1 /* Copyright (c) 2015 Xiph.Org Foundation 2 Written by Viswanath Puttagunta */ 3 /** 4 @file celt_ne10_mdct.c 5 @brief ARM Neon optimizations for mdct using NE10 library 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 #ifndef SKIP_CONFIG_H 34 #ifdef HAVE_CONFIG_H 35 #include "config.h" 36 #endif 37 #endif 38 39 #include "kiss_fft.h" 40 #include "_kiss_fft_guts.h" 41 #include "mdct.h" 42 #include "stack_alloc.h" 43 44 void clt_mdct_forward_neon(const mdct_lookup *l, 45 kiss_fft_scalar *in, 46 kiss_fft_scalar * OPUS_RESTRICT out, 47 const opus_val16 *window, 48 int overlap, int shift, int stride, int arch) 49 { 50 int i; 51 int N, N2, N4; 52 VARDECL(kiss_fft_scalar, f); 53 VARDECL(kiss_fft_cpx, f2); 54 const kiss_fft_state *st = l->kfft[shift]; 55 const kiss_twiddle_scalar *trig; 56 57 SAVE_STACK; 58 59 N = l->n; 60 trig = l->trig; 61 for (i=0;i<shift;i++) 62 { 63 N >>= 1; 64 trig += N; 65 } 66 N2 = N>>1; 67 N4 = N>>2; 68 69 ALLOC(f, N2, kiss_fft_scalar); 70 ALLOC(f2, N4, kiss_fft_cpx); 71 72 /* Consider the input to be composed of four blocks: [a, b, c, d] */ 73 /* Window, shuffle, fold */ 74 { 75 /* Temp pointers to make it really clear to the compiler what we're doing */ 76 const kiss_fft_scalar * OPUS_RESTRICT xp1 = in+(overlap>>1); 77 const kiss_fft_scalar * OPUS_RESTRICT xp2 = in+N2-1+(overlap>>1); 78 kiss_fft_scalar * OPUS_RESTRICT yp = f; 79 const opus_val16 * OPUS_RESTRICT wp1 = window+(overlap>>1); 80 const opus_val16 * OPUS_RESTRICT wp2 = window+(overlap>>1)-1; 81 for(i=0;i<((overlap+3)>>2);i++) 82 { 83 /* Real part arranged as -d-cR, Imag part arranged as -b+aR*/ 84 *yp++ = MULT16_32_Q15(*wp2, xp1[N2]) + MULT16_32_Q15(*wp1,*xp2); 85 *yp++ = MULT16_32_Q15(*wp1, *xp1) - MULT16_32_Q15(*wp2, xp2[-N2]); 86 xp1+=2; 87 xp2-=2; 88 wp1+=2; 89 wp2-=2; 90 } 91 wp1 = window; 92 wp2 = window+overlap-1; 93 for(;i<N4-((overlap+3)>>2);i++) 94 { 95 /* Real part arranged as a-bR, Imag part arranged as -c-dR */ 96 *yp++ = *xp2; 97 *yp++ = *xp1; 98 xp1+=2; 99 xp2-=2; 100 } 101 for(;i<N4;i++) 102 { 103 /* Real part arranged as a-bR, Imag part arranged as -c-dR */ 104 *yp++ = -MULT16_32_Q15(*wp1, xp1[-N2]) + MULT16_32_Q15(*wp2, *xp2); 105 *yp++ = MULT16_32_Q15(*wp2, *xp1) + MULT16_32_Q15(*wp1, xp2[N2]); 106 xp1+=2; 107 xp2-=2; 108 wp1+=2; 109 wp2-=2; 110 } 111 } 112 /* Pre-rotation */ 113 { 114 kiss_fft_scalar * OPUS_RESTRICT yp = f; 115 const kiss_twiddle_scalar *t = &trig[0]; 116 for(i=0;i<N4;i++) 117 { 118 kiss_fft_cpx yc; 119 kiss_twiddle_scalar t0, t1; 120 kiss_fft_scalar re, im, yr, yi; 121 t0 = t[i]; 122 t1 = t[N4+i]; 123 re = *yp++; 124 im = *yp++; 125 yr = S_MUL(re,t0) - S_MUL(im,t1); 126 yi = S_MUL(im,t0) + S_MUL(re,t1); 127 yc.r = yr; 128 yc.i = yi; 129 f2[i] = yc; 130 } 131 } 132 133 opus_fft(st, f2, (kiss_fft_cpx *)f, arch); 134 135 /* Post-rotate */ 136 { 137 /* Temp pointers to make it really clear to the compiler what we're doing */ 138 const kiss_fft_cpx * OPUS_RESTRICT fp = (kiss_fft_cpx *)f; 139 kiss_fft_scalar * OPUS_RESTRICT yp1 = out; 140 kiss_fft_scalar * OPUS_RESTRICT yp2 = out+stride*(N2-1); 141 const kiss_twiddle_scalar *t = &trig[0]; 142 /* Temp pointers