8 *
> sizeof(int16_t));
> > + }
> > +
> > + es = width + 8 - x0; // 8, ...
> > + if (es > 8)
> > + p->row_fdct(block + 8 * 8, p->src + y * stride + 8 + x0 + 2
> - (y & 1), stride, (es - 4) >> 2);
> > +
> > + p->column_fidct((int16_t*)(&p->threshold_mtx[0]), block,
> block3, es&(~1));
> > + p->row_idct(block3 + 0 * 8, p->temp + (y & 15) * stride + x0 +
> 2 - (y & 1), stride, es >> 2);
> > +
> > + if (!(y1 & 7) && y1) {
> > + if (y1 & 8)
> > + p->store_slice(dst + (y1 - 8) * dst_stride, p->temp + 8
> + 8 * stride,
> > + dst_stride, stride, width, 8, 5 -
> p->log2_count);
> > + else
> > + p->store_slice2(dst + (y1 - 8) * dst_stride, p->temp +
> 8 + 0 * stride,
> > + dst_stride, stride, width, 8, 5 -
> p->log2_count);
> > + }
> > + }
> > +
> > + if (y & 7) { // == height & 7
> > + if (y & 8)
> > + p->store_slice(dst + ((y - 8) & ~7) * dst_stride, p->temp +
> 8 + 8 * stride,
> > + dst_stride, stride, width, y&7, 5 -
> p->log2_count);
> > + else
> > + p->store_slice2(dst + ((y - 8) & ~7) * dst_stride, p->temp
> + 8 + 0 * stride,
> > + dst_stride, stride, width, y&7, 5 -
> p->log2_count);
> > + }
> > +}
> > +
> > +static void column_fidct_c(int16_t* thr_adr, int16_t *data, int16_t
> *output, int cnt)
>
> consistency nit: TYPE* VAR => TYPE *VAR
>
> here and below
>
> > +{
> > + int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
> > + int_simd16_t tmp10, tmp11, tmp12, tmp13;
> > + int_simd16_t z1,z2,z3,z4,z5, z10, z11, z12, z13;
> > + int_simd16_t d0, d1, d2, d3, d4, d5, d6, d7;
> > +
> > + int16_t* dataptr;
> > + int16_t* wsptr;
> > + int16_t* threshold;
> > + int ctr;
> > +
> > + dataptr = data;
> > + wsptr = output;
> > +
> > + for (; cnt > 0; cnt -= 2) { //start positions
> > + threshold = (int16_t*)thr_adr;//threshold_mtx
> > + for (ctr = DCTSIZE; ctr > 0; ctr--) {
> > + // Process columns from input, add to output.
> > + tmp0 = dataptr[DCTSIZE * 0] + dataptr[DCTSIZE * 7];
> > + tmp7 = dataptr[DCTSIZE * 0] - dataptr[DCTSIZE * 7];
> > +
> > + tmp1 = dataptr[DCTSIZE * 1] + dataptr[DCTSIZE * 6];
> > + tmp6 = dataptr[DCTSIZE * 1] - dataptr[DCTSIZE * 6];
> > +
> > + tmp2 = dataptr[DCTSIZE * 2] + dataptr[DCTSIZE * 5];
> > + tmp5 = dataptr[DCTSIZE * 2] - dataptr[DCTSIZE * 5];
> > +
