1 /* Copyright (c) 2014-2015 Xiph.Org Foundation 2 Written by Viswanath Puttagunta */ 3 /** 4 @file celt_neon_intr.c 5 @brief ARM Neon Intrinsic optimizations for celt 6 */ 7 8 /* 9 Redistribution and use in source and binary forms, with or without 10 modification, are permitted provided that the following conditions 11 are met: 12 13 - Redistributions of source code must retain the above copyright 14 notice, this list of conditions and the following disclaimer. 15 16 - Redistributions in binary form must reproduce the above copyright 17 notice, this list of conditions and the following disclaimer in the 18 documentation and/or other materials provided with the distribution. 19 20 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 21 ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 22 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 23 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER 24 OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, 25 EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 26 PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR 27 PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF 28 LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING 29 NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 30 SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 31 */ 32 33 #ifdef HAVE_CONFIG_H 34 #include "config.h" 35 #endif 36 37 #include <arm_neon.h> 38 #include "../pitch.h" 39 40 #if defined(FIXED_POINT) 41 void xcorr_kernel_neon_fixed(const opus_val16 * x, const opus_val16 * y, opus_val32 sum[4], int len) 42 { 43 int j; 44 int32x4_t a = vld1q_s32(sum); 45 /* Load y[0...3] */ 46 /* This requires len>0 to always be valid (which we assert in the C code). */ 47 int16x4_t y0 = vld1_s16(y); 48 y += 4; 49 50 for (j = 0; j + 8 <= len; j += 8) 51 { 52 /* Load x[0...7] */ 53 int16x8_t xx = vld1q_s16(x); 54 int16x4_t x0 = vget_low_s16(xx); 55 int16x4_t x4 = vget_high_s16(xx); 56 /* Load y[4...11] */ 57 int16x8_t yy = vld1q_s16(y); 58 int16x4_t y4 = vget_low_s16(yy); 59 int16x4_t y8 = vget_high_s16(yy); 60 int32x4_t a0 = vmlal_lane_s16(a, y0, x0, 0); 61 int32x4_t a1 = vmlal_lane_s16(a0, y4, x4, 0); 62 63 int16x4_t y1 = vext_s16(y0, y4, 1); 64 int16x4_t y5 = vext_s16(y4, y8, 1); 65 int32x4_t a2 = vmlal_lane_s16(a1, y1, x0, 1); 66 int32x4_t a3 = vmlal_lane_s16(a2, y5, x4, 1); 67 68 int16x4_t y2 = vext_s16(y0, y4, 2); 69 int16x4_t y6 = vext_s16(y4, y8, 2); 70 int32x4_t a4 = vmlal_lane_s16(a3, y2, x0, 2); 71 int32x4_t a5 = vmlal_lane_s16(a4, y6, x4, 2); 72 73 int16x4_t y3 = vext_s16(y0, y4, 3); 74 int16x4_t y7 = vext_s16(y4, y8, 3); 75 int32x4_t a6 = vmlal_lane_s16(a5, y3, x0, 3); 76 int32x4_t a7 = vmlal_lane_s16(a6, y7, x4, 3); 77 78 y0 = y8; 79 a = a7; 80 x += 8; 81 y += 8; 82 } 83 84 for (; j < len; j++) 85 { 86 int16x4_t x0 = vld1_dup_s16(x); /* load next x */ 87 int32x4_t a0 = vmlal_s16(a, y0, x0); 88 89 int16x4_t y4 = vld1_dup_s16(y); /* load next y */ 90 y0 = vext_s16(y0, y4, 1); 91 a = a0; 92 x++; 93 y++; 94 } 95 96 vst1q_s32(sum, a); 97 } 98 99 #else 100 /* 101 * Function: xcorr_kernel_neon_float 102 * --------------------------------- 103 * Computes 4 correlation values and stores them in sum[4] 104 */ 105 static void xcorr_kernel_neon_float(const float32_t *x, const float32_t *y, 106 float32_t