/*
* SSE Convolution
* Copyright (C) 2012, 2013 Thomas Tsou <tom@tsou.cc>
*
* This library 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.
*
* This library 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 this library; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include <malloc.h>
#include <string.h>
#include <stdio.h>
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#ifdef HAVE_SSE3
#include <xmmintrin.h>
#include <pmmintrin.h>
/* 4-tap SSE complex-real convolution */
static void sse_conv_real4(float *restrict x,
float *restrict h,
float *restrict y,
int len)
{
__m128 m0, m1, m2, m3, m4, m5, m6, m7;
/* Load (aligned) filter taps */
m0 = _mm_load_ps(&h[0]);
m1 = _mm_load_ps(&h[4]);
m7 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(0, 2, 0, 2));
for (int i = 0; i < len; i++) {
/* Load (unaligned) input data */
m0 = _mm_loadu_ps(&x[2 * i + 0]);
m1 = _mm_loadu_ps(&x[2 * i + 4]);
m2 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(0, 2, 0, 2));
m3 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(1, 3, 1, 3));
/* Quad multiply */
m4 = _mm_mul_ps(m2, m7);
m5 = _mm_mul_ps(m3, m7);
/* Sum and store */
m6 = _mm_hadd_ps(m4, m5);
m0 = _mm_hadd_ps(m6, m6);
_mm_store_ss(&y[2 * i + 0], m0);
m0 = _mm_shuffle_ps(m0, m0, _MM_SHUFFLE(0, 3, 2, 1));
_mm_store_ss(&y[2 * i + 1], m0);
}
}
/* 8-tap SSE complex-real convolution */
static void sse_conv_real8(float *restrict x,
float *restrict h,
float *restrict y,
int len)
{
__m128 m0, m1, m2, m3, m4, m5, m6, m7, m8, m9;
/* Load (aligned) filter taps */
m0 = _mm_load_ps(&h[0]);
m1 = _mm_load_ps(&h[4]);
m2 = _mm_load_ps(&h[8]);
m3 = _mm_load_ps(&h[12]);
m4 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(0, 2, 0, 2));
m5 = _mm_shuffle_ps(m2, m3, _MM_SHUFFLE(0, 2, 0, 2));
for (int i = 0; i < len; i++) {
/* Load (unaligned) input data */
m0 = _mm_loadu_ps(&x[2 * i + 0]);
m1 = _mm_loadu_ps(&x[2 * i + 4]);
m2 = _mm_loadu_ps(&x[2 * i + 8]);
m3 = _mm_loadu_ps(&x[2 * i + 12]);
m6 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(0, 2, 0, 2));
m7 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(1, 3, 1, 3));
m8 = _mm_shuffle_ps(m2, m3, _MM_SHUFFLE(0, 2, 0, 2));
m9 = _mm_shuffle_ps(m2, m3, _MM_SHUFFLE(1, 3, 1, 3));
/* Quad multiply */
m6 = _mm_mul_ps(m6, m4);
m7 = _mm_mul_ps(m7, m4);
m8 = _mm_mul_ps(m8, m5);
m9 = _mm_mul_ps(m9, m5);
/* Sum and store */
m6 = _mm_add_ps(m6, m8);
m7 = _mm_add_ps(m7, m9);
m6 = _mm_hadd_ps(m6, m7);
m6 = _mm_hadd_ps(m6, m6);
_mm_store_ss(&y[2 * i + 0], m6);
m6 = _mm_shuffle_ps(m6, m6, _MM_SHUFFLE(0, 3, 2, 1));
_mm_store_ss(&y[2 * i + 1], m6);
}
}
/* 12-tap SSE complex-real convolution */
static void sse_conv_real12(float *restrict x,
float *restrict h,
float *restrict y,
int len)
{
__m128 m0, m1, m2, m3, m4, m5, m6, m7;
__m128 m8, m9, m10, m11, m12, m13, m14;
/* Load (aligned) filter taps */
m0 = _mm_load_ps(&h[0]);
m1 = _mm_load_ps(&h[4]);
m2 = _mm_load_ps(&h[8]);
m3 = _mm_load_ps(&h[12]);
m4 = _mm_load_ps(&h[16]);
m5 = _mm_load_ps(&h[20]);
m12 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(0, 2, 0, 2));
m13 = _mm_shuffle_ps(m2, m3, _MM_SHUFFLE(0, 2, 0, 2));
m14 = _mm_shuffle_ps(m4, m5, _MM_SHUFFLE(0, 2, 0, 2));
for (int i = 0; i < len; i++) {
/* Load (unaligned) input data */
m0 = _mm_loadu_ps(&x[2 * i + 0]);
m1 = _mm_loadu_ps(&x[2 * i + 4]);
m2 = _mm_loadu_ps(&x[2 * i + 8]);
m3 = _mm_loadu_ps(&x[2 * i + 12]);
m4 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(0, 2, 0, 2));
m5 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(1, 3, 1, 3));
m6 = _mm_shuffle_ps(m2, m3, _MM_SHUFFLE(0, 2, 0, 2));
m7 = _mm_shuffle_ps(m2, m3, _MM_SHUFFLE(1, 3, 1, 3));
m0 = _mm_loadu_ps(&x[2 * i + 16]);
m1 = _mm_loadu_ps(&x[2 * i + 20]);
m8 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(0, 2, 0, 2));
m9 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(1, 3, 1, 3));
/* Quad multiply */
m0 = _mm_mul_ps(m4, m12);
m1 = _mm_mul_ps(m5, m12);
m2 = _mm_mul_ps(m6, m13);
m3 = _mm_mul_ps(m7, m13);
m4 = _mm_mul_ps(m8, m14);
m5 = _mm_mul_ps(m9, m14);
/* Sum and store */
m8 = _mm_add_ps(m0, m2);
m9 = _mm_add_ps(m1, m3);
m10 = _mm_add_ps(m8, m4);
m11 = _mm_add_ps(m9, m5);
m2 = _mm_hadd_ps(m10, m11);
m3 = _mm_hadd_ps(m2, m2);
_mm_store_ss(&y[2 * i + 0], m3);
m3 = _mm_shuffle_ps(m3, m3, _MM_SHUFFLE(0, 3, 2, 1));
_mm_store_ss(&y[2 * i + 1], m3);
}
}
/* 16-tap SSE complex-real convolution */
static void sse_conv_real16(float *restrict x,
float *restrict h,
float *restrict y,
int len)
{
__m128 m0, m1, m2, m3, m4, m5, m6, m7;
__m128 m8, m9, m10, m11, m12, m13, m14, m15;
/* Load (aligned) filter taps */
m0 = _mm_load_ps(&h[0]);
m1 = _mm_load_ps(&h[4]);
m2 = _mm_load_ps(&h[8]);
m3 = _mm_load_ps(&h[12]);
m4 = _mm_load_ps(&h[16]);
m5 = _mm_load_ps(&h[20]);
m6 = _mm_load_ps(&h[24]);
m7 = _mm_load_ps(&h[28]);
m12 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(0, 2, 0, 2));
m13 = _mm_shuffle_ps(m2, m3, _MM_SHUFFLE(0, 2, 0, 2));
m14 = _mm_shuffle_ps(m4, m5, _MM_SHUFFLE(0, 2, 0, 2));
m15 = _mm_shuffle_ps(m6, m7, _MM_SHUFFLE(0, 2, 0, 2));
for (int i = 0; i < len; i++) {
/* Load (unaligned) input data */
m0 = _mm_loadu_ps(&x[2 * i + 0]);
m1 = _mm_loadu_ps(&x[2 * i + 4]);
m2 = _mm_loadu_ps(&x[2 * i + 8]);
m3 = _mm_loadu_ps(&x[2 * i + 12]);
m4 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(0, 2, 0, 2));
m5 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(1, 3, 1, 3));
m6 = _mm_shuffle_ps(m2, m3, _MM_SHUFFLE(0, 2, 0, 2));
m7 = _mm_shuffle_ps(m2, m3, _MM_SHUFFLE(1, 3, 1, 3));
m0 = _mm_loadu_ps(&x[2 * i + 16]);
m1 = _mm_loadu_ps(&x[2 * i + 20]);
m2 = _mm_loadu_ps(&x[2 * i + 24]);
m3 = _mm_loadu_ps(&x[2 * i + 28]);
m8 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(0, 2, 0, 2));
m9 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(1, 3, 1, 3));
m10 = _mm_shuffle_ps(m2, m3, _MM_SHUFFLE(0, 2, 0, 2));
m11 = _mm_shuffle_ps(m2, m3, _MM_SHUFFLE(1, 3, 1, 3));
/* Quad multiply */
m0 = _mm_mul_ps(m4, m12);
m1 = _mm_mul_ps(m5, m12);
m2 = _mm_mul_ps(m6, m13);
m3 = _mm_mul_ps(m7, m13);
m4 = _mm_mul_ps(m8, m14);
m5 = _mm_mul_ps(m9, m14);
m6 = _mm_mul_ps(m10, m15);
m7 = _mm_mul_ps(m11, m15);
/* Sum and store */
m8 = _mm_add_ps(m0, m2);
m9 = _mm_add_ps(m1, m3);
m10 = _mm_add_ps(m4, m6);
m11 = _mm_add_ps(m5, m7);
m0 = _mm_add_ps(m8, m10);
m1 = _mm_add_ps(m9, m11);
m2 = _mm_hadd_ps(m0, m1);
m3 = _mm_hadd_ps(m2, m2);
_mm_store_ss(&y[2 * i + 0], m3);
m3 = _mm_shuffle_ps(m3, m3, _MM_SHUFFLE(0, 3, 2, 1));
_mm_store_ss(&y[2 * i + 1], m3);
}
}
/* 20-tap SSE complex-real convolution */
static void sse_conv_real20(float *restrict x,
float *restrict h,
float *restrict y,
int len)
{
__m128 m0, m1, m2, m3, m4, m5, m6, m7;
__m128 m8, m9, m11, m12, m13, m14, m15;
/* Load (aligned) filter taps */
m0 = _mm_load_ps(&h[0]);
m1 = _mm_load_ps(&h[4]);
m2 = _mm_load_ps(&h[8]);
m3 = _mm_load_ps(&h[12]);
m4 = _mm_load_ps(&h[16]);
m5 = _mm_load_ps(&h[20]);
m6 = _mm_load_ps(&h[24]);
m7 = _mm_load_ps(&h[28]);
m8 = _mm_load_ps(&h[32]);
m9 = _mm_load_ps(&h[36]);
m11 = _mm_shuffle_ps(m0, m1, _MM_SHUFFLE(0, 2, 0, 2));
m12 = _mm_shuffle_ps(m2, m3, _MM_SHUFFLE(0, 2, 0, 2));
m13 = _mm_shuffle_ps(m4, m5, _MM_SHUFFLE(0, 2, 0, 2));
m14 = _mm_shuffle_ps(m6, m7, _MM_SHUFFLE(0, 2, 0, 2));
m15 = _mm_shuffle_ps(m8, m9, _MM_SHUFFLE(0, 2, 0, 2));
for (int i = 0; i < len; i++) {
/* Multiply-accumulate first 12 taps */
m0 = _mm_loadu_ps(