> > + tmp3 = dataptr[DCTSIZE * 3] + dataptr[DCTSIZE * 4];
> > + tmp4 = dataptr[DCTSIZE * 3] - dataptr[DCTSIZE * 4];
> > +
> > + // Even part of FDCT
> > +
> > + tmp10 = tmp0 + tmp3;
> > + tmp13 = tmp0 - tmp3;
> > + tmp11 = tmp1 + tmp2;
> > + tmp12 = tmp1 - tmp2;
> > +
> > + d0 = tmp10 + tmp11;
> > + d4 = tmp10 - tmp11;
> > +
> > + z1 = MULTIPLY16H((tmp12 + tmp13) << 2, FIX_0_707106781);
> > + d2 = tmp13 + z1;
> > + d6 = tmp13 - z1;
> > +
> > + // Even part of IDCT
> > +
> > + THRESHOLD(tmp0, d0, threshold[0 * 8]);
> > + THRESHOLD(tmp1, d2, threshold[2 * 8]);
> > + THRESHOLD(tmp2, d4, threshold[4 * 8]);
> > + THRESHOLD(tmp3, d6, threshold[6 * 8]);
> > + tmp0 += 2;
> > + tmp10 = (tmp0 + tmp2) >> 2;
> > + tmp11 = (tmp0 - tmp2) >> 2;
> > +
> > + tmp13 = (tmp1 + tmp3) >>2; //+2 ! (psnr decides)
> > + tmp12 = MULTIPLY16H((tmp1 - tmp3), FIX_1_414213562_A) -
> tmp13; //<<2
> > +
> > + tmp0 = tmp10 + tmp13; //->temps
> > + tmp3 = tmp10 - tmp13; //->temps
> > + tmp1 = tmp11 + tmp12; //->temps
> > + tmp2 = tmp11 - tmp12; //->temps
> > +
> > + // Odd part of FDCT
> > +
> > + tmp10 = tmp4 + tmp5;
> > + tmp11 = tmp5 + tmp6;
> > + tmp12 = tmp6 + tmp7;
> > +
> > + z5 = MULTIPLY16H((tmp10 - tmp12) << 2, FIX_0_382683433);
> > + z2 = MULTIPLY16H(tmp10 << 2, FIX_0_541196100) + z5;
> > + z4 = MULTIPLY16H(tmp12 << 2, FIX_1_306562965) + z5;
> > + z3 = MULTIPLY16H(tmp11 << 2, FIX_0_707106781);
> > +
> > + z11 = tmp7 + z3;
> > + z13 = tmp7 - z3;
> > +
> > + d5 = z13 + z2;
> > + d3 = z13 - z2;
> > + d1 = z11 + z4;
> > + d7 = z11 - z4;
> > +
> > + // Odd part of IDCT
> > +
> > + THRESHOLD(tmp4, d1, threshold[1 * 8]);
> > + THRESHOLD(tmp5, d3, threshold[3 * 8]);
> > + THRESHOLD(tmp6, d5, threshold[5 * 8]);
> > + THRESHOLD(tmp7, d7, threshold[7 * 8]);
> > +
> > + //Simd version uses here a shortcut for the tmp5,tmp6,tmp7
> == 0
> > + z13 = tmp6 + tmp5;
> > + z10 = (tmp6 - tmp5) << 1;
> > + z11 = tmp4 + tmp7;
> > + z12 = (tmp4 - tmp7) << 1;
> > +
> > + tmp7 = (z11 + z13) >> 2; //+2 !
> > + tmp11 = MULTIPLY16H((z11 - z13) << 1, FIX_1_414213562);
> > + z5 = MULTIPLY16H(z10 + z12, FIX_1_847759065);
> > + tmp10 = MULTIPLY16H(z12, FIX_1_082392200) - z5;
> > + tmp12 = MULTIPLY16H(z10, FIX_2_613125930) +
> z5; // - !!
