virtualenv/FFmpeg
1
0
Fork 0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2025-11-23 21:54:53 +02:00
Code Issues Releases Activity

avfilter/vf_fspp: Add DSPCtx, move DSP functions to file of their own

Browse Source 
This is in preparation for adding checkasm tests; without it,
checkasm would pull all of libavfilter in.

Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
This commit is contained in:
Andreas Rheinhardt
2025-11-09 17:06:46 +01:00
parent 57d6898730
commit 9b34088c4d
5 changed files with 455 additions and 404 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
libavfilter/Makefile
View File

@@ -329,7 +329,7 @@ OBJS-$(CONFIG_FRAMESTEP_FILTER) += vf_framestep.o
OBJS-$(CONFIG_FREEZEDETECT_FILTER) += vf_freezedetect.o
OBJS-$(CONFIG_FREEZEFRAMES_FILTER) += vf_freezeframes.o
OBJS-$(CONFIG_FREI0R_FILTER) += vf_frei0r.o
OBJS-$(CONFIG_FSPP_FILTER) += vf_fspp.o qp_table.o
OBJS-$(CONFIG_FSPP_FILTER) += vf_fspp.o vf_fsppdsp.o qp_table.o
OBJS-$(CONFIG_FSYNC_FILTER) += vf_fsync.o
OBJS-$(CONFIG_GBLUR_FILTER) += vf_gblur.o
OBJS-$(CONFIG_GBLUR_VULKAN_FILTER) += vf_gblur_vulkan.o vulkan.o vulkan_filter.o

399
libavfilter/vf_fspp.c
View File

@@ -41,12 +41,40 @@
#include "libavutil/mem_internal.h"
#include "libavutil/opt.h"
#include "libavutil/pixdesc.h"
#include "libavutil/video_enc_params.h"
#include "avfilter.h"
#include "filters.h"
#include "qp_table.h"
#include "vf_fspp.h"
#include "vf_fsppdsp.h"
#include "video.h"
#define BLOCKSZ 12
#define MAX_LEVEL 5
typedef struct FSPPContext {
const struct AVClass *class;
uint64_t threshold_mtx_noq[8 * 2];
uint64_t threshold_mtx[8 * 2]; //used in both C & MMX (& later SSE2) versions
int log2_count;
int strength;
int hsub;
int vsub;
int temp_stride;
int qp;
enum AVVideoEncParamsType qscale_type;
int prev_q;
uint8_t *src;
int16_t *temp;
int8_t *non_b_qp_table;
int non_b_qp_stride;
int use_bframe_qp;
FSPPDSPContext dsp;
} FSPPContext;
#define OFFSET(x) offsetof(FSPPContext, x)
#define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
static const AVOption fspp_options[] = {
@@ -59,17 +87,6 @@ static const AVOption fspp_options[] = {
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
@@ -84,73 +101,6 @@ static const short custom_threshold[64] = {
20, 27, 26, 23, 20, 15, 11, 5
};
//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(int16_t *thr_adr_noq, int16_t *thr_adr, int q)
{
int a;
for (a = 0; a < 64; a++)
thr_adr[a] = q * thr_adr_noq[a];
}
static void filter(FSPPContext *p, uint8_t *dst, uint8_t *src,
int dst_stride, int src_stride,
int width, int height,
@@ -197,13 +147,13 @@ static void filter(FSPPContext *p, uint8_t *dst, uint8_t *src,
if (qy < 0) qy = 0;
qy = (qy >> qpsv) * qp_stride;
p->row_fdct(block, p->src + y * stride + 2 - (y&1), stride, 2);
p->dsp.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));
p->dsp.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
p->dsp.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
@@ -213,288 +163,42 @@ static void filter(FSPPContext *p, uint8_t *dst, uint8_t *src,
t = qp_store[qy + (t >> qpsh)];
t = ff_norm_qscale(t, p->qscale_type);
if (t != p->prev_q) p->prev_q = t, p->mul_thrmat((int16_t *)(&p->threshold_mtx_noq[0]), (int16_t *)(&p->threshold_mtx[0]), t);
p->column_fidct((int16_t *)(&p->threshold_mtx[0]), block + x * 8, block3 + x * 8, 8); //yes, this is a HOTSPOT
if (t != p->prev_q) p->prev_q = t, p->dsp.mul_thrmat((int16_t *)(&p->threshold_mtx_noq[0]), (int16_t *)(&p->threshold_mtx[0]), t);
p->dsp.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));
p->dsp.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->dsp.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->dsp.column_fidct((int16_t *)(&p->threshold_mtx[0]), block, block3, es&(~1));
if (es > 3)
p->row_idct(block3 + 0 * 8, p->temp + (y & 15) * stride + x0 + 2 - (y & 1), stride, es >> 2);
p->dsp.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);
p->dsp.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);
p->dsp.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 % 8 != 0
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);
p->dsp.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,
p->dsp.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)
{
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, ptrdiff_t 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, ptrdiff_t 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;
// 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 are read in 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 const enum AVPixelFormat pix_fmts[] = {
AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV422P,
AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV411P,
@@ -522,16 +226,7 @@ static int config_input(AVFilterLink *inlink)
if (!fspp->temp || !fspp->src)
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);
#endif
ff_fsppdsp_init(&fspp->dsp);
return 0;
}
@@ -567,7 +262,7 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in)
}
if (fspp->qp)
fspp->prev_q = fspp->qp, fspp->mul_thrmat((int16_t *)(&fspp->threshold_mtx_noq[0]), (int16_t *)(&fspp->threshold_mtx[0]), fspp->qp);
fspp->prev_q = fspp->qp, fspp->dsp.mul_thrmat((int16_t *)(&fspp->threshold_mtx_noq[0]), (int16_t *)(&fspp->threshold_mtx[0]), fspp->qp);
/* if we are not in a constant user quantizer mode and we don't want to use
* the quantizers from the B-frames (B-frames often have a higher QP), we

