46ae64cd60
This reverts commit e00426c512.
The way we handle platform-specific intrinsics is not good, so the
current state will not compile on armv8. This commit also requires
SSE4.1 support, which is likely not a good idea for portable binaries.
We'll have to redo this with more caution after 3.2 is released, or
we might simply drop opus as we're only using it as dependency for
theora right now.
Fixes #33606.
258 lines
7.9 KiB
C
258 lines
7.9 KiB
C
/* Copyright (c) 2015 Xiph.Org Foundation
|
|
Written by Viswanath Puttagunta */
|
|
/**
|
|
@file celt_ne10_mdct.c
|
|
@brief ARM Neon optimizations for mdct using NE10 library
|
|
*/
|
|
|
|
/*
|
|
Redistribution and use in source and binary forms, with or without
|
|
modification, are permitted provided that the following conditions
|
|
are met:
|
|
|
|
- Redistributions of source code must retain the above copyright
|
|
notice, this list of conditions and the following disclaimer.
|
|
|
|
- Redistributions in binary form must reproduce the above copyright
|
|
notice, this list of conditions and the following disclaimer in the
|
|
documentation and/or other materials provided with the distribution.
|
|
|
|
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
|
|
``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
|
|
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
|
|
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
|
|
OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
|
EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
|
PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
|
PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
|
|
LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
|
|
NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
|
|
SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
|
*/
|
|
|
|
#ifndef SKIP_CONFIG_H
|
|
#ifdef HAVE_CONFIG_H
|
|
#include "config.h"
|
|
#endif
|
|
#endif
|
|
|
|
#include "kiss_fft.h"
|
|
#include "_kiss_fft_guts.h"
|
|
#include "mdct.h"
|
|
#include "stack_alloc.h"
|
|
|
|
void clt_mdct_forward_neon(const mdct_lookup *l,
|
|
kiss_fft_scalar *in,
|
|
kiss_fft_scalar * OPUS_RESTRICT out,
|
|
const opus_val16 *window,
|
|
int overlap, int shift, int stride, int arch)
|
|
{
|
|
int i;
|
|
int N, N2, N4;
|
|
VARDECL(kiss_fft_scalar, f);
|
|
VARDECL(kiss_fft_cpx, f2);
|
|
const kiss_fft_state *st = l->kfft[shift];
|
|
const kiss_twiddle_scalar *trig;
|
|
|
|
SAVE_STACK;
|
|
|
|
N = l->n;
|
|
trig = l->trig;
|
|
for (i=0;i<shift;i++)
|
|
{
|
|
N >>= 1;
|
|
trig += N;
|
|
}
|
|
N2 = N>>1;
|
|
N4 = N>>2;
|
|
|
|
ALLOC(f, N2, kiss_fft_scalar);
|
|
ALLOC(f2, N4, kiss_fft_cpx);
