Chris@69: /* Copyright (c) 2014, Cisco Systems, INC
Chris@69:    Written by XiangMingZhu WeiZhou MinPeng YanWang
Chris@69: 
Chris@69:    Redistribution and use in source and binary forms, with or without
Chris@69:    modification, are permitted provided that the following conditions
Chris@69:    are met:
Chris@69: 
Chris@69:    - Redistributions of source code must retain the above copyright
Chris@69:    notice, this list of conditions and the following disclaimer.
Chris@69: 
Chris@69:    - Redistributions in binary form must reproduce the above copyright
Chris@69:    notice, this list of conditions and the following disclaimer in the
Chris@69:    documentation and/or other materials provided with the distribution.
Chris@69: 
Chris@69:    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
Chris@69:    ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
Chris@69:    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
Chris@69:    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
Chris@69:    OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
Chris@69:    EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
Chris@69:    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
Chris@69:    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
Chris@69:    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
Chris@69:    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
Chris@69:    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
Chris@69: */
Chris@69: 
Chris@69: #ifdef HAVE_CONFIG_H
Chris@69: #include "config.h"
Chris@69: #endif
Chris@69: 
Chris@69: #include "macros.h"
Chris@69: #include "celt_lpc.h"
Chris@69: #include "stack_alloc.h"
Chris@69: #include "mathops.h"
Chris@69: #include "pitch.h"
Chris@69: 
Chris@69: #if defined(OPUS_X86_MAY_HAVE_SSE) && !defined(FIXED_POINT)
Chris@69: 
Chris@69: #include <xmmintrin.h>
Chris@69: #include "arch.h"
Chris@69: 
Chris@69: void xcorr_kernel_sse(const opus_val16 *x, const opus_val16 *y, opus_val32 sum[4], int len)
Chris@69: {
Chris@69:    int j;
Chris@69:    __m128 xsum1, xsum2;
Chris@69:    xsum1 = _mm_loadu_ps(sum);
Chris@69:    xsum2 = _mm_setzero_ps();
Chris@69: 
Chris@69:    for (j = 0; j < len-3; j += 4)
Chris@69:    {
Chris@69:       __m128 x0 = _mm_loadu_ps(x+j);
Chris@69:       __m128 yj = _mm_loadu_ps(y+j);
Chris@69:       __m128 y3 = _mm_loadu_ps(y+j+3);
Chris@69: 
Chris@69:       xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0x00),yj));
Chris@69:       xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0x55),
Chris@69:                                           _mm_shuffle_ps(yj,y3,0x49)));
Chris@69:       xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0xaa),
Chris@69:                                           _mm_shuffle_ps(yj,y3,0x9e)));
Chris@69:       xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0xff),y3));
Chris@69:    }
Chris@69:    if (j < len)
Chris@69:    {
Chris@69:       xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_load1_ps(x+j),_mm_loadu_ps(y+j)));
Chris@69:       if (++j < len)
Chris@69:       {
Chris@69:          xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(_mm_load1_ps(x+j),_mm_loadu_ps(y+j)));
Chris@69:          if (++j < len)
Chris@69:          {
Chris@69:             xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_load1_ps(x+j),_mm_loadu_ps(y+j)));
Chris@69:          }
Chris@69:       }
Chris@69:    }
Chris@69:    _mm_storeu_ps(sum,_mm_add_ps(xsum1,xsum2));
Chris@69: }
Chris@69: 
Chris@69: 
Chris@69: void dual_inner_prod_sse(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02,
Chris@69:       int N, opus_val32 *xy1, opus_val32 *xy2)
Chris@69: {
Chris@69:    int i;
Chris@69:    __m128 xsum1, xsum2;
Chris@69:    xsum1 = _mm_setzero_ps();
Chris@69:    xsum2 = _mm_setzero_ps();
Chris@69:    for (i=0;i<N-3;i+=4)
Chris@69:    {
Chris@69:       __m128 xi = _mm_loadu_ps(x+i);
Chris@69:       __m128 y1i = _mm_loadu_ps(y01+i);
Chris@69:       __m128 y2i = _mm_loadu_ps(y02+i);
Chris@69:       xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(xi, y1i));
Chris@69:       xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(xi, y2i));
Chris@69:    }
Chris@69:    /* Horizontal sum */
Chris@69:    xsum1 = _mm_add_ps(xsum1, _mm_movehl_ps(xsum1, xsum1));
Chris@69:    xsum1 = _mm_add_ss(xsum1, _mm_shuffle_ps(xsum1, xsum1, 0x55));
Chris@69:    _mm_store_ss(xy1, xsum1);
Chris@69:    xsum2 = _mm_add_ps(xsum2, _mm_movehl_ps(xsum2, xsum2));
Chris@69:    xsum2 = _mm_add_ss(xsum2, _mm_shuffle_ps(xsum2, xsum2, 0x55));
Chris@69:    _mm_store_ss(xy2, xsum2);
Chris@69:    for (;i<N;i++)
Chris@69:    {
Chris@69:       *xy1 = MAC16_16(*xy1, x[i], y01[i]);
Chris@69:       *xy2 = MAC16_16(*xy2, x[i], y02[i]);
Chris@69:    }
Chris@69: }
Chris@69: 
Chris@69: opus_val32 celt_inner_prod_sse(const opus_val16 *x, const opus_val16 *y,
Chris@69:       int N)
Chris@69: {
Chris@69:    int i;
Chris@69:    float xy;
Chris@69:    __m128 sum;
Chris@69:    sum = _mm_setzero_ps();
Chris@69:    /* FIXME: We should probably go 8-way and use 2 sums. */
Chris@69:    for (i=0;i<N-3;i+=4)
Chris@69:    {
Chris@69:       __m128 xi = _mm_loadu_ps(x+i);
Chris@69:       __m128 yi = _mm_loadu_ps(y+i);
Chris@69:       sum = _mm_add_ps(sum,_mm_mul_ps(xi, yi));
Chris@69:    }
Chris@69:    /* Horizontal sum */
Chris@69:    sum = _mm_add_ps(sum, _mm_movehl_ps(sum, sum));
Chris@69:    sum = _mm_add_ss(sum, _mm_shuffle_ps(sum, sum, 0x55));
Chris@69:    _mm_store_ss(&xy, sum);
Chris@69:    for (;i<N;i++)
Chris@69:    {
Chris@69:       xy = MAC16_16(xy, x[i], y[i]);
Chris@69:    }
Chris@69:    return xy;
Chris@69: }
Chris@69: 
Chris@69: void comb_filter_const_sse(opus_val32 *y, opus_val32 *x, int T, int N,
Chris@69:       opus_val16 g10, opus_val16 g11, opus_val16 g12)
Chris@69: {
Chris@69:    int i;
Chris@69:    __m128 x0v;
Chris@69:    __m128 g10v, g11v, g12v;
Chris@69:    g10v = _mm_load1_ps(&g10);
Chris@69:    g11v = _mm_load1_ps(&g11);
Chris@69:    g12v = _mm_load1_ps(&g12);
Chris@69:    x0v = _mm_loadu_ps(&x[-T-2]);
Chris@69:    for (i=0;i<N-3;i+=4)
Chris@69:    {
Chris@69:       __m128 yi, yi2, x1v, x2v, x3v, x4v;
Chris@69:       const opus_val32 *xp = &x[i-T-2];
Chris@69:       yi = _mm_loadu_ps(x+i);
Chris@69:       x4v = _mm_loadu_ps(xp+4);
Chris@69: #if 0
Chris@69:       /* Slower version with all loads */
Chris@69:       x1v = _mm_loadu_ps(xp+1);
Chris@69:       x2v = _mm_loadu_ps(xp+2);
Chris@69:       x3v = _mm_loadu_ps(xp+3);
Chris@69: #else
Chris@69:       x2v = _mm_shuffle_ps(x0v, x4v, 0x4e);
Chris@69:       x1v = _mm_shuffle_ps(x0v, x2v, 0x99);
Chris@69:       x3v = _mm_shuffle_ps(x2v, x4v, 0x99);
Chris@69: #endif
Chris@69: 
Chris@69:       yi = _mm_add_ps(yi, _mm_mul_ps(g10v,x2v));
Chris@69: #if 0 /* Set to 1 to make it bit-exact with the non-SSE version */
Chris@69:       yi = _mm_add_ps(yi, _mm_mul_ps(g11v,_mm_add_ps(x3v,x1v)));
Chris@69:       yi = _mm_add_ps(yi, _mm_mul_ps(g12v,_mm_add_ps(x4v,x0v)));
Chris@69: #else
Chris@69:       /* Use partial sums */
Chris@69:       yi2 = _mm_add_ps(_mm_mul_ps(g11v,_mm_add_ps(x3v,x1v)),
Chris@69:                        _mm_mul_ps(g12v,_mm_add_ps(x4v,x0v)));
Chris@69:       yi = _mm_add_ps(yi, yi2);
Chris@69: #endif
Chris@69:       x0v=x4v;
Chris@69:       _mm_storeu_ps(y+i, yi);
Chris@69:    }
Chris@69: #ifdef CUSTOM_MODES
Chris@69:    for (;i<N;i++)
Chris@69:    {
Chris@69:       y[i] = x[i]
Chris@69:                + MULT16_32_Q15(g10,x[i-T])
Chris@69:                + MULT16_32_Q15(g11,ADD32(x[i-T+1],x[i-T-1]))
Chris@69:                + MULT16_32_Q15(g12,ADD32(x[i-T+2],x[i-T-2]));
Chris@69:    }
Chris@69: #endif
Chris@69: }
Chris@69: 
Chris@69: 
Chris@69: #endif