369
libavfilter/vf_fsppdsp.c Normal file
View File

@@ -0,0 +1,369 @@
/*
* Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
* Copyright (C) 2005 Nikolaj Poroshin <porosh3@psu.ru>
* Copyright (c) 2014 Arwa Arif <arwaarif1994@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.
*/
#include <stdint.h>
#include "vf_fsppdsp.h"
#include "libavutil/mathematics.h"
#include "libavutil/mem_internal.h"
#define DCTSIZE 8
#define FIX(x,s) ((x) * (1 << s) + 0.5)
#define MULTIPLY16H(x,k) (((x) * (k)) >> 16)
#define THRESHOLD(r,x,t) \
if(((unsigned)((x) + t)) > t * 2) r = (x); \
else r = 0;
#define DESCALE(x,n) (((x) + (1 << ((n) - 1))) >> n)
typedef int32_t int_simd16_t;
static const int16_t FIX_0_382683433 = FIX(0.382683433, 14);
static const int16_t FIX_0_541196100 = FIX(0.541196100, 14);
static const int16_t FIX_0_707106781 = FIX(M_SQRT1_2 , 14);
static const int16_t FIX_1_306562965 = FIX(1.306562965, 14);
static const int16_t FIX_1_414213562_A = FIX(M_SQRT2 , 14);
static const int16_t FIX_1_847759065 = FIX(1.847759065, 13);
static const int16_t FIX_2_613125930 = FIX(-2.613125930, 13);
static const int16_t FIX_1_414213562 = FIX(M_SQRT2 , 13);
static const int16_t FIX_1_082392200 = FIX(1.082392200, 13);
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, },
};
//This func reads from 1 slice, 1 and clears 0 & 1
void ff_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)
{
#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 (int y = 0; y < height; y++) {
const uint8_t *d = dither[y];
for (int 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
void ff_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)
{
#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 (int y = 0; y < height; y++) {
const uint8_t *d = dither[y];
for (int 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;
}
}
void ff_mul_thrmat_c(int16_t *thr_adr_noq, int16_t *thr_adr, int q)
{
for (int a = 0; a < 64; a++)
thr_adr[a] = q * thr_adr_noq[a];
}
void ff_column_fidct_c(int16_t *thr_adr, int16_t *data, int16_t *output, 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, z10, z11, z12, z13;
int_simd16_t d0, d1, d2, d3, d4, d5, d6, d7;
int16_t *dataptr;
int16_t *wsptr;
int16_t *threshold;
dataptr = data;
wsptr = output;
for (; cnt > 0; cnt -= 2) { //start positions
threshold = (int16_t *)thr_adr;//threshold_mtx
for (int 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;
}
}
void ff_row_idct_c(int16_t *workspace, int16_t *output_adr, ptrdiff_t 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
}
}
void ff_row_fdct_c(int16_t *data, const uint8_t *pixels, ptrdiff_t 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;
// 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 are read in 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;
}
}