sum[4], int len) { 107 float32x4_t YY[3]; 108 float32x4_t YEXT[3]; 109 float32x4_t XX[2]; 110 float32x2_t XX_2; 111 float32x4_t SUMM; 112 const float32_t *xi = x; 113 const float32_t *yi = y; 114 115 celt_assert(len>0); 116 117 YY[0] = vld1q_f32(yi); 118 SUMM = vdupq_n_f32(0); 119 120 /* Consume 8 elements in x vector and 12 elements in y 121 * vector. However, the 12'th element never really gets 122 * touched in this loop. So, if len == 8, then we only 123 * must access y[0] to y[10]. y[11] must not be accessed 124 * hence make sure len > 8 and not len >= 8 125 */ 126 while (len > 8) { 127 yi += 4; 128 YY[1] = vld1q_f32(yi); 129 yi += 4; 130 YY[2] = vld1q_f32(yi); 131 132 XX[0] = vld1q_f32(xi); 133 xi += 4; 134 XX[1] = vld1q_f32(xi); 135 xi += 4; 136 137 SUMM = vmlaq_lane_f32(SUMM, YY[0], vget_low_f32(XX[0]), 0); 138 YEXT[0] = vextq_f32(YY[0], YY[1], 1); 139 SUMM = vmlaq_lane_f32(SUMM, YEXT[0], vget_low_f32(XX[0]), 1); 140 YEXT[1] = vextq_f32(YY[0], YY[1], 2); 141 SUMM = vmlaq_lane_f32(SUMM, YEXT[1], vget_high_f32(XX[0]), 0); 142 YEXT[2] = vextq_f32(YY[0], YY[1], 3); 143 SUMM = vmlaq_lane_f32(SUMM, YEXT[2], vget_high_f32(XX[0]), 1); 144 145 SUMM = vmlaq_lane_f32(SUMM, YY[1], vget_low_f32(XX[1]), 0); 146 YEXT[0] = vextq_f32(YY[1], YY[2], 1); 147 SUMM = vmlaq_lane_f32(SUMM, YEXT[0], vget_low_f32(XX[1]), 1); 148 YEXT[1] = vextq_f32(YY[1], YY[2], 2); 149 SUMM = vmlaq_lane_f32(SUMM, YEXT[1], vget_high_f32(XX[1]), 0); 150 YEXT[2] = vextq_f32(YY[1], YY[2], 3); 151 SUMM = vmlaq_lane_f32(SUMM, YEXT[2], vget_high_f32(XX[1]), 1); 152 153 YY[0] = YY[2]; 154 len -= 8; 155 } 156 157 /* Consume 4 elements in x vector and 8 elements in y 158 * vector. However, the 8'th element in y never really gets 159 * touched in this loop. So, if len == 4, then we only 160 * must access y[0] to y[6]. y[7] must not be accessed 161 * hence make sure len>4 and not len>=4 162 */ 163 if (len > 4) { 164 yi += 4; 165 YY[1] = vld1q_f32(yi); 166 167 XX[0] = vld1q_f32(xi); 168 xi += 4; 169 170 SUMM = vmlaq_lane_f32(SUMM, YY[0], vget_low_f32(XX[0]), 0); 171 YEXT[0] = vextq_f32(YY[0], YY[1], 1); 172 SUMM = vmlaq_lane_f32(SUMM, YEXT[0], vget_low_f32(XX[0]), 1); 173 YEXT[1] = vextq_f32(YY[0], YY[1], 2); 174 SUMM = vmlaq_lane_f32(SUMM, YEXT[1], vget_high_f32(XX[0]), 0); 175 YEXT[2] = vextq_f32(YY[0], YY[1], 3); 176 SUMM = vmlaq_lane_f32(SUMM, YEXT[2], vget_high_f32(XX[0]), 1); 177 178 YY[0] = YY[1]; 179 len -= 4; 180 } 181 182 while (--len > 0) { 183 XX_2 = vld1_dup_f32(xi++); 184 SUMM = vmlaq_lane_f32(SUMM, YY[0], XX_2, 0); 185 YY[0]= vld1q_f32(++yi); 186 } 187 188 XX_2 = vld1_dup_f32(xi); 189 SUMM = vmlaq_lane_f32(SUMM, YY[0], XX_2, 0); 190 191 vst1q_f32(sum, SUMM); 192 } 193 194 void celt_pitch_xcorr_float_neon(const opus_val16 *_x, const opus_val16 *_y, 195 opus_val32 *xcorr, int len, int max_pitch, int arch) { 196 int i; 197 (void)arch; 198 celt_assert(max_pitch > 0); 199 celt_sig_assert((((unsigned char *)_x-(unsigned char *)NULL)&3)==0); 200 201 for (i = 0; i < (max_pitch-3); i += 4) { 202 xcorr_kernel_neon_float((const float32_t *)_x, (const float32_t *)_y+i, 203 (float32_t *)xcorr+i, len); 204 } 205 206 /* In case max_pitch isn't a multiple of 4, do non-unrolled version. */ 207 for (; i < max_pitch; i++) { 208 xcorr[i] = celt_inner_prod_neon(_x, _y+i, len); 209 } 210 } 211 #endif 212