|
|
|
|
/* Consider the input to be composed of four blocks: [a, b, c, d] */
|
|
/* Window, shuffle, fold */
|
|
{
|
|
/* Temp pointers to make it really clear to the compiler what we're doing */
|
|
const kiss_fft_scalar * OPUS_RESTRICT xp1 = in+(overlap>>1);
|
|
const kiss_fft_scalar * OPUS_RESTRICT xp2 = in+N2-1+(overlap>>1);
|
|
kiss_fft_scalar * OPUS_RESTRICT yp = f;
|
|
const opus_val16 * OPUS_RESTRICT wp1 = window+(overlap>>1);
|
|
const opus_val16 * OPUS_RESTRICT wp2 = window+(overlap>>1)-1;
|
|
for(i=0;i<((overlap+3)>>2);i++)
|
|
{
|
|
/* Real part arranged as -d-cR, Imag part arranged as -b+aR*/
|
|
*yp++ = MULT16_32_Q15(*wp2, xp1[N2]) + MULT16_32_Q15(*wp1,*xp2);
|
|
*yp++ = MULT16_32_Q15(*wp1, *xp1) - MULT16_32_Q15(*wp2, xp2[-N2]);
|
|
xp1+=2;
|
|
xp2-=2;
|
|
wp1+=2;
|
|
wp2-=2;
|
|
}
|
|
wp1 = window;
|
|
wp2 = window+overlap-1;
|
|
for(;i<N4-((overlap+3)>>2);i++)
|
|
{
|
|
/* Real part arranged as a-bR, Imag part arranged as -c-dR */
|
|
*yp++ = *xp2;
|
|
*yp++ = *xp1;
|
|
xp1+=2;
|
|
xp2-=2;
|
|
}
|
|
for(;i<N4;i++)
|
|
{
|
|
/* Real part arranged as a-bR, Imag part arranged as -c-dR */
|
|
*yp++ = -MULT16_32_Q15(*wp1, xp1[-N2]) + MULT16_32_Q15(*wp2, *xp2);
|
|
*yp++ = MULT16_32_Q15(*wp2, *xp1) + MULT16_32_Q15(*wp1, xp2[N2]);
|
|
xp1+=2;
|
|
xp2-=2;
|
|
wp1+=2;
|
|
wp2-=2;
|
|
}
|
|
}
|
|
/* Pre-rotation */
|
|
{
|
|
kiss_fft_scalar * OPUS_RESTRICT yp = f;
|
|
const kiss_twiddle_scalar *t = &trig[0];
|
|
for(i=0;i<N4;i++)
|
|
{
|
|
kiss_fft_cpx yc;
|
|
kiss_twiddle_scalar t0, t1;
|
|
kiss_fft_scalar re, im, yr, yi;
|
|
t0 = t[i];
|
|
t1 = t[N4+i];
|
|
re = *yp++;
|
|
im = *yp++;
|
|
yr = S_MUL(re,t0) - S_MUL(im,t1);
|
|
yi = S_MUL(im,t0) + S_MUL(re,t1);
|
|
yc.r = yr;
|
|
yc.i = yi;
|
|
f2[i] = yc;
|
|
}
|
|
}
|
|
|
|
opus_fft(st, f2, (kiss_fft_cpx *)f, arch);
|
|
|
|
/* Post-rotate */
|
|
{
|
|
/* Temp pointers to make it really clear to the compiler what we're doing */
|
|
const kiss_fft_cpx * OPUS_RESTRICT fp = (kiss_fft_cpx *)f;
|
|
kiss_fft_scalar * OPUS_RESTRICT yp1 = out;
|
|
kiss_fft_scalar * OPUS_RESTRICT yp2 = out+stride*(N2-1);
|
|
const kiss_twiddle_scalar *t = &trig[0];
|
|
/* Temp pointers to make it really clear to the compiler what we're doing */
|
|
for(i=0;i<N4;i++)
|
|
{
|
|
kiss_fft_scalar yr, yi;
|
|
yr = S_MUL(fp->i,t[N4+i]) - S_MUL(fp->r,t[i]);
|
|
yi = S_MUL(fp->r,t[N4+i]) + S_MUL(fp->i,t[i]);
|
|
*yp1 = yr;
|
|
*yp2 = yi;
|
|
fp++;
|
|
yp1 += 2*stride;
|
|
yp2 -= 2*stride;
|
|
}
|
|
}
|
|
RESTORE_STACK;
|
|
}
|
|
|
|
void clt_mdct_backward_neon(const mdct_lookup *l,
|
|
kiss_fft_scalar *in,
|
|
kiss_fft_scalar * OPUS_RESTRICT out,
|
|
const opus_val16 * OPUS_RESTRICT window,
