Skip to content

F
ffmpeg.wasm-core
Commit 46803f4f authored by Loren Merritt's avatar Loren Merritt
Browse files
Options

optimize imdct_half: 

remove tmp buffer.
skip fft reinterleave pass, leaving data in a format more convenient for simd.
merge post-rotate with post-reorder.

Originally committed as revision 14700 to svn://svn.ffmpeg.org/ffmpeg/trunk
parent 49c0dd75
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 241 additions and 352 deletions
libavcodec/dsputil.h
View file @ 46803f4f
......@@ -645,7 +645,7 @@ typedef struct FFTContext {
void (*imdct_calc)(struct MDCTContext *s, FFTSample *output,
const FFTSample *input, FFTSample *tmp);
void (*imdct_half)(struct MDCTContext *s, FFTSample *output,
const FFTSample *input, FFTSample *tmp);
const FFTSample *input);
} FFTContext;
int ff_fft_init(FFTContext *s, int nbits, int inverse);
......@@ -696,16 +696,16 @@ void ff_sine_window_init(float *window, int n);
int ff_mdct_init(MDCTContext *s, int nbits, int inverse);
void ff_imdct_calc(MDCTContext *s, FFTSample *output,
const FFTSample *input, FFTSample *tmp);
void ff_imdct_half(MDCTContext *s, FFTSample *output,
const FFTSample *input, FFTSample *tmp);
void ff_imdct_half(MDCTContext *s, FFTSample *output, const FFTSample *input);
void ff_imdct_calc_3dn(MDCTContext *s, FFTSample *output,
const FFTSample *input, FFTSample *tmp);
void ff_imdct_half_3dn(MDCTContext *s, FFTSample *output, const FFTSample *input);
void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output,
const FFTSample *input, FFTSample *tmp);
void ff_imdct_half_3dn2(MDCTContext *s, FFTSample *output,
const FFTSample *input, FFTSample *tmp);
void ff_imdct_half_3dn2(MDCTContext *s, FFTSample *output, const FFTSample *input);
void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output,
const FFTSample *input, FFTSample *tmp);
void ff_imdct_half_sse(MDCTContext *s, FFTSample *output,
const FFTSample *input, FFTSample *tmp);
void ff_imdct_half_sse(MDCTContext *s, FFTSample *output, const FFTSample *input);
void ff_mdct_calc(MDCTContext *s, FFTSample *out,
const FFTSample *input, FFTSample *tmp);
void ff_mdct_end(MDCTContext *s);
......
libavcodec/fft.c
View file @ 46803f4f
......@@ -106,6 +106,8 @@ int ff_fft_init(FFTContext *s, int nbits, int inverse)
s->fft_calc = ff_fft_calc_3dn2;
} else if (has_vectors & MM_3DNOW) {
/* 3DNow! for K6-2/3 */
s->imdct_calc = ff_imdct_calc_3dn;
s->imdct_half = ff_imdct_half_3dn;
s->fft_calc = ff_fft_calc_3dn;
}
#elif defined HAVE_ALTIVEC && !defined ALTIVEC_USE_REFERENCE_C_CODE
......
libavcodec/i386/fft_3dn2.c
View file @ 46803f4f
/*
* FFT/MDCT transform with Extended 3DNow! optimizations
* Copyright (c) 2006 Zuxy MENG Jie, Loren Merritt
* Based on fft_sse.c copyright (c) 2002 Fabrice Bellard.
* Copyright (c) 2006-2008 Zuxy MENG Jie, Loren Merritt
*
* This file is part of FFmpeg.
