1 /* 2 * Copyright (c) 2011 The WebRTC project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11 /* 12 * lattice.c 13 * 14 * Contains the normalized lattice filter routines (MA and AR) for iSAC codec 15 * 16 */ 17 18 #include "codec.h" 19 #include "settings.h" 20 21 #define LATTICE_MUL_32_32_RSFT16(a32a, a32b, b32) \ 22 ((int32_t)(WEBRTC_SPL_MUL(a32a, b32) + (WEBRTC_SPL_MUL_16_32_RSFT16(a32b, b32)))) 23 /* This macro is FORBIDDEN to use elsewhere than in a function in this file and 24 its corresponding neon version. It might give unpredictable results, since a 25 general int32_t*int32_t multiplication results in a 64 bit value. 26 The result is then shifted just 16 steps to the right, giving need for 48 27 bits, i.e. in the generel case, it will NOT fit in a int32_t. In the 28 cases used in here, the int32_t will be enough, since (for a good 29 reason) the involved multiplicands aren't big enough to overflow a 30 int32_t after shifting right 16 bits. I have compared the result of a 31 multiplication between t32 and tmp32, done in two ways: 32 1) Using (int32_t) (((float)(tmp32))*((float)(tmp32b))/65536.0); 33 2) Using LATTICE_MUL_32_32_RSFT16(t16a, t16b, tmp32b); 34 By running 25 files, I haven't found any bigger diff than 64 - this was in the 35 case when method 1) gave 650235648 and 2) gave 650235712. 36 */ 37 38 /* Function prototype: filtering ar_g_Q0[] and ar_f_Q0[] through an AR filter 39 with coefficients cth_Q15[] and sth_Q15[]. 40 Implemented for both generic and ARMv7 platforms. 41 */ 42 void WebRtcIsacfix_FilterArLoop(int16_t* ar_g_Q0, 43 int16_t* ar_f_Q0, 44 int16_t* cth_Q15, 45 int16_t* sth_Q15, 46 size_t order_coef); 47 48 /* Inner loop used for function WebRtcIsacfix_NormLatticeFilterMa(). It does: 49 for 0 <= n < HALF_SUBFRAMELEN - 1: 50 *ptr2 = input2 * (*ptr2) + input0 * (*ptr0)); 51 *ptr1 = input1 * (*ptr0) + input0 * (*ptr2); 52 Note, function WebRtcIsacfix_FilterMaLoopNeon and WebRtcIsacfix_FilterMaLoopC 53 are not bit-exact. The accuracy by the ARM Neon function is same or better. 54 */ WebRtcIsacfix_FilterMaLoopC(int16_t input0,int16_t input1,int32_t input2,int32_t * ptr0,int32_t * ptr1,int32_t * ptr2)55 void WebRtcIsacfix_FilterMaLoopC(int16_t input0, // Filter coefficient 56 int16_t input1, // Filter coefficient 57 int32_t input2, // Inverse coeff. (1/input1) 58 int32_t* ptr0, // Sample buffer 59 int32_t* ptr1, // Sample buffer 60 int32_t* ptr2) { // Sample buffer 61 int n = 0; 62 63 // Separate the 32-bit variable input2 into two 16-bit integers (high 16 and 64 // low 16 bits), for using LATTICE_MUL_32_32_RSFT16 in the loop. 65 int16_t t16a = (int16_t)(input2 >> 16); 66 int16_t t16b = (int16_t)input2; 67 if (t16b < 0) t16a++; 68 69 // The loop filtering the samples *ptr0, *ptr1, *ptr2 with filter coefficients 70 // input0, input1, and input2. 