to make it really clear to the compiler what we're doing */ 143 for(i=0;i<N4;i++) 144 { 145 kiss_fft_scalar yr, yi; 146 yr = S_MUL(fp->i,t[N4+i]) - S_MUL(fp->r,t[i]); 147 yi = S_MUL(fp->r,t[N4+i]) + S_MUL(fp->i,t[i]); 148 *yp1 = yr; 149 *yp2 = yi; 150 fp++; 151 yp1 += 2*stride; 152 yp2 -= 2*stride; 153 } 154 } 155 RESTORE_STACK; 156 } 157 158 void clt_mdct_backward_neon(const mdct_lookup *l, 159 kiss_fft_scalar *in, 160 kiss_fft_scalar * OPUS_RESTRICT out, 161 const opus_val16 * OPUS_RESTRICT window, 162 int overlap, int shift, int stride, int arch) 163 { 164 int i; 165 int N, N2, N4; 166 VARDECL(kiss_fft_scalar, f); 167 const kiss_twiddle_scalar *trig; 168 const kiss_fft_state *st = l->kfft[shift]; 169 170 N = l->n; 171 trig = l->trig; 172 for (i=0;i<shift;i++) 173 { 174 N >>= 1; 175 trig += N; 176 } 177 N2 = N>>1; 178 N4 = N>>2; 179 180 ALLOC(f, N2, kiss_fft_scalar); 181 182 /* Pre-rotate */ 183 { 184 /* Temp pointers to make it really clear to the compiler what we're doing */ 185 const kiss_fft_scalar * OPUS_RESTRICT xp1 = in; 186 const kiss_fft_scalar * OPUS_RESTRICT xp2 = in+stride*(N2-1); 187 kiss_fft_scalar * OPUS_RESTRICT yp = f; 188 const kiss_twiddle_scalar * OPUS_RESTRICT t = &trig[0]; 189 for(i=0;i<N4;i++) 190 { 191 kiss_fft_scalar yr, yi; 192 yr = S_MUL(*xp2, t[i]) + S_MUL(*xp1, t[N4+i]); 193 yi = S_MUL(*xp1, t[i]) - S_MUL(*xp2, t[N4+i]); 194 yp[2*i] = yr; 195 yp[2*i+1] = yi; 196 xp1+=2*stride; 197 xp2-=2*stride; 198 } 199 } 200 201 opus_ifft(st, (kiss_fft_cpx *)f, (kiss_fft_cpx*)(out+(overlap>>1)), arch); 202 203 /* Post-rotate and de-shuffle from both ends of the buffer at once to make 204 it in-place. */ 205 { 206 kiss_fft_scalar * yp0 = out+(overlap>>1); 207 kiss_fft_scalar * yp1 = out+(overlap>>1)+N2-2; 208 const kiss_twiddle_scalar *t = &trig[0]; 209 /* Loop to (N4+1)>>1 to handle odd N4. When N4 is odd, the 210 middle pair will be computed twice. */ 211 for(i=0;i<(N4+1)>>1;i++) 212 { 213 kiss_fft_scalar re, im, yr, yi; 214 kiss_twiddle_scalar t0, t1; 215 re = yp0[0]; 216 im = yp0[1]; 217 t0 = t[i]; 218 t1 = t[N4+i]; 219 /* We'd scale up by 2 here, but instead it's done when mixing the windows */ 220 yr = S_MUL(re,t0) + S_MUL(im,t1); 221 yi = S_MUL(re,t1) - S_MUL(im,t0); 222 re = yp1[0]; 223 im = yp1[1]; 224 yp0[0] = yr; 225 yp1[1] = yi; 226 227 t0 = t[(N4-i-1)]; 228 t1 = t[(N2-i-1)]; 229 /* We'd scale up by 2 here, but instead it's done when mixing the windows */ 230 yr = S_MUL(re,t0) + S_MUL(im,t1); 231 yi = S_MUL(re,t1) - S_MUL(im,t0); 232 yp1[0] = yr; 233 yp0[1] = yi; 234 yp0 += 2; 235 yp1 -= 2; 236 } 237 } 238 239 /* Mirror on both sides for TDAC */ 240 { 241 kiss_fft_scalar * OPUS_RESTRICT xp1 = out+overlap-1; 242 kiss_fft_scalar * OPUS_RESTRICT yp1 = out; 243 const opus_val16 * OPUS_RESTRICT wp1 = window; 244 const opus_val16 * OPUS_RESTRICT wp2 = window+overlap-1; 245 246 for(i = 0; i < overlap/2; i++) 247 { 248 kiss_fft_scalar x1, x2; 249 x1 = *xp1; 250 x2 = *yp1; 251 *yp1++ = MULT16_32_Q15(*wp2, x2) - MULT16_32_Q15(*wp1, x1); 252 *xp1-- = MULT16_32_Q15(*wp1, x2) + MULT16_32_Q15(*wp2, x1); 253 wp1++; 254 wp2--; 255 } 256 } 257 RESTORE_STACK; 258 } 259