*
......@@ -23,16 +22,23 @@
#include "libavutil/x86_cpu.h"
#include "libavcodec/dsputil.h"
DECLARE_ALIGNED_8(static const int, m1m1[2]) = { 1<<31, 1<<31 };
#ifdef EMULATE_3DNOWEXT
#define PSWAPD(s,d)\
"movq "#s","#d"\n"\
"psrlq $32,"#d"\n"\
"punpckldq "#s","#d"\n"
#define ff_fft_calc_3dn2 ff_fft_calc_3dn
#define ff_fft_dispatch_3dn2 ff_fft_dispatch_3dn
#define ff_fft_dispatch_interleave_3dn2 ff_fft_dispatch_interleave_3dn
#define ff_imdct_calc_3dn2 ff_imdct_calc_3dn
#define ff_imdct_half_3dn2 ff_imdct_half_3dn
#else
#define PSWAPD(s,d) "pswapd "#s","#d"\n"
#endif
void ff_fft_dispatch_3dn2(FFTComplex *z, int nbits);
void ff_fft_dispatch_interleave_3dn2(FFTComplex *z, int nbits);
void ff_fft_calc_3dn2(FFTContext *s, FFTComplex *z)
{
......@@ -45,35 +51,45 @@ void ff_fft_calc_3dn2(FFTContext *s, FFTComplex *z)
FFSWAP(FFTSample, z[i].im, z[i+1].re);
}
static void imdct_3dn2(MDCTContext *s, const FFTSample *input, FFTSample *tmp)
void ff_imdct_half_3dn2(MDCTContext *s, FFTSample *output, const FFTSample *input)
{
long n4, n2, n;
x86_reg k;
x86_reg j, k;
long n = 1 << s->nbits;
long n2 = n >> 1;
long n4 = n >> 2;
long n8 = n >> 3;
const uint16_t *revtab = s->fft.revtab;
const FFTSample *tcos = s->tcos;
const FFTSample *tsin = s->tsin;
const FFTSample *in1, *in2;
FFTComplex *z = (FFTComplex *)tmp;
n = 1 << s->nbits;
n2 = n >> 1;
n4 = n >> 2;
FFTComplex *z = (FFTComplex *)output;
/* pre rotation */
in1 = input;
in2 = input + n2 - 1;
#ifdef EMULATE_3DNOWEXT
asm volatile("movd %0, %%mm7" ::"r"(1<<31));
#endif
for(k = 0; k < n4; k++) {
// FIXME a single block is faster, but gcc 2.95 and 3.4.x on 32bit can't compile it
asm volatile(
"movd %0, %%mm0 \n\t"
"movd %2, %%mm1 \n\t"
"punpckldq %1, %%mm0 \n\t"
"punpckldq %3, %%mm1 \n\t"
"movq %%mm0, %%mm2 \n\t"
"pfmul %%mm1, %%mm0 \n\t"
"pswapd %%mm1, %%mm1 \n\t"
"pfmul %%mm1, %%mm2 \n\t"
"pfpnacc %%mm2, %%mm0 \n\t"
"movd %0, %%mm0 \n"
"movd %2, %%mm1 \n"
"punpckldq %1, %%mm0 \n"
"punpckldq %3, %%mm1 \n"
"movq %%mm0, %%mm2 \n"
PSWAPD( %%mm1, %%mm3 )
"pfmul %%mm1, %%mm0 \n"
"pfmul %%mm3, %%mm2 \n"
#ifdef EMULATE_3DNOWEXT
"movq %%mm0, %%mm1 \n"
"punpckhdq %%mm2, %%mm0 \n"
"punpckldq %%mm2, %%mm1 \n"
"pxor %%mm7, %%mm0 \n"
"pfadd %%mm1, %%mm0 \n"
#else
"pfpnacc %%mm2, %%mm0 \n"
#endif
::"m"(in2[-2*k]), "m"(in1[2*k]),
"m"(tcos[k]), "m"(tsin[k])
);
......@@ -83,101 +99,74 @@ static void imdct_3dn2(MDCTContext *s, const FFTSample *input, FFTSample *tmp)
);
}
ff_fft_calc_3dn2(&s->fft, z);
/* post rotation + reordering */
for(k = 0; k < n4; k++) {
asm volatile(
"movq %0, %%mm0 \n\t"
"movd %1, %%mm1 \n\t"
"punpckldq %2, %%mm1 \n\t"
"movq %%mm0, %%mm2 \n\t"
"pfmul %%mm1, %%mm0 \n\t"
"pswapd %%mm1, %%mm1 \n\t"
"pfmul %%mm1, %%mm2 \n\t"
"pfpnacc %%mm2, %%mm0 \n\t"
"movq %%mm0, %0 \n\t"
:"+m"(z[k])
:"m"(tcos[k]), "m"(tsin[k])
);
}
}
void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output,
const FFTSample *input, FFTSample *tmp)
{
x86_reg k;
long n8, n2, n;
FFTComplex *z = (FFTComplex *)tmp;