71 for(n = 0; n < HALF_SUBFRAMELEN - 1; n++, ptr0++, ptr1++, ptr2++) { 72 int32_t tmp32a = 0; 73 int32_t tmp32b = 0; 74 75 // Calculate *ptr2 = input2 * (*ptr2 + input0 * (*ptr0)); 76 tmp32a = WEBRTC_SPL_MUL_16_32_RSFT15(input0, *ptr0); // Q15 * Q15 >> 15 = Q15 77 tmp32b = *ptr2 + tmp32a; // Q15 + Q15 = Q15 78 *ptr2 = LATTICE_MUL_32_32_RSFT16(t16a, t16b, tmp32b); 79 80 // Calculate *ptr1 = input1 * (*ptr0) + input0 * (*ptr2); 81 tmp32a = WEBRTC_SPL_MUL_16_32_RSFT15(input1, *ptr0); // Q15*Q15>>15 = Q15 82 tmp32b = WEBRTC_SPL_MUL_16_32_RSFT15(input0, *ptr2); // Q15*Q15>>15 = Q15 83 *ptr1 = tmp32a + tmp32b; // Q15 + Q15 = Q15 84 } 85 } 86 87 /* filter the signal using normalized lattice filter */ 88 /* MA filter */ WebRtcIsacfix_NormLatticeFilterMa(size_t orderCoef,int32_t * stateGQ15,int16_t * lat_inQ0,int16_t * filt_coefQ15,int32_t * gain_lo_hiQ17,int16_t lo_hi,int16_t * lat_outQ9)89 void WebRtcIsacfix_NormLatticeFilterMa(size_t orderCoef, 90 int32_t *stateGQ15, 91 int16_t *lat_inQ0, 92 int16_t *filt_coefQ15, 93 int32_t *gain_lo_hiQ17, 94 int16_t lo_hi, 95 int16_t *lat_outQ9) 96 { 97 int16_t sthQ15[MAX_AR_MODEL_ORDER]; 98 int16_t cthQ15[MAX_AR_MODEL_ORDER]; 99 100 int u, n; 101 size_t i, k; 102 int16_t temp2,temp3; 103 size_t ord_1 = orderCoef+1; 104 int32_t inv_cthQ16[MAX_AR_MODEL_ORDER]; 105 106 int32_t gain32, fQtmp; 107 int16_t gain16; 108 int16_t gain_sh; 109 110 int32_t tmp32, tmp32b; 111 int32_t fQ15vec[HALF_SUBFRAMELEN]; 112 int32_t gQ15[MAX_AR_MODEL_ORDER+1][HALF_SUBFRAMELEN]; 113 int16_t sh; 114 int16_t t16a; 115 int16_t t16b; 116 117 for (u=0;u<SUBFRAMES;u++) 118 { 119 int32_t temp1 = u * HALF_SUBFRAMELEN; 120 121 /* set the Direct Form coefficients */ 122 temp2 = (int16_t)(u * orderCoef); 123 temp3 = (int16_t)(2 * u + lo_hi); 124 125 /* compute lattice filter coefficients */ 126 memcpy(sthQ15, &filt_coefQ15[temp2], orderCoef * sizeof(int16_t)); 127 128 WebRtcSpl_SqrtOfOneMinusXSquared(sthQ15, orderCoef, cthQ15); 129 130 /* compute the gain */ 131 gain32 = gain_lo_hiQ17[temp3]; 132 gain_sh = WebRtcSpl_NormW32(gain32); 133 gain32 <<= gain_sh; // Q(17+gain_sh) 134 135 for (k=0;k<orderCoef;k++) 136 { 137 gain32 = WEBRTC_SPL_MUL_16_32_RSFT15(cthQ15[k], gain32); //Q15*Q(17+gain_sh)>>15 = Q(17+gain_sh) 138 inv_cthQ16[k] = WebRtcSpl_DivW32W16((int32_t)2147483647, cthQ15[k]); // 1/cth[k] in Q31/Q15 = Q16 139 } 140 gain16 = (int16_t)(gain32 >> 16); // Q(1+gain_sh). 