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 ((WebRtc_Word32)(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 WebRtc_Word32*WebRtc_Word32 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 WebRtc_Word32. In the 28 cases used in here, the WebRtc_Word32 will be enough, since (for a good 29 reason) the involved multiplicands aren't big enough to overflow a 30 WebRtc_Word32 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 (WebRtc_Word32) (((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 int16_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 */ 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 // Declare a function pointer. 88 FilterMaLoopFix WebRtcIsacfix_FilterMaLoopFix; 89 90 /* filter the signal using normalized lattice filter */ 91 /* MA filter */ 92 void WebRtcIsacfix_NormLatticeFilterMa(WebRtc_Word16 orderCoef, 93 WebRtc_Word32 *stateGQ15, 94 WebRtc_Word16 *lat_inQ0, 95 WebRtc_Word16 *filt_coefQ15, 96 WebRtc_Word32 *gain_lo_hiQ17, 97 WebRtc_Word16 lo_hi, 98 WebRtc_Word16 *lat_outQ9) 99 { 100 WebRtc_Word16 sthQ15[MAX_AR_MODEL_ORDER]; 101 WebRtc_Word16 cthQ15[MAX_AR_MODEL_ORDER]; 102 103 int u, i, k, n; 104 WebRtc_Word16 temp2,temp3; 105 WebRtc_Word16 ord_1 = orderCoef+1; 106 WebRtc_Word32 inv_cthQ16[MAX_AR_MODEL_ORDER]; 107 108 WebRtc_Word32 gain32, fQtmp; 109 WebRtc_Word16 gain16; 110 WebRtc_Word16 gain_sh; 111 112 WebRtc_Word32 tmp32, tmp32b; 113 WebRtc_Word32 fQ15vec[HALF_SUBFRAMELEN]; 114 WebRtc_Word32 gQ15[MAX_AR_MODEL_ORDER+1][HALF_SUBFRAMELEN]; 115 WebRtc_Word16 sh; 116 WebRtc_Word16 t16a; 117 WebRtc_Word16 t16b; 118 119 for (u=0;u<SUBFRAMES;u++) 120 { 121 int32_t temp1 = WEBRTC_SPL_MUL_16_16(u, HALF_SUBFRAMELEN); 122 123 /* set the Direct Form coefficients */ 124 temp2 = (WebRtc_Word16)WEBRTC_SPL_MUL_16_16(u, orderCoef); 125 temp3 = (WebRtc_Word16)WEBRTC_SPL_MUL_16_16(2, u)+lo_hi; 126 127 /* compute lattice filter coefficients */ 128 memcpy(sthQ15, &filt_coefQ15[temp2], orderCoef * sizeof(WebRtc_Word16)); 129 130 WebRtcSpl_SqrtOfOneMinusXSquared(sthQ15, orderCoef, cthQ15); 131 132 /* compute the gain */ 133 gain32 = gain_lo_hiQ17[temp3]; 134 gain_sh = WebRtcSpl_NormW32(gain32); 135 gain32 = WEBRTC_SPL_LSHIFT_W32(gain32, gain_sh); //Q(17+gain_sh) 136 137 for (k=0;k<orderCoef;k++) 138 { 139 gain32 = WEBRTC_SPL_MUL_16_32_RSFT15(cthQ15[k], gain32); //Q15*Q(17+gain_sh)>>15 = Q(17+gain_sh) 140 inv_cthQ16[k] = WebRtcSpl_DivW32W16((WebRtc_Word32)2147483647, cthQ15[k]); // 1/cth[k] in Q31/Q15 = Q16 141 } 142 gain16 = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(gain32, 16); //Q(1+gain_sh) 143 144 /* normalized lattice filter */ 145 /*****************************/ 146 147 /* initial conditions */ 148 for (i=0;i<HALF_SUBFRAMELEN;i++) 149 { 150 fQ15vec[i] = WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)lat_inQ0[i + temp1], 15); //Q15 151 gQ15[0][i] = WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)lat_inQ0[i + temp1], 15); //Q15 152 } 153 154 155 fQtmp = fQ15vec[0]; 156 157 /* get the state of f&g for the first input, for all orders */ 158 for (i=1;i<ord_1;i++) 159 { 160 // Calculate f[i][0] = inv_cth[i-1]*(f[i-1][0] + sth[i-1]*stateG[i-1]); 161 tmp32 = WEBRTC_SPL_MUL_16_32_RSFT15(sthQ15[i-1], stateGQ15[i-1]);//Q15*Q15>>15 = Q15 162 tmp32b= fQtmp + tmp32; //Q15+Q15=Q15 163 tmp32 = inv_cthQ16[i-1]; //Q16 164 t16a = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(tmp32, 