n = 1 << s->nbits;
n2 = n >> 1;
n8 = n >> 3;
imdct_3dn2(s, input, tmp);
k = n-8;
asm volatile("movd %0, %%mm7" ::"r"(1<<31));
ff_fft_dispatch_3dn2(z, s->fft.nbits);
#define CMUL(j,mm0,mm1)\
"movq (%2,"#j",2), %%mm6 \n"\
"movq 8(%2,"#j",2), "#mm0"\n"\
"movq %%mm6, "#mm1"\n"\
"movq "#mm0",%%mm7 \n"\
"pfmul (%3,"#j"), %%mm6 \n"\
"pfmul (%4,"#j"), "#mm0"\n"\
"pfmul (%4,"#j"), "#mm1"\n"\
"pfmul (%3,"#j"), %%mm7 \n"\
"pfsub %%mm6, "#mm0"\n"\
"pfadd %%mm7, "#mm1"\n"
/* post rotation */
j = -n2;
k = n2-8;
asm volatile(
"1: \n\t"
"movq (%4,%0), %%mm0 \n\t" // z[n8+k]
"neg %0 \n\t"
"pswapd -8(%4,%0), %%mm1 \n\t" // z[n8-1-k]
"movq %%mm0, %%mm2 \n\t"
"pxor %%mm7, %%mm2 \n\t"
"punpckldq %%mm1, %%mm2 \n\t"
"pswapd %%mm2, %%mm3 \n\t"
"punpckhdq %%mm1, %%mm0 \n\t"
"pswapd %%mm0, %%mm4 \n\t"
"pxor %%mm7, %%mm0 \n\t"
"pxor %%mm7, %%mm4 \n\t"
"movq %%mm3, -8(%3,%0) \n\t" // output[n-2-2*k] = { z[n8-1-k].im, -z[n8+k].re }
"movq %%mm4, -8(%2,%0) \n\t" // output[n2-2-2*k]= { -z[n8-1-k].re, z[n8+k].im }
"neg %0 \n\t"
"movq %%mm0, (%1,%0) \n\t" // output[2*k] = { -z[n8+k].im, z[n8-1-k].re }
"movq %%mm2, (%2,%0) \n\t" // output[n2+2*k] = { -z[n8+k].re, z[n8-1-k].im }
"sub $8, %0 \n\t"
"jge 1b \n\t"
:"+r"(k)
:"r"(output), "r"(output+n2), "r"(output+n), "r"(z+n8)
"1: \n"
CMUL(%0, %%mm0, %%mm1)
CMUL(%1, %%mm2, %%mm3)
"movd %%mm0, (%2,%0,2) \n"
"movd %%mm1,12(%2,%1,2) \n"
"movd %%mm2, (%2,%1,2) \n"
"movd %%mm3,12(%2,%0,2) \n"
"psrlq $32, %%mm0 \n"
"psrlq $32, %%mm1 \n"
"psrlq $32, %%mm2 \n"
"psrlq $32, %%mm3 \n"
"movd %%mm0, 8(%2,%0,2) \n"
"movd %%mm1, 4(%2,%1,2) \n"
"movd %%mm2, 8(%2,%1,2) \n"
"movd %%mm3, 4(%2,%0,2) \n"
"sub $8, %1 \n"
"add $8, %0 \n"
"jl 1b \n"
:"+r"(j), "+r"(k)
:"r"(z+n8), "r"(tcos+n8), "r"(tsin+n8)
:"memory"
);
asm volatile("femms");
}
void ff_imdct_half_3dn2(MDCTContext *s, FFTSample *output,
void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output,
const FFTSample *input, FFTSample *tmp)
{
x86_reg j, k;
long n8, n4, n;
FFTComplex *z = (FFTComplex *)tmp;
long n = 1 << s->nbits;
long n4 = n >> 2;
n = 1 << s->nbits;
n4 = n >> 2;
n8 = n >> 3;
imdct_3dn2(s, input, tmp);
ff_imdct_half_3dn2(s, output+n4, input);
j = -n;
k = n-8;
asm volatile("movd %0, %%mm7" ::"r"(1<<31));
asm volatile(
"1: \n\t"
"movq (%3,%1), %%mm0 \n\t" // z[n8+k]
"pswapd (%3,%0), %%mm1 \n\t" // z[n8-1-k]
"movq %%mm0, %%mm2 \n\t"
"punpckldq %%mm1, %%mm0 \n\t"
"punpckhdq %%mm2, %%mm1 \n\t"
"pxor %%mm7, %%mm0 \n\t"
"pxor %%mm7, %%mm1 \n\t"
"movq %%mm0, (%2,%1) \n\t" // output[n4+2*k] = { -z[n8+k].re, z[n8-1-k].im }
"movq %%mm1, (%2,%0) \n\t" // output[n4-2-2*k] = { -z[n8-1-k].re, z[n8+k].im }
"sub $8, %1 \n\t"
"add $8, %0 \n\t"
"jl 1b \n\t"
"movq %4, %%mm7 \n"
"1: \n"
PSWAPD((%2,%1), %%mm0)
PSWAPD((%3,%0), %%mm1)
"pxor %%mm7, %%mm0 \n"
"movq %%mm1, (%3,%1) \n"
"movq %%mm0, (%2,%0) \n"
"sub $8, %1 \n"
"add $8, %0 \n"
"jl 1b \n"
:"+r"(j), "+r"(k)
:"r"(output+n4), "r"(z+n8)
:"memory"
:"r"(output+n4), "r"(output+n4*3),
"m"(*m1m1)
);
asm volatile("femms");
}
......