85
libavfilter/vf_fspp.h → libavfilter/vf_fsppdsp.h
View File

@@ -20,56 +20,17 @@
* 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
*/
#ifndef AVFILTER_FSPP_H
#define AVFILTER_FSPP_H
#ifndef AVFILTER_FSPPDSP_H
#define AVFILTER_FSPPDSP_H
#include "libavutil/video_enc_params.h"
#include "avfilter.h"
#include <stddef.h>
#include <stdint.h>
#define BLOCKSZ 12
#define MAX_LEVEL 5
#include "config.h"
#define DCTSIZE 8
#define DCTSIZE_S "8"
#define FIX(x,s) ((x) * (1 << s) + 0.5)
#define MULTIPLY16H(x,k) (((x) * (k)) >> 16)
#define THRESHOLD(r,x,t) \
if(((unsigned)((x) + t)) > t * 2) r = (x); \
else r = 0;
#define DESCALE(x,n) (((x) + (1 << ((n) - 1))) >> n)
typedef int32_t int_simd16_t;
static const int16_t FIX_0_382683433 = FIX(0.382683433, 14);
static const int16_t FIX_0_541196100 = FIX(0.541196100, 14);
static const int16_t FIX_0_707106781 = FIX(M_SQRT1_2 , 14);
static const int16_t FIX_1_306562965 = FIX(1.306562965, 14);
static const int16_t FIX_1_414213562_A = FIX(M_SQRT2 , 14);
static const int16_t FIX_1_847759065 = FIX(1.847759065, 13);
static const int16_t FIX_2_613125930 = FIX(-2.613125930, 13);
static const int16_t FIX_1_414213562 = FIX(M_SQRT2 , 13);
static const int16_t FIX_1_082392200 = FIX(1.082392200, 13);
typedef struct FSPPContext {
AVClass *class;
uint64_t threshold_mtx_noq[8 * 2];
uint64_t threshold_mtx[8 * 2]; //used in both C & MMX (& later SSE2) versions
int log2_count;
int strength;
int hsub;
int vsub;
int temp_stride;
int qp;
enum AVVideoEncParamsType qscale_type;
int prev_q;
uint8_t *src;
int16_t *temp;
int8_t *non_b_qp_table;
int non_b_qp_stride;
int use_bframe_qp;
#include "libavutil/attributes_internal.h"
typedef struct FSPPDSPContext {
void (*store_slice)(uint8_t *dst, int16_t *src,
ptrdiff_t dst_stride, ptrdiff_t src_stride,
ptrdiff_t width, ptrdiff_t height, ptrdiff_t log2_scale);
@@ -88,9 +49,35 @@ typedef struct FSPPContext {
void (*row_fdct)(int16_t *data, const uint8_t *pixels,
ptrdiff_t line_size, int cnt);
} FSPPDSPContext;
} FSPPContext;
FF_VISIBILITY_PUSH_HIDDEN
void ff_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);
void ff_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);
void ff_mul_thrmat_c(int16_t *thr_adr_noq, int16_t *thr_adr, int q);
void ff_column_fidct_c(int16_t *thr_adr, int16_t *data, int16_t *output, int cnt);
void ff_row_idct_c(int16_t *workspace, int16_t *output_adr, ptrdiff_t output_stride, int cnt);
void ff_row_fdct_c(int16_t *data, const uint8_t *pixels, ptrdiff_t line_size, int cnt);
void ff_fspp_init_x86(FSPPContext *fspp);
void ff_fsppdsp_init_x86(FSPPDSPContext *fspp);
FF_VISIBILITY_POP_HIDDEN
#endif /* AVFILTER_FSPP_H */
static inline void ff_fsppdsp_init(FSPPDSPContext *fspp)
{
fspp->store_slice = ff_store_slice_c;
fspp->store_slice2 = ff_store_slice2_c;
fspp->mul_thrmat = ff_mul_thrmat_c;
fspp->column_fidct = ff_column_fidct_c;
fspp->row_idct = ff_row_idct_c;
fspp->row_fdct = ff_row_fdct_c;
#if ARCH_X86
ff_fsppdsp_init_x86(fspp);
#endif
}
#endif /* AVFILTER_FSPPDSP_H */

4
libavfilter/x86/vf_fspp_init.c
View File

@@ -21,7 +21,7 @@
#include "libavutil/attributes.h"
#include "libavutil/x86/cpu.h"
#include "libavfilter/vf_fspp.h"
#include "libavfilter/vf_fsppdsp.h"
void ff_store_slice_mmx(uint8_t *dst, int16_t *src,
ptrdiff_t dst_stride, ptrdiff_t src_stride,
@@ -34,7 +34,7 @@ void ff_column_fidct_mmx(int16_t *thr_adr, int16_t *data, int16_t *output, int c
void ff_row_idct_mmx(int16_t *workspace, int16_t *output_adr, ptrdiff_t output_stride, int cnt);
void ff_row_fdct_mmx(int16_t *data, const uint8_t *pixels, ptrdiff_t line_size, int cnt);
av_cold void ff_fspp_init_x86(FSPPContext *s)
av_cold void ff_fsppdsp_init_x86(FSPPDSPContext *s)
{
int cpu_flags = av_get_cpu_flags();