wrapper/FFmpeg
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
FFmpeg/libavcodec/x86/fft_sse.c
Vitor Sessak 4dcc4f8eaa SSE optimized 32-point DCT 
Originally committed as revision 24077 to svn://svn.ffmpeg.org/ffmpeg/trunk
2010-07-06 16:58:54 +00:00

469 lines
17 KiB
C
Raw Blame History

/*
* FFT/MDCT transform with SSE optimizations
* Copyright (c) 2008 Loren Merritt
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 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
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser 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 "libavutil/x86_cpu.h"
#include "libavutil/common.h"
#include "libavcodec/dsputil.h"
#include "fft.h"
DECLARE_ALIGNED(16, static const int, m1m1m1m1)[4] =
{ 1 << 31, 1 << 31, 1 << 31, 1 << 31 };
void ff_fft_dispatch_sse(FFTComplex *z, int nbits);
void ff_fft_dispatch_interleave_sse(FFTComplex *z, int nbits);
void ff_fft_calc_sse(FFTContext *s, FFTComplex *z)
{
int n = 1 << s->nbits;
ff_fft_dispatch_interleave_sse(z, s->nbits);
if(n <= 16) {
x86_reg i = -8*n;
__asm__ volatile(
"1: \n"
"movaps (%0,%1), %%xmm0 \n"
"movaps %%xmm0, %%xmm1 \n"
"unpcklps 16(%0,%1), %%xmm0 \n"
"unpckhps 16(%0,%1), %%xmm1 \n"
"movaps %%xmm0, (%0,%1) \n"
"movaps %%xmm1, 16(%0,%1) \n"
"add $32, %0 \n"
"jl 1b \n"
:"+r"(i)
:"r"(z+n)
:"memory"
);
}
}
void ff_fft_permute_sse(FFTContext *s, FFTComplex *z)
{
int n = 1 << s->nbits;
int i;
for(i=0; i<n; i+=2) {
__asm__ volatile(
"movaps %2, %%xmm0 \n"
"movlps %%xmm0, %0 \n"
"movhps %%xmm0, %1 \n"
:"=m"(s->tmp_buf[s->revtab[i]]),
"=m"(s->tmp_buf[s->revtab[i+1]])
:"m"(z[i])
);
}
memcpy(z, s->tmp_buf, n*sizeof(FFTComplex));
}
void ff_imdct_half_sse(FFTContext *s, FFTSample *output, const FFTSample *input)
{
av_unused x86_reg i, j, k, l;
long n = 1 << s->mdct_bits;
long n2 = n >> 1;
long n4 = n >> 2;
long n8 = n >> 3;
const uint16_t *revtab = s->revtab + n8;
const FFTSample *tcos = s->tcos;
const FFTSample *tsin = s->tsin;
FFTComplex *z = (FFTComplex *)output;
/* pre rotation */
for(k=n8-2; k>=0; k-=2) {
__asm__ volatile(
"movaps (%2,%1,2), %%xmm0 \n" // { z[k].re, z[k].im, z[k+1].re, z[k+1].im }
"movaps -16(%2,%0,2), %%xmm1 \n" // { z[-k-2].re, z[-k-2].im, z[-k-1].re, z[-k-1].im }
"movaps %%xmm0, %%xmm2 \n"
"shufps $0x88, %%xmm1, %%xmm0 \n" // { z[k].re, z[k+1].re, z[-k-2].re, z[-k-1].re }
"shufps $0x77, %%xmm2, %%xmm1 \n" // { z[-k-1].im, z[-k-2].im, z[k+1].im, z[k].im }