libavcodec/i386/fft_sse.c
View file @ 46803f4f
This diff is collapsed.
libavcodec/mdct.c
View file @ 46803f4f
......@@ -100,18 +100,25 @@ int ff_mdct_init(MDCTContext *s, int nbits, int inverse)
(pim) = _are * _bim + _aim * _bre;\
}
static void imdct_c(MDCTContext *s, const FFTSample *input, FFTSample *tmp)
/**
* Compute the middle half of the inverse MDCT of size N = 2^nbits,
* thus excluding the parts that can be derived by symmetry
* @param output N/2 samples
* @param input N/2 samples
*/
void ff_imdct_half(MDCTContext *s, FFTSample *output, const FFTSample *input)
{
int k, n4, n2, n, j;
int k, n8, n4, n2, n, j;
const uint16_t *revtab = s->fft.revtab;
const FFTSample *tcos = s->tcos;
const FFTSample *tsin = s->tsin;
const FFTSample *in1, *in2;
FFTComplex *z = (FFTComplex *)tmp;
FFTComplex *z = (FFTComplex *)output;
n = 1 << s->nbits;
n2 = n >> 1;
n4 = n >> 2;
n8 = n >> 3;
/* pre rotation */
in1 = input;
......@@ -125,9 +132,15 @@ static void imdct_c(MDCTContext *s, const FFTSample *input, FFTSample *tmp)
ff_fft_calc(&s->fft, z);
/* post rotation + reordering */
/* XXX: optimize */
for(k = 0; k < n4; k++) {
CMUL(z[k].re, z[k].im, z[k].re, z[k].im, tcos[k], tsin[k]);
output += n4;
for(k = 0; k < n8; k++) {
FFTSample r0, i0, r1, i1;
CMUL(r0, i1, z[n8-k-1].im, z[n8-k-1].re, tsin[n8-k-1], tcos[n8-k-1]);
CMUL(r1, i0, z[n8+k ].im, z[n8+k ].re, tsin[n8+k ], tcos[n8+k ]);
z[n8-k-1].re = r0;
z[n8-k-1].im = i0;
z[n8+k ].re = r1;
z[n8+k ].im = i1;
}
}
......@@ -140,52 +153,16 @@ static void imdct_c(MDCTContext *s, const FFTSample *input, FFTSample *tmp)
void ff_imdct_calc(MDCTContext *s, FFTSample *output,
const FFTSample *input, FFTSample *tmp)
{
int k, n8, n2, n;
FFTComplex *z = (FFTComplex *)tmp;
n = 1 << s->nbits;
n2 = n >> 1;
n8 = n >> 3;
imdct_c(s, input, tmp);
for(k = 0; k < n8; k++) {
output[2*k] = -z[n8 + k].im;
output[n2-1-2*k] = z[n8 + k].im;
int k;
int n = 1 << s->nbits;
int n2 = n >> 1;
int n4 = n >> 2;
output[2*k+1] = z[n8-1-k].re;
output[n2-1-2*k-1] = -z[n8-1-k].re;
output[n2 + 2*k]=-z[k+n8].re;
output[n-1- 2*k]=-z[k+n8].re;
output[n2 + 2*k+1]=z[n8-k-1].im;
output[n-2 - 2 * k] = z[n8-k-1].im;
}
}
ff_imdct_half(s, output+n4, input);
/**
* Compute the middle half of the inverse MDCT of size N = 2^nbits,
* thus excluding the parts that can be derived by symmetry
* @param output N/2 samples
* @param input N/2 samples
* @param tmp N/2 samples
*/
void ff_imdct_half(MDCTContext *s, FFTSample *output,
const FFTSample *input, FFTSample *tmp)
{
int k, n8, n4, n;
FFTComplex *z = (FFTComplex *)tmp;
n = 1 << s->nbits;
n4 = n >> 2;
n8 = n >> 3;
imdct_c(s, input, tmp);
for(k = 0; k < n8; k++) {
output[n4-1-2*k] = z[n8+k].im;