> > +
> > + tmp6 = tmp12 - tmp7;
> > + tmp5 = tmp11 - tmp6;
> > + tmp4 = tmp10 + tmp5;
> > +
> > + wsptr[DCTSIZE * 0] += (tmp0 + tmp7);
> > + wsptr[DCTSIZE * 1] += (tmp1 + tmp6);
> > + wsptr[DCTSIZE * 2] += (tmp2 + tmp5);
> > + wsptr[DCTSIZE * 3] += (tmp3 - tmp4);
> > + wsptr[DCTSIZE * 4] += (tmp3 + tmp4);
> > + wsptr[DCTSIZE * 5] += (tmp2 - tmp5);
> > + wsptr[DCTSIZE * 6] = (tmp1 - tmp6);
> > + wsptr[DCTSIZE * 7] = (tmp0 - tmp7);
> > + //
> > + dataptr++; //next column
> > + wsptr++;
> > + threshold++;
> > + }
> > + dataptr += 8; //skip each second start pos
> > + wsptr += 8;
> > + }
> > +}
> > +
> > +static void row_idct_c(int16_t* workspace, int16_t* output_adr, int
> output_stride, int cnt)
> > +{
> > + int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
> > + int_simd16_t tmp10, tmp11, tmp12, tmp13;
> > + int_simd16_t z5, z10, z11, z12, z13;
> > + int16_t* outptr;
> > + int16_t* wsptr;
> > +
> > + cnt *= 4;
> > + wsptr = workspace;
> > + outptr = output_adr;
> > + for (; cnt > 0; cnt--) {
> > + // Even part
> > + //Simd version reads 4x4 block and transposes it
> > + tmp10 = wsptr[2] + wsptr[3];
> > + tmp11 = wsptr[2] - wsptr[3];
> > +
> > + tmp13 = wsptr[0] + wsptr[1];
> > + tmp12 = (MULTIPLY16H(wsptr[0] - wsptr[1], FIX_1_414213562_A) <<
> 2) - tmp13;//this shift order to avoid overflow
> > +
> > + tmp0 = tmp10 + tmp13; //->temps
> > + tmp3 = tmp10 - tmp13; //->temps
> > + tmp1 = tmp11 + tmp12;
> > + tmp2 = tmp11 - tmp12;
> > +
> > + // Odd part
> > + //Also transpose, with previous:
> > + // ---- ---- ||||
> > + // ---- ---- idct ||||
> > + // ---- ---- ---> ||||
> > + // ---- ---- ||||
> > + z13 = wsptr[4] + wsptr[5];
> > + z10 = wsptr[4] - wsptr[5];
> > + z11 = wsptr[6] + wsptr[7];
> > + z12 = wsptr[6] - wsptr[7];
> > +
> > + tmp7 = z11 + z13;
> > + tmp11 = MULTIPLY16H(z11 - z13, FIX_1_414213562);
> > +
> > + z5 = MULTIPLY16H(z10 + z12, FIX_1_847759065);
> > + tmp10 = MULTIPLY16H(z12, FIX_1_082392200) - z5;
> > + tmp12 = MULTIPLY16H(z10, FIX_2_613125930) + z5; // - FIX_
> > +
> > + tmp6 = (tmp12 << 3) - tmp7;
> > + tmp5 = (tmp11 << 3) - tmp6;
> > + tmp4 = (tmp10 << 3) + tmp5;
> > +
> > + // Final output stage: descale and write column
> > + outptr[0 * output_stride] += DESCALE(tmp0 + tmp7, 3);
> > + outptr[1 * output_stride] += DESCALE(tmp1 + tmp6, 3);
> > + outptr[2 * output_stride] += DESCALE(tmp2 + tmp5, 3);
> > + outptr[3 * output_stride] += DESCALE(tmp3 - tmp4, 3);
> > + outptr[4 * output_stride] += DESCALE(tmp3 + tmp4, 3);
> > + outptr[5 * output_stride] += DESCALE(tmp2 - tmp5, 3);
> > + outptr[6 * output_stride] += DESCALE(tmp1 - tmp6, 3); //no += ?
> > + outptr[7 * output_stride] += DESCALE(tmp0 - tmp7, 3); //no += ?
> > + outptr++;
> > +
> > + wsptr += DCTSIZE; // advance pointer to next row
> > + }
> > +}
> > +
> > +static void row_fdct_c(int16_t *data, const uint8_t *pixels, int
> line_size, int cnt)
> > +{
> > + int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
> > + int_simd16_t tmp10, tmp11, tmp12, tmp13;
> > + int_simd16_t z1, z2, z3, z4, z5, z11, z13;
> > + int16_t *dataptr;
> > +
> > + cnt*=4;
>
> nit: cnt *= 4;
>
> > + // Pass 1: process rows.