"movlps (%3,%1), %%xmm4 \n"
"movlps (%4,%1), %%xmm5 \n"
"movhps -8(%3,%0), %%xmm4 \n" // { cos[k], cos[k+1], cos[-k-2], cos[-k-1] }
"movhps -8(%4,%0), %%xmm5 \n" // { sin[k], sin[k+1], sin[-k-2], sin[-k-1] }
"movaps %%xmm0, %%xmm2 \n"
"movaps %%xmm1, %%xmm3 \n"
"mulps %%xmm5, %%xmm0 \n" // re*sin
"mulps %%xmm4, %%xmm1 \n" // im*cos
"mulps %%xmm4, %%xmm2 \n" // re*cos
"mulps %%xmm5, %%xmm3 \n" // im*sin
"subps %%xmm0, %%xmm1 \n" // -> re
"addps %%xmm3, %%xmm2 \n" // -> im
"movaps %%xmm1, %%xmm0 \n"
"unpcklps %%xmm2, %%xmm1 \n" // { z[k], z[k+1] }
"unpckhps %%xmm2, %%xmm0 \n" // { z[-k-2], z[-k-1] }
::"r"(-4*k), "r"(4*k),
"r"(input+n4), "r"(tcos+n8), "r"(tsin+n8)
);
#if ARCH_X86_64
// if we have enough regs, don't let gcc make the luts latency-bound
// but if not, latency is faster than spilling
__asm__("movlps %%xmm0, %0 \n"
"movhps %%xmm0, %1 \n"
"movlps %%xmm1, %2 \n"
"movhps %%xmm1, %3 \n"
:"=m"(z[revtab[-k-2]]),
"=m"(z[revtab[-k-1]]),
"=m"(z[revtab[ k ]]),
"=m"(z[revtab[ k+1]])
);
#else
__asm__("movlps %%xmm0, %0" :"=m"(z[revtab[-k-2]]));
__asm__("movhps %%xmm0, %0" :"=m"(z[revtab[-k-1]]));
__asm__("movlps %%xmm1, %0" :"=m"(z[revtab[ k ]]));
__asm__("movhps %%xmm1, %0" :"=m"(z[revtab[ k+1]]));
#endif
}
ff_fft_dispatch_sse(z, s->nbits);
/* post rotation + reinterleave + reorder */
#define CMUL(j,xmm0,xmm1)\
"movaps (%2,"#j",2), %%xmm6 \n"\
"movaps 16(%2,"#j",2), "#xmm0"\n"\
"movaps %%xmm6, "#xmm1"\n"\
"movaps "#xmm0",%%xmm7 \n"\
"mulps (%3,"#j"), %%xmm6 \n"\
"mulps (%4,"#j"), "#xmm0"\n"\
"mulps (%4,"#j"), "#xmm1"\n"\
"mulps (%3,"#j"), %%xmm7 \n"\
"subps %%xmm6, "#xmm0"\n"\
"addps %%xmm7, "#xmm1"\n"
j = -n2;
k = n2-16;
__asm__ volatile(
"1: \n"
CMUL(%0, %%xmm0, %%xmm1)
CMUL(%1, %%xmm4, %%xmm5)
"shufps $0x1b, %%xmm1, %%xmm1 \n"
"shufps $0x1b, %%xmm5, %%xmm5 \n"
"movaps %%xmm4, %%xmm6 \n"
"unpckhps %%xmm1, %%xmm4 \n"
"unpcklps %%xmm1, %%xmm6 \n"
"movaps %%xmm0, %%xmm2 \n"
"unpcklps %%xmm5, %%xmm0 \n"
"unpckhps %%xmm5, %%xmm2 \n"
"movaps %%xmm6, (%2,%1,2) \n"
"movaps %%xmm4, 16(%2,%1,2) \n"
"movaps %%xmm0, (%2,%0,2) \n"
"movaps %%xmm2, 16(%2,%0,2) \n"
"sub $16, %1 \n"
"add $16, %0 \n"
"jl 1b \n"
:"+&r"(j), "+&r"(k)
:"r"(z+n8), "r"(tcos+n8), "r"(tsin+n8)
:"memory"
);
}
void ff_imdct_calc_sse(FFTContext *s, FFTSample *output, const FFTSample *input)
{
x86_reg j, k;
long n = 1 << s->mdct_bits;
long n4 = n >> 2;
ff_imdct_half_sse(s, output+n4, input);
j = -n;
k = n-16;