output[n4-1-2*k-1] = -z[n8-k-1].re;
output[n4 + 2*k] = -z[n8+k].re;
output[n4 + 2*k+1] = z[n8-k-1].im;
for(k = 0; k < n4; k++) {
output[k] = -output[n2-k-1];
output[n-k-1] = output[n2+k];
}
}
......@@ -203,7 +180,7 @@ void ff_mdct_calc(MDCTContext *s, FFTSample *out,
const uint16_t *revtab = s->fft.revtab;
const FFTSample *tcos = s->tcos;
const FFTSample *tsin = s->tsin;
FFTComplex *x = (FFTComplex *)tmp;
FFTComplex *x = (FFTComplex *)out;
n = 1 << s->nbits;
n2 = n >> 1;
......@@ -227,12 +204,14 @@ void ff_mdct_calc(MDCTContext *s, FFTSample *out,
ff_fft_calc(&s->fft, x);
/* post rotation */
for(i=0;i<n4;i++) {
re = x[i].re;
im = x[i].im;
CMUL(re1, im1, re, im, -tsin[i], -tcos[i]);
out[2*i] = im1;
out[n2-1-2*i] = re1;
for(i=0;i<n8;i++) {
FFTSample r0, i0, r1, i1;
CMUL(i1, r0, x[n8-i-1].re, x[n8-i-1].im, -tsin[n8-i-1], -tcos[n8-i-1]);
CMUL(i0, r1, x[n8+i ].re, x[n8+i ].im, -tsin[n8+i ], -tcos[n8+i ]);
x[n8-i-1].re = r0;
x[n8-i-1].im = i0;
x[n8+i ].re = r1;
x[n8+i ].im = i1;
}
}
......
libavcodec/vorbis_dec.c
View file @ 46803f4f
......@@ -1517,18 +1517,18 @@ static int vorbis_parse_audio_packet(vorbis_context *vc) {
// MDCT, overlap/add, save data for next overlapping FPMATH
retlen = (blocksize + vc->blocksize[previous_window])/4;
dir = retlen <= blocksize/2; // pick an order so that ret[] can reuse residues[] without stepping on any data we need
dir = retlen <= blocksize/2; // pick an order so that ret[] can reuse floors[] without stepping on any data we need
for(j=dir?0:vc->audio_channels-1; (unsigned)j<vc->audio_channels; j+=dir*2-1) {
uint_fast16_t bs0=vc->blocksize[0];
uint_fast16_t bs1=vc->blocksize[1];
float *residue=vc->channel_residues+res_chan[j]*blocksize/2;
float *floor=vc->channel_floors+j*blocksize/2;
float *saved=vc->saved+j*bs1/4;
float *ret=vc->channel_residues+j*retlen;
float *buf=floor;
float *ret=vc->channel_floors+j*retlen;
float *buf=residue;
const float *win=vc->win[blockflag&previous_window];
vc->mdct[0].fft.imdct_half(&vc->mdct[blockflag], buf, floor, residue);
vc->mdct[0].fft.imdct_half(&vc->mdct[blockflag], buf, floor);
if(blockflag == previous_window) {
vc->dsp.vector_fmul_window(ret, saved, buf, win, fadd_bias, blocksize/4);
......@@ -1583,7 +1583,7 @@ static int vorbis_decode_frame(AVCodecContext *avccontext,
AV_DEBUG("parsed %d bytes %d bits, returned %d samples (*ch*bits) \n", get_bits_count(gb)/8, get_bits_count(gb)%8, len);
for(i=0; i<vc->audio_channels; i++)
channel_ptrs[i] = vc->channel_residues+i*len;
channel_ptrs[i] = vc->channel_floors+i*len;
vc->dsp.float_to_int16_interleave(data, channel_ptrs, len, vc->audio_channels);
*data_size=len*2*vc->audio_channels;
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment