1 /* Copyright (c) 2014, Cisco Systems, INC 2 Written by XiangMingZhu WeiZhou MinPeng YanWang 3 4 Redistribution and use in source and binary forms, with or without 5 modification, are permitted provided that the following conditions 6 are met: 7 8 - Redistributions of source code must retain the above copyright 9 notice, this list of conditions and the following disclaimer. 10 11 - Redistributions in binary form must reproduce the above copyright 12 notice, this list of conditions and the following disclaimer in the 13 documentation and/or other materials provided with the distribution. 14 15 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 16 ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 17 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 18 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER 19 OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, 20 EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 21 PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR 22 PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF 23 LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING 24 NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 25 SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 26 */ 27 28 #ifdef HAVE_CONFIG_H 29 #include "config.h" 30 #endif 31 32 #include "macros.h" 33 #include "celt_lpc.h" 34 #include "stack_alloc.h" 35 #include "mathops.h" 36 #include "pitch.h" 37 38 #if defined(OPUS_X86_MAY_HAVE_SSE) && !defined(FIXED_POINT) 39 40 #include <xmmintrin.h> 41 #include "arch.h" 42 43 void xcorr_kernel_sse(const opus_val16 *x, const opus_val16 *y, opus_val32 sum[4], int len) 44 { 45 int j; 46 __m128 xsum1, xsum2; 47 xsum1 = _mm_loadu_ps(sum); 48 xsum2 = _mm_setzero_ps(); 49 50 for (j = 0; j < len-3; j += 4) 51 { 52 __m128 x0 = _mm_loadu_ps(x+j); 53 __m128 yj = _mm_loadu_ps(y+j); 54 __m128 y3 = _mm_loadu_ps(y+j+3); 55 56 xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0x00),yj)); 57 xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0x55), 58 _mm_shuffle_ps(yj,y3,0x49))); 59 xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0xaa), 60 _mm_shuffle_ps(yj,y3,0x9e))); 61 xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0xff),y3)); 62 } 63 if (j < len) 64 { 65 xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_load1_ps(x+j),_mm_loadu_ps(y+j))); 66 if (++j < len) 67 { 68 xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(_mm_load1_ps(x+j),_mm_loadu_ps(y+j))); 69 if (++j < len) 70 { 71 xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_load1_ps(x+j),_mm_loadu_ps(y+j))); 72 } 73 } 74 } 75 _mm_storeu_ps(sum,_mm_add_ps(xsum1,xsum2)); 76 } 77 78 79 void dual_inner_prod_sse(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02, 80 int N, opus_val32 *xy1, opus_val32 *xy2) 81 { 82 int i; 83 __m128 xsum1, xsum2; 84 xsum1 = _mm_setzero_ps(); 85 xsum2 = _mm_setzero_ps(); 86 for (i=0;i<N-3;i+=4) 87 { 88 __m128 xi = _mm_loadu_ps(x+i); 89 __m128 y1i = _mm_loadu_ps(y01+i); 90 __m128 y2i = _mm_loadu_ps(y02+i); 91 xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(xi, y1i)); 92 xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(xi, y2i)); 93 } 94 /* Horizontal sum */ 95 xsum1 = _mm_add_ps(xsum1, _mm_movehl_ps(xsum1, xsum1)); 96 xsum1 = _mm_add_ss(xsum1, _mm_shuffle_ps(xsum1, xsum1, 0x55)); 97 _mm_store_ss(xy1, xsum1); 98 xsum2 = _mm_add_ps(xsum2, _mm_movehl_ps(xsum2, xsum2)); 99 xsum2 = _mm_add_ss(xsum2, _mm_shuffle_ps(xsum2, xsum2, 0x55)); 100 _mm_store_ss(xy2, xsum2); 101 for (;i<N;i++) 102 { 103 *xy1 = MAC16_16(*xy1, x[i], y01[i]); 104 *xy2 = MAC16_16(*xy2, x[i], y02[i]); 105 } 106 } 107 108 opus_val32 celt_inner_prod_sse(const opus_val16 *x, const opus_val16 *y, 109 int N) 110 { 111 int i; 112 float xy; 113 __m128 sum; 114 sum = _mm_setzero_ps(); 115 /* FIXME: We should probably go 8-way and use 2 sums. */ 116 for (i=0;i<N-3;i+=4) 117 { 118 __m128 xi = _mm_loadu_ps(x+i); 119 __m128 yi = _mm_loadu_ps(y+i); 120 sum = _mm_add_ps(sum,_mm_mul_ps(xi, yi)); 121 } 122 /* Horizontal sum */ 123 sum = _mm_add_ps(sum, _mm_movehl_ps(sum, sum)); 124 sum = _mm_add_ss(sum, _mm_shuffle_ps(sum, sum, 0x55)); 125 _mm_store_ss(&xy, sum); 126 for (;i<N;i++) 127 { 128 xy = MAC16_16(xy, x[i], y[i]); 129 } 130 return xy; 131 } 132 133 void comb_filter_const_sse(opus_val32 *y, opus_val32 *x, int T, int N, 134 opus_val16 g10, opus_val16 g11, opus_val16 g12) 135 { 136 int i; 137 __m128 x0v; 138 __m128 g10v, g11v, g12v; 139 g10v = _mm_load1_ps(&g10); 140 g11v = _mm_load1_ps(&g11); 141 g12v = _mm_load1_ps(&g12); 142 x0v = _mm_loadu_ps(&x[-T-2]); 143 for (i=0;i<N-3;i+=4) 144 { 145 __m128 yi, yi2, x1v, x2v, x3v, x4v; 146 const opus_val32 *xp = &x[i-T-2]; 147 yi = _mm_loadu_ps(x+i); 148 x4v = _mm_loadu_ps(xp+4); 149 #if 0 150 /* Slower version with all loads */ 151 x1v = _mm_loadu_ps(xp+1); 152 x2v = _mm_loadu_ps(xp+2); 153 x3v = _mm_loadu_ps(xp+3); 154 #else 155 x2v = _mm_shuffle_ps(x0v, x4v, 0x4e); 156 x1v = _mm_shuffle_ps(x0v, x2v, 0x99); 157 x3v = _mm_shuffle_ps(x2v, x4v, 0x99); 158 #endif 159 160 yi = _mm_add_ps(yi, _mm_mul_ps(g10v,x2v)); 161 #if 0 /* Set to 1 to make it bit-exact with the non-SSE version */ 162 yi = _mm_add_ps(yi, _mm_mul_ps(g11v,_mm_add_ps(x3v,x1v))); 163 yi = _mm_add_ps(yi, _mm_mul_ps(g12v,_mm_add_ps(x4v,x0v))); 164 #else 165 /* Use partial sums */ 166 yi2 = _mm_add_ps(_mm_mul_ps(g11v,_mm_add_ps(x3v,x1v)), 167 _mm_mul_ps(g12v,_mm_add_ps(x4v,x0v))); 168 yi = _mm_add_ps(yi, yi2); 169 #endif 170 x0v=x4v; 171 _mm_storeu_ps(y+i, yi); 172 } 173 #ifdef CUSTOM_MODES 174 for (;i<N;i++) 175 { 176 y[i] = x[i] 177 + MULT16_32_Q15(g10,x[i-T]) 178 + MULT16_32_Q15(g11,ADD32(x[i-T+1],x[i-T-1])) 179 + MULT16_32_Q15(g12,ADD32(x[i-T+2],x[i-T-2])); 180 } 181 #endif 182 } 183 184 185 #endif 186