__asm__ volatile(
"movaps %4, %%xmm7 \n"
"1: \n"
"movaps (%2,%1), %%xmm0 \n"
"movaps (%3,%0), %%xmm1 \n"
"shufps $0x1b, %%xmm0, %%xmm0 \n"
"shufps $0x1b, %%xmm1, %%xmm1 \n"
"xorps %%xmm7, %%xmm0 \n"
"movaps %%xmm1, (%3,%1) \n"
"movaps %%xmm0, (%2,%0) \n"
"sub $16, %1 \n"
"add $16, %0 \n"
"jl 1b \n"
:"+r"(j), "+r"(k)
:"r"(output+n4), "r"(output+n4*3),
"m"(*m1m1m1m1)
);
}
DECLARE_ALIGNED(16, static const float, b1)[] = {
0.500603, 0.505471, 0.515447, 0.531043,
0.553104, 0.582935, 0.622504, 0.674808,
-1.169440, -0.972568, -0.839350, -0.744536,
-10.190008, -3.407609, -2.057781, -1.484165,
0.502419, 0.522499, 0.566944, 0.646822,
0.788155, 1.060678, 1.722447, 5.101149,
0.509796, 0.601345, 0.899976, 2.562916,
1.000000, 1.000000, 1.306563, 0.541196,
1.000000, 0.707107, 1.000000, -0.707107
};
DECLARE_ALIGNED(16, static const int32_t, smask)[4] = {
0, 0, 0x80000000, 0x80000000
};
/* butterfly operator */
#define BUTTERFLY(a,b,c,tmp) \
"movaps %%" #a ", %%" #tmp " \n\t" \
"subps %%" #b ", %%" #a " \n\t" \
"addps %%" #tmp ", %%" #b " \n\t" \
"mulps " #c ", %%" #a " \n\t"
///* Same as BUTTERFLY when vectors a and b overlap */
#define BUTTERFLY0(val, mask, cos, tmp, shuf) \
"movaps %%" #val ", %%" #tmp " \n\t" \
"shufps " #shuf ", %%" #val ",%%" #val " \n\t" \
"xorps %%" #mask ", %%" #tmp " \n\t" /* flip signs */ \
"addps %%" #tmp ", %%" #val " \n\t" \
"mulps %%" #cos ", %%" #val " \n\t"
#define BUTTERFLY2(val, mask, cos, tmp) BUTTERFLY0(val, mask, cos, tmp, $0x1b)
#define BUTTERFLY3(val, mask, cos, tmp) BUTTERFLY0(val, mask, cos, tmp, $0xb1)
void ff_dct32_float_sse(FFTSample *out, const FFTSample *in)
{
int32_t tmp1 = 0;
__asm__ volatile(
/* pass 1 */
"movaps (%4), %%xmm0 \n\t"
"movaps 112(%4), %%xmm1 \n\t"
"shufps $0x1b, %%xmm1, %%xmm1 \n\t"
BUTTERFLY(xmm0, xmm1, (%2), xmm3)
"movaps 64(%4), %%xmm7 \n\t"
"movaps 48(%4), %%xmm4 \n\t"
"shufps $0x1b, %%xmm4, %%xmm4 \n\t"
BUTTERFLY(xmm7, xmm4, 48(%2), xmm3)
/* pass 2 */
"movaps 64(%2), %%xmm2 \n\t"
BUTTERFLY(xmm1, xmm4, %%xmm2, xmm3)
"movaps %%xmm1, 48(%1) \n\t"
"movaps %%xmm4, (%1) \n\t"
/* pass 1 */
"movaps 16(%4), %%xmm1 \n\t"
"movaps 96(%4), %%xmm6 \n\t"
"shufps $0x1b, %%xmm6, %%xmm6 \n\t"
BUTTERFLY(xmm1, xmm6, 16(%2), xmm3)
"movaps 80(%4), %%xmm4 \n\t"
"movaps 32(%4), %%xmm5 \n\t"
"shufps $0x1b, %%xmm5, %%xmm5 \n\t"
BUTTERFLY(xmm4, xmm5, 32(%2), xmm3)
/* pass 2 */
BUTTERFLY(xmm0, xmm7, %%xmm2, xmm3)
"movaps 80(%2), %%xmm2 \n\t"
BUTTERFLY(xmm6, xmm5, %%xmm2, xmm3)
BUTTERFLY(xmm1, xmm4, %%xmm2, xmm3)
/* pass 3 */
"movaps 96(%2), %%xmm2 \n\t"
"shufps $0x1b, %%xmm1, %%xmm1 \n\t"
BUTTERFLY(xmm0, xmm1, %%xmm2, xmm3)
"movaps %%xmm0, 112(%1) \n\t"
"movaps %%xmm1, 96(%1) \n\t"
"movaps 0(%1), %%xmm0 \n\t"
"shufps $0x1b, %%xmm5, %%xmm5 \n\t"
BUTTERFLY(xmm0, xmm5, %%xmm2, xmm3)
"movaps 48(%1), %%xmm1 \n\t"
"shufps $0x1b, %%xmm6, %%xmm6 \n\t"
BUTTERFLY(xmm1, xmm6, %%xmm2, xmm3)
"movaps %%xmm1, 48(%1) \n\t"
"shufps $0x1b, %%xmm4, %%xmm4 \n\t"
BUTTERFLY(xmm7, xmm4, %%xmm2, xmm3)
/* pass 4 */
"movaps (%3), %%xmm3 \n\t"
"movaps 112(%2), %%xmm2 \n\t"
BUTTERFLY2(xmm5, xmm3, xmm2, xmm1)
BUTTERFLY2(xmm0, xmm3, xmm2, xmm1)
"movaps %%xmm0, 16(%1) \n\t"
BUTTERFLY2(xmm6, xmm3, xmm2, xmm1)
"movaps %%xmm6, 32(%1) \n\t"
"movaps 48(%1), %%xmm0 \n\t"
BUTTERFLY2(xmm0, xmm3, xmm2, xmm1)
"movaps %%xmm0, 48(%1) \n\t"
BUTTERFLY2(xmm4, xmm3, xmm2, xmm1)
BUTTERFLY2(xmm7, xmm3, xmm2, xmm1)
"movaps 96(%1), %%xmm6 \n\t"
BUTTERFLY2(xmm6, xmm3, xmm2, xmm1)
"movaps 112(%1), %%xmm0 \n\t"
BUTTERFLY2(xmm0, xmm3, xmm2, xmm1)
/* pass 5 */
"movaps 128(%2), %%xmm2 \n\t"
"shufps $0xCC, %%xmm3,%%xmm3 \n\t"
BUTTERFLY3(xmm5, xmm3, xmm2, xmm1)
"movaps %%xmm5, (%1) \n\t"
"movaps 16(%1), %%xmm1 \n\t"
BUTTERFLY3(xmm1, xmm3, xmm2, xmm5)
"movaps %%xmm1, 16(%1) \n\t"
BUTTERFLY3(xmm4, xmm3, xmm2, xmm5)
"movaps %%xmm4, 64(%1) \n\t"
BUTTERFLY3(xmm7, xmm3, xmm2, xmm5)
"movaps %%xmm7, 80(%1) \n\t"
"movaps 32(%1), %%xmm5 \n\t"
BUTTERFLY3(xmm5, xmm3, xmm2, xmm7)
"movaps %%xmm5, 32(%1) \n\t"
"movaps 48(%1), %%xmm4 \n\t"
BUTTERFLY3(xmm4, xmm3, xmm2, xmm7)
"movaps %%xmm4, 48(%1) \n\t"
BUTTERFLY3(xmm6, xmm3, xmm2, xmm7)
"movaps %%xmm6, 96(%1) \n\t"
BUTTERFLY3(xmm0, xmm3, xmm2, xmm7)
"movaps %%xmm0, 112(%1) \n\t"
/* pass 6, no SIMD... */
"movss 56(%1), %%xmm3 \n\t"
"movl 4(%1), %0 \n\t"
"addss 60(%1), %%xmm3 \n\t"
"movss 72(%1), %%xmm7 \n\t"
"addss %%xmm3, %%xmm4 \n\t"
"movss 52(%1), %%xmm2 \n\t"
"addss %%xmm3, %%xmm2 \n\t"
"movss 24(%1), %%xmm3 \n\t"
"addss 28(%1), %%xmm3 \n\t"
"addss 76(%1), %%xmm7 \n\t"
"addss %%xmm3, %%xmm1 \n\t"
"addss %%xmm4, %%xmm5 \n\t"
"movss %%xmm1, 16(%1) \n\t"
"movss 20(%1), %%xmm1 \n\t"
"addss %%xmm3, %%xmm1 \n\t"
"movss 40(%1), %%xmm3 \n\t"
"movss %%xmm1, 48(%1) \n\t"
"addss 44(%1), %%xmm3 \n\t"
"movss 20(%1), %%xmm1 \n\t"
"addss %%xmm3, %%xmm4 \n\t"
"addss %%xmm2, %%xmm3 \n\t"
"addss 28(%1), %%xmm1 \n\t"
"movss %%xmm3, 40(%1) \n\t"
"addss 36(%1), %%xmm2 \n\t"