141 142 /* normalized lattice filter */ 143 /*****************************/ 144 145 /* initial conditions */ 146 for (i=0;i<HALF_SUBFRAMELEN;i++) 147 { 148 fQ15vec[i] = lat_inQ0[i + temp1] << 15; // Q15 149 gQ15[0][i] = lat_inQ0[i + temp1] << 15; // Q15 150 } 151 152 153 fQtmp = fQ15vec[0]; 154 155 /* get the state of f&g for the first input, for all orders */ 156 for (i=1;i<ord_1;i++) 157 { 158 // Calculate f[i][0] = inv_cth[i-1]*(f[i-1][0] + sth[i-1]*stateG[i-1]); 159 tmp32 = WEBRTC_SPL_MUL_16_32_RSFT15(sthQ15[i-1], stateGQ15[i-1]);//Q15*Q15>>15 = Q15 160 tmp32b= fQtmp + tmp32; //Q15+Q15=Q15 161 tmp32 = inv_cthQ16[i-1]; //Q16 162 t16a = (int16_t)(tmp32 >> 16); 163 t16b = (int16_t)(tmp32 - (t16a << 16)); 164 if (t16b<0) t16a++; 165 tmp32 = LATTICE_MUL_32_32_RSFT16(t16a, t16b, tmp32b); 166 fQtmp = tmp32; // Q15 167 168 // Calculate g[i][0] = cth[i-1]*stateG[i-1] + sth[i-1]* f[i][0]; 169 tmp32 = WEBRTC_SPL_MUL_16_32_RSFT15(cthQ15[i-1], stateGQ15[i-1]); //Q15*Q15>>15 = Q15 170 tmp32b = WEBRTC_SPL_MUL_16_32_RSFT15(sthQ15[i-1], fQtmp); //Q15*Q15>>15 = Q15 171 tmp32 = tmp32 + tmp32b;//Q15+Q15 = Q15 172 gQ15[i][0] = tmp32; // Q15 173 } 174 175 /* filtering */ 176 /* save the states */ 177 for(k=0;k<orderCoef;k++) 178 { 179 // for 0 <= n < HALF_SUBFRAMELEN - 1: 180 // f[k+1][n+1] = inv_cth[k]*(f[k][n+1] + sth[k]*g[k][n]); 181 // g[k+1][n+1] = cth[k]*g[k][n] + sth[k]* f[k+1][n+1]; 182 WebRtcIsacfix_FilterMaLoopFix(sthQ15[k], cthQ15[k], inv_cthQ16[k], 183 &gQ15[k][0], &gQ15[k+1][1], &fQ15vec[1]); 184 } 185 186 fQ15vec[0] = fQtmp; 187 188 for(n=0;n<HALF_SUBFRAMELEN;n++) 189 { 190 //gain32 >>= gain_sh; // Q(17+gain_sh) -> Q17 191 tmp32 = WEBRTC_SPL_MUL_16_32_RSFT16(gain16, fQ15vec[n]); //Q(1+gain_sh)*Q15>>16 = Q(gain_sh) 192 sh = 9-gain_sh; //number of needed shifts to reach Q9 193 t16a = (int16_t) WEBRTC_SPL_SHIFT_W32(tmp32, sh); 194 lat_outQ9[n + temp1] = t16a; 195 } 196 197 /* save the states */ 198 for (i=0;i<ord_1;i++) 199 { 200 stateGQ15[i] = gQ15[i][HALF_SUBFRAMELEN-1]; 201 } 202 //process next frame 203 } 204 205 return; 206 } 207 208 209 210 211 212 /* ----------------AR filter-------------------------*/ 213 /* filter the signal using normalized lattice filter */ WebRtcIsacfix_NormLatticeFilterAr(size_t orderCoef,int16_t * stateGQ0,int32_t * lat_inQ25,int16_t * filt_coefQ15,int32_t * gain_lo_hiQ17,int16_t lo_hi,int16_t * lat_outQ0)214 void WebRtcIsacfix_NormLatticeFilterAr(size_t orderCoef, 215 int16_t *stateGQ0, 216 int32_t *lat_inQ25, 217 int16_t *filt_coefQ15, 218 int32_t *gain_lo_hiQ17, 219 int16_t lo_hi, 220 int16_t *lat_outQ0) 221 { 222 size_t ii, k, i; 223 int n, u; 224 int16_t sthQ15[MAX_AR_MODEL_ORDER]; 225 int16_t cthQ15[MAX_AR_MODEL_ORDER]; 226 int32_t tmp32; 227 228 229 int16_t tmpAR; 230 int16_t ARfQ0vec[HALF_SUBFRAMELEN]; 