> > +
> > + dataptr = data;
> > + for (; cnt > 0; cnt--) {
> > + tmp0 = pixels[line_size * 0] + pixels[line_size * 7];
> > + tmp7 = pixels[line_size * 0] - pixels[line_size * 7];
> > + tmp1 = pixels[line_size * 1] + pixels[line_size * 6];
> > + tmp6 = pixels[line_size * 1] - pixels[line_size * 6];
> > + tmp2 = pixels[line_size * 2] + pixels[line_size * 5];
> > + tmp5 = pixels[line_size * 2] - pixels[line_size * 5];
> > + tmp3 = pixels[line_size * 3] + pixels[line_size * 4];
> > + tmp4 = pixels[line_size * 3] - pixels[line_size * 4];
> > +
> > + // Even part
> > +
> > + tmp10 = tmp0 + tmp3;
> > + tmp13 = tmp0 - tmp3;
> > + tmp11 = tmp1 + tmp2;
> > + tmp12 = tmp1 - tmp2;
> > + //Even columns are written first, this leads to different order
> of columns
> > + //in column_fidct(), but they are processed independently, so
> all ok.
> > + //Later in the row_idct() columns readed at the same order.
> > + dataptr[2] = tmp10 + tmp11;
> > + dataptr[3] = tmp10 - tmp11;
> > +
> > + z1 = MULTIPLY16H((tmp12 + tmp13) << 2, FIX_0_707106781);
> > + dataptr[0] = tmp13 + z1;
> > + dataptr[1] = tmp13 - z1;
> > +
> > + // Odd part
> > +
> > + tmp10 = (tmp4 + tmp5) << 2;
> > + tmp11 = (tmp5 + tmp6) << 2;
> > + tmp12 = (tmp6 + tmp7) << 2;
> > +
> > + z5 = MULTIPLY16H(tmp10 - tmp12, FIX_0_382683433);
> > + z2 = MULTIPLY16H(tmp10, FIX_0_541196100) + z5;
> > + z4 = MULTIPLY16H(tmp12, FIX_1_306562965) + z5;
> > + z3 = MULTIPLY16H(tmp11, FIX_0_707106781);
> > +
> > + z11 = tmp7 + z3;
> > + z13 = tmp7 - z3;
> > +
> > + dataptr[4] = z13 + z2;
> > + dataptr[5] = z13 - z2;
> > + dataptr[6] = z11 + z4;
> > + dataptr[7] = z11 - z4;
> > +
> > + pixels++; // advance pointer to next column
> > + dataptr += DCTSIZE;
> > + }
> > +}
> > +
> > +static int query_formats(AVFilterContext *ctx)
> > +{
> > + static const enum PixelFormat pix_fmts[] = {
> > + AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV422P,
> > + AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV411P,
> > + AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV440P,
> > + AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ422P,
> > + AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ440P,
> > + AV_PIX_FMT_NONE
> > + };
> > + ff_set_common_formats(ctx, ff_make_format_list(pix_fmts));
> > + return 0;
> > +}
> > +
> > +static int config_input(AVFilterLink *inlink)
> > +{
> > + AVFilterContext *ctx = inlink->dst;
> > + FSPPContext *fspp = ctx->priv;
> > + const int h = FFALIGN(inlink->h + 16, 16);
> > + const AVPixFmtDescriptor *desc =
> av_pix_fmt_desc_get(inlink->format);
> > +
> > + fspp->hsub = desc->log2_chroma_w;
> > + fspp->vsub = desc->log2_chroma_h;
> > +
> > + fspp->temp_stride = FFALIGN(inlink->w + 16, 16);
> > + fspp->temp = av_malloc_array(fspp->temp_stride, h *
> sizeof(*fspp->temp));
> > + fspp->src = av_malloc_array(fspp->temp_stride, h *
> sizeof(*fspp->src));
> > +
> > + if (!fspp->temp || !fspp->src)
> > + return AVERROR(ENOMEM);
> > +
> > + if (!fspp->use_bframe_qp) {
> > + fspp->non_b_qp_alloc_size = FF_CEIL_RSHIFT(inlink->w, 4) *
> FF_CEIL_RSHIFT(inlink->h, 4);
> > + fspp->non_b_qp_table = av_calloc(fspp->non_b_qp_alloc_size,
> sizeof(*fspp->non_b_qp_table));
> > + if (!fspp->non_b_qp_table)
> > + return AVERROR(ENOMEM);
> > + }
> > +
> > + fspp->store_slice = store_slice_c;
> > + fspp->store_slice2 = store_slice2_c;
> > + fspp->mul_thrmat = mul_thrmat_c;
> > + fspp->column_fidct = column_fidct_c;
> > + fspp->row_idct = row_idct_c;
> > + fspp->row_fdct = row_fdct_c;
> > +
> > + if (ARCH_X86)
> > + ff_fspp_init_x86(fspp);
> > +
> > + return 0;
> > +}
> > +
> > +static int filter_frame(AVFilterLink *inlink, AVFrame *in)
> > +{
> > + AVFilterContext *ctx = inlink->dst;
> > + FSPPContext *fspp = ctx->priv;
> > + AVFilterLink *outlink = ctx->outputs[0];
> > + AVFrame *out = in;
> > +
> > + int qp_stride = 0;
> > + uint8_t *qp_table = NULL;
> > + int i, bias;
> > + int custom_threshold_m[64];
> > +
> > + bias = (1 << 4) + fspp->strength;
> > +
> > + for (i = 0; i < 64; i++) //FIXME: tune custom_threshold[] and
> remove this !
> > + custom_threshold_m[i] = (uint64_t)(custom_threshold[i] * (bias /
> 71) + 0.5);
>
> style: indent
>
> 71. -> 71
>
> Also why are you not using int as in the original code?
>
I am using int. Why shouldn't 71. be used?
>
> > +
> > + for (i = 0; i < 8; i++) {
> > +
> > + fspp->threshold_mtx_noq[2 * i] = (uint64_t)custom_threshold_m[i
> * 8 + 2]
> > + |(((uint64_t)custom_threshold_m[i
> * 8 + 6]) << 16)
> > + |(((uint64_t)custom_threshold_m[i
> * 8 + 0]) << 32)
> > + |(((uint64_t)custom_threshold_m[i
> * 8 + 4]) << 48);
> > +
> > + fspp->threshold_mtx_noq[2 * i + 1] =
> (uint64_t)custom_threshold_m[i * 8 + 5]
> > +
> |(((uint64_t)custom_threshold_m[i * 8 + 3]) << 16)
> > +
> |(((uint64_t)custom_threshold_m[i * 8 + 1]) << 32)
> > +
> |(((uint64_t)custom_threshold_m[i * 8 + 7]) << 48);
> > + }
> > +
>
> > + if (fspp->qp)
> > + fspp->prev_q = fspp->qp, fspp->mul_thrmat(fspp, fspp->qp);
>
> style
>
> [...]
>
> > +typedef struct fsppContext {
>
> what if you do:
>
> typedef struct FSPPContext {
> ...
> ... struct FSPPContext ...
> } FSPPContext;
>
> ?
> (I hate arbitrary name mangling)
>
> [...]
> --
> FFmpeg = Friendly Fabulous Maxi Proud Eretic Gorilla
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
The output differs a lot with and without -cpuflags 0.
Updated the patch.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-lavfi-port-mp-fspp-to-a-native-libavfilter-filter.patch
Type: text/x-patch
Size: 93051 bytes
Desc: not available
URL: <https://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20141223/b6123aa7/attachment.bin>
More information about the ffmpeg-devel
mailing list