"movss 8(%1), %%xmm3 \n\t"
"movss %%xmm2, 56(%1) \n\t"
"addss 12(%1), %%xmm3 \n\t"
"movss %%xmm5, 8(%1) \n\t"
"movss %%xmm3, 32(%1) \n\t"
"movss 52(%1), %%xmm2 \n\t"
"movss 80(%1), %%xmm3 \n\t"
"movss 120(%1), %%xmm5 \n\t"
"movss %%xmm1, 80(%1) \n\t"
"movss %%xmm4, 24(%1) \n\t"
"addss 124(%1), %%xmm5 \n\t"
"movss 64(%1), %%xmm1 \n\t"
"addss 60(%1), %%xmm2 \n\t"
"addss %%xmm5, %%xmm0 \n\t"
"addss 116(%1), %%xmm5 \n\t"
"movl %0, 64(%1) \n\t"
"addss %%xmm0, %%xmm6 \n\t"
"addss %%xmm6, %%xmm1 \n\t"
"movl 12(%1), %0 \n\t"
"movss %%xmm1, 4(%1) \n\t"
"movss 88(%1), %%xmm1 \n\t"
"movl %0, 96(%1) \n\t"
"addss 92(%1), %%xmm1 \n\t"
"movss 104(%1), %%xmm4 \n\t"
"movl 28(%1), %0 \n\t"
"addss 108(%1), %%xmm4 \n\t"
"addss %%xmm4, %%xmm0 \n\t"
"addss %%xmm1, %%xmm3 \n\t"
"addss 84(%1), %%xmm1 \n\t"
"addss %%xmm5, %%xmm4 \n\t"
"addss %%xmm3, %%xmm6 \n\t"
"addss %%xmm0, %%xmm3 \n\t"
"addss %%xmm7, %%xmm0 \n\t"
"addss 100(%1), %%xmm5 \n\t"
"addss %%xmm4, %%xmm7 \n\t"
"movl %0, 112(%1) \n\t"
"movss %%xmm0, 28(%1) \n\t"
"movss 36(%1), %%xmm0 \n\t"
"movss %%xmm7, 36(%1) \n\t"
"addss %%xmm1, %%xmm4 \n\t"
"movss 116(%1), %%xmm7 \n\t"
"addss %%xmm2, %%xmm0 \n\t"
"addss 124(%1), %%xmm7 \n\t"
"movss %%xmm0, 72(%1) \n\t"
"movss 44(%1), %%xmm0 \n\t"
"movss %%xmm6, 12(%1) \n\t"
"movss %%xmm3, 20(%1) \n\t"
"addss %%xmm0, %%xmm2 \n\t"
"movss %%xmm4, 44(%1) \n\t"
"movss %%xmm2, 88(%1) \n\t"
"addss 60(%1), %%xmm0 \n\t"
"movl 60(%1), %0 \n\t"
"movl %0, 120(%1) \n\t"
"movss %%xmm0, 104(%1) \n\t"
"addss %%xmm5, %%xmm1 \n\t"
"addss 68(%1), %%xmm5 \n\t"
"movss %%xmm1, 52(%1) \n\t"
"movss %%xmm5, 60(%1) \n\t"
"movss 68(%1), %%xmm1 \n\t"
"movss 100(%1), %%xmm5 \n\t"
"addss %%xmm7, %%xmm5 \n\t"
"addss 108(%1), %%xmm7 \n\t"
"addss %%xmm5, %%xmm1 \n\t"
"movss 84(%1), %%xmm2 \n\t"
"addss 92(%1), %%xmm2 \n\t"
"addss %%xmm2, %%xmm5 \n\t"
"movss %%xmm1, 68(%1) \n\t"
"addss %%xmm7, %%xmm2 \n\t"
"movss 76(%1), %%xmm1 \n\t"
"movss %%xmm2, 84(%1) \n\t"
"movss %%xmm5, 76(%1) \n\t"
"movss 108(%1), %%xmm2 \n\t"
"addss %%xmm1, %%xmm7 \n\t"
"addss 124(%1), %%xmm2 \n\t"
"addss %%xmm2, %%xmm1 \n\t"
"addss 92(%1), %%xmm2 \n\t"
"movss %%xmm1, 100(%1) \n\t"
"movss %%xmm2, 108(%1) \n\t"
"movss 92(%1), %%xmm2 \n\t"
"movss %%xmm7, 92(%1) \n\t"
"addss 124(%1), %%xmm2 \n\t"
"movss %%xmm2, 116(%1) \n\t"
:"+&r"(tmp1)
:"r"(out), "r"(b1), "r"(smask), "r"(in)
:"memory"
);
}
Reference in a new issue View git blame Copy permalink