|
|
int overlap, int shift, int stride, int arch)
|
|
{
|
|
int i;
|
|
int N, N2, N4;
|
|
VARDECL(kiss_fft_scalar, f);
|
|
const kiss_twiddle_scalar *trig;
|
|
const kiss_fft_state *st = l->kfft[shift];
|
|
|
|
N = l->n;
|
|
trig = l->trig;
|
|
for (i=0;i<shift;i++)
|
|
{
|
|
N >>= 1;
|
|
trig += N;
|
|
}
|
|
N2 = N>>1;
|
|
N4 = N>>2;
|
|
|
|
ALLOC(f, N2, kiss_fft_scalar);
|
|
|
|
/* Pre-rotate */
|
|
{
|
|
/* Temp pointers to make it really clear to the compiler what we're doing */
|
|
const kiss_fft_scalar * OPUS_RESTRICT xp1 = in;
|
|
const kiss_fft_scalar * OPUS_RESTRICT xp2 = in+stride*(N2-1);
|
|
kiss_fft_scalar * OPUS_RESTRICT yp = f;
|
|
const kiss_twiddle_scalar * OPUS_RESTRICT t = &trig[0];
|
|
for(i=0;i<N4;i++)
|
|
{
|
|
kiss_fft_scalar yr, yi;
|
|
yr = S_MUL(*xp2, t[i]) + S_MUL(*xp1, t[N4+i]);
|
|
yi = S_MUL(*xp1, t[i]) - S_MUL(*xp2, t[N4+i]);
|
|
yp[2*i] = yr;
|
|
yp[2*i+1] = yi;
|
|
xp1+=2*stride;
|
|
xp2-=2*stride;
|
|
}
|
|
}
|
|
|
|
opus_ifft(st, (kiss_fft_cpx *)f, (kiss_fft_cpx*)(out+(overlap>>1)), arch);
|
|
|
|
/* Post-rotate and de-shuffle from both ends of the buffer at once to make
|
|
it in-place. */
|
|
{
|
|
kiss_fft_scalar * yp0 = out+(overlap>>1);
|
|
kiss_fft_scalar * yp1 = out+(overlap>>1)+N2-2;
|
|
const kiss_twiddle_scalar *t = &trig[0];
|
|
/* Loop to (N4+1)>>1 to handle odd N4. When N4 is odd, the
|
|
middle pair will be computed twice. */
|
|
for(i=0;i<(N4+1)>>1;i++)
|
|
{
|
|
kiss_fft_scalar re, im, yr, yi;
|
|
kiss_twiddle_scalar t0, t1;
|
|
re = yp0[0];
|
|
im = yp0[1];
|
|
t0 = t[i];
|
|
t1 = t[N4+i];
|
|
/* We'd scale up by 2 here, but instead it's done when mixing the windows */
|
|
yr = S_MUL(re,t0) + S_MUL(im,t1);
|
|
yi = S_MUL(re,t1) - S_MUL(im,t0);
|
|
re = yp1[0];
|
|
im = yp1[1];
|
|
yp0[0] = yr;
|
|
yp1[1] = yi;
|
|
|
|
t0 = t[(N4-i-1)];
|
|
t1 = t[(N2-i-1)];
|
|
/* We'd scale up by 2 here, but instead it's done when mixing the windows */
|
|
yr = S_MUL(re,t0) + S_MUL(im,t1);
|
|
yi = S_MUL(re,t1) - S_MUL(im,t0);
|
|
yp1[0] = yr;
|
|
yp0[1] = yi;
|
|
yp0 += 2;
|
|
yp1 -= 2;
|
|
}
|
|
}
|
|
|
|
/* Mirror on both sides for TDAC */
|
|
{
|
|
kiss_fft_scalar * OPUS_RESTRICT xp1 = out+overlap-1;
|
|
kiss_fft_scalar * OPUS_RESTRICT yp1 = out;
|
|
const opus_val16 * OPUS_RESTRICT wp1 = window;
|
|
const opus_val16 * OPUS_RESTRICT wp2 = window+overlap-1;
|
|
|
|
for(i = 0; i < overlap/2; i++)
|
|
{
|
|
kiss_fft_scalar x1, x2;
|
|
x1 = *xp1;
|
|
x2 = *yp1;
|
|
*yp1++ = MULT16_32_Q15(*wp2, x2) - MULT16_32_Q15(*wp1, x1);
|
|
*xp1-- = MULT16_32_Q15(*wp1, x2) + MULT16_32_Q15(*wp2, x1);
|
|
wp1++;
|
|
wp2--;
|
|
}
|
|
}
|
|
RESTORE_STACK;
|
|
}
|