231 int16_t ARgQ0vec[MAX_AR_MODEL_ORDER+1]; 232 233 int32_t inv_gain32; 234 int16_t inv_gain16; 235 int16_t den16; 236 int16_t sh; 237 238 int16_t temp2,temp3; 239 size_t ord_1 = orderCoef+1; 240 241 for (u=0;u<SUBFRAMES;u++) 242 { 243 int32_t temp1 = u * HALF_SUBFRAMELEN; 244 245 //set the denominator and numerator of the Direct Form 246 temp2 = (int16_t)(u * orderCoef); 247 temp3 = (int16_t)(2 * u + lo_hi); 248 249 for (ii=0; ii<orderCoef; ii++) { 250 sthQ15[ii] = filt_coefQ15[temp2+ii]; 251 } 252 253 WebRtcSpl_SqrtOfOneMinusXSquared(sthQ15, orderCoef, cthQ15); 254 255 /* Simulation of the 25 files shows that maximum value in 256 the vector gain_lo_hiQ17[] is 441344, which means that 257 it is log2((2^31)/441344) = 12.2 shifting bits from 258 saturation. Therefore, it should be safe to use Q27 instead 259 of Q17. */ 260 261 tmp32 = gain_lo_hiQ17[temp3] << 10; // Q27 262 263 for (k=0;k<orderCoef;k++) { 264 tmp32 = WEBRTC_SPL_MUL_16_32_RSFT15(cthQ15[k], tmp32); // Q15*Q27>>15 = Q27 265 } 266 267 sh = WebRtcSpl_NormW32(tmp32); // tmp32 is the gain 268 den16 = (int16_t) WEBRTC_SPL_SHIFT_W32(tmp32, sh-16); //Q(27+sh-16) = Q(sh+11) (all 16 bits are value bits) 269 inv_gain32 = WebRtcSpl_DivW32W16((int32_t)2147483647, den16); // 1/gain in Q31/Q(sh+11) = Q(20-sh) 270 271 //initial conditions 272 inv_gain16 = (int16_t)(inv_gain32 >> 2); // 1/gain in Q(20-sh-2) = Q(18-sh) 273 274 for (i=0;i<HALF_SUBFRAMELEN;i++) 275 { 276 277 tmp32 = lat_inQ25[i + temp1] << 1; // Q25->Q26 278 tmp32 = WEBRTC_SPL_MUL_16_32_RSFT16(inv_gain16, tmp32); //lat_in[]*inv_gain in (Q(18-sh)*Q26)>>16 = Q(28-sh) 279 tmp32 = WEBRTC_SPL_SHIFT_W32(tmp32, -(28-sh)); // lat_in[]*inv_gain in Q0 280 281 ARfQ0vec[i] = (int16_t)WebRtcSpl_SatW32ToW16(tmp32); // Q0 282 } 283 284 // Get the state of f & g for the first input, for all orders. 285 for (i = orderCoef; i > 0; i--) 286 { 287 tmp32 = (cthQ15[i - 1] * ARfQ0vec[0] - sthQ15[i - 1] * stateGQ0[i - 1] + 288 16384) >> 15; 289 tmpAR = (int16_t)WebRtcSpl_SatW32ToW16(tmp32); // Q0 290 291 tmp32 = (sthQ15[i - 1] * ARfQ0vec[0] + cthQ15[i - 1] * stateGQ0[i - 1] + 292 16384) >> 15; 293 ARgQ0vec[i] = (int16_t)WebRtcSpl_SatW32ToW16(tmp32); // Q0 294 ARfQ0vec[0] = tmpAR; 295 } 296 ARgQ0vec[0] = ARfQ0vec[0]; 297 298 // Filter ARgQ0vec[] and ARfQ0vec[] through coefficients cthQ15[] and sthQ15[]. 299 WebRtcIsacfix_FilterArLoop(ARgQ0vec, ARfQ0vec, cthQ15, sthQ15, orderCoef); 300 301 for(n=0;n<HALF_SUBFRAMELEN;n++) 302 { 303 lat_outQ0[n + temp1] = ARfQ0vec[n]; 304 } 305 306 307 /* cannot use memcpy in the following */ 308 309 for (i=0;i<ord_1;i++) 310 { 311 stateGQ0[i] = ARgQ0vec[i]; 312 } 313 } 314 315 return; 316 } 317