16); 165 t16b = (WebRtc_Word16) (tmp32-WEBRTC_SPL_LSHIFT_W32(((WebRtc_Word32)t16a), 16)); 166 if (t16b<0) t16a++; 167 tmp32 = LATTICE_MUL_32_32_RSFT16(t16a, t16b, tmp32b); 168 fQtmp = tmp32; // Q15 169 170 // Calculate g[i][0] = cth[i-1]*stateG[i-1] + sth[i-1]* f[i][0]; 171 tmp32 = WEBRTC_SPL_MUL_16_32_RSFT15(cthQ15[i-1], stateGQ15[i-1]); //Q15*Q15>>15 = Q15 172 tmp32b = WEBRTC_SPL_MUL_16_32_RSFT15(sthQ15[i-1], fQtmp); //Q15*Q15>>15 = Q15 173 tmp32 = tmp32 + tmp32b;//Q15+Q15 = Q15 174 gQ15[i][0] = tmp32; // Q15 175 } 176 177 /* filtering */ 178 /* save the states */ 179 for(k=0;k<orderCoef;k++) 180 { 181 // for 0 <= n < HALF_SUBFRAMELEN - 1: 182 // f[k+1][n+1] = inv_cth[k]*(f[k][n+1] + sth[k]*g[k][n]); 183 // g[k+1][n+1] = cth[k]*g[k][n] + sth[k]* f[k+1][n+1]; 184 WebRtcIsacfix_FilterMaLoopFix(sthQ15[k], cthQ15[k], inv_cthQ16[k], 185 &gQ15[k][0], &gQ15[k+1][1], &fQ15vec[1]); 186 } 187 188 fQ15vec[0] = fQtmp; 189 190 for(n=0;n<HALF_SUBFRAMELEN;n++) 191 { 192 //gain32 = WEBRTC_SPL_RSHIFT_W32(gain32, gain_sh); // Q(17+gain_sh) -> Q17 193 tmp32 = WEBRTC_SPL_MUL_16_32_RSFT16(gain16, fQ15vec[n]); //Q(1+gain_sh)*Q15>>16 = Q(gain_sh) 194 sh = 9-gain_sh; //number of needed shifts to reach Q9 195 t16a = (WebRtc_Word16) WEBRTC_SPL_SHIFT_W32(tmp32, sh); 196 lat_outQ9[n + temp1] = t16a; 197 } 198 199 /* save the states */ 200 for (i=0;i<ord_1;i++) 201 { 202 stateGQ15[i] = gQ15[i][HALF_SUBFRAMELEN-1]; 203 } 204 //process next frame 205 } 206 207 return; 208 } 209 210 211 212 213 214 /* ----------------AR filter-------------------------*/ 215 /* filter the signal using normalized lattice filter */ 216 void WebRtcIsacfix_NormLatticeFilterAr(WebRtc_Word16 orderCoef, 217 WebRtc_Word16 *stateGQ0, 218 WebRtc_Word32 *lat_inQ25, 219 WebRtc_Word16 *filt_coefQ15, 220 WebRtc_Word32 *gain_lo_hiQ17, 221 WebRtc_Word16 lo_hi, 222 WebRtc_Word16 *lat_outQ0) 223 { 224 int ii,n,k,i,u; 225 WebRtc_Word16 sthQ15[MAX_AR_MODEL_ORDER]; 226 WebRtc_Word16 cthQ15[MAX_AR_MODEL_ORDER]; 227 WebRtc_Word32 tmp32; 228 229 230 WebRtc_Word16 tmpAR; 231 WebRtc_Word16 ARfQ0vec[HALF_SUBFRAMELEN]; 232 WebRtc_Word16 ARgQ0vec[MAX_AR_MODEL_ORDER+1]; 233 234 WebRtc_Word32 inv_gain32; 235 WebRtc_Word16 inv_gain16; 236 WebRtc_Word16 den16; 237 WebRtc_Word16 sh; 238 239 WebRtc_Word16 temp2,temp3; 240 WebRtc_Word16 ord_1 = orderCoef+1; 241 242 for (u=0;u<SUBFRAMES;u++) 243 { 244 int32_t temp1 = WEBRTC_SPL_MUL_16_16(u, HALF_SUBFRAMELEN); 245 246 //set the denominator and numerator of the Direct Form 247 temp2 = (WebRtc_Word16)WEBRTC_SPL_MUL_16_16(u, orderCoef); 248 temp3 = (WebRtc_Word16)WEBRTC_SPL_MUL_16_16(2, u) + lo_hi; 249 250 for (ii=0; ii<orderCoef; ii++) { 251 sthQ15[ii] = filt_coefQ15[temp2+ii]; 252 } 253 254 WebRtcSpl_SqrtOfOneMinusXSquared(sthQ15, orderCoef, cthQ15); 255 256 /* Simulation of the 25 files shows that maximum value in 257 the vector gain_lo_hiQ17[] is 441344, which means that 258 it is log2((2^31)/441344) = 12.2 shifting bits from 259 saturation. Therefore, it should be safe to use Q27 instead 260 of Q17. */ 261 262 tmp32 = WEBRTC_SPL_LSHIFT_W32(gain_lo_hiQ17[temp3], 10); // Q27 263 264 for (k=0;k<orderCoef;k++) { 265 tmp32 = WEBRTC_SPL_MUL_16_32_RSFT15(cthQ15[k], tmp32); // Q15*Q27>>15 = Q27 266 } 267 268 sh = WebRtcSpl_NormW32(tmp32); // tmp32 is the gain 269 den16 = (WebRtc_Word16) WEBRTC_SPL_SHIFT_W32(tmp32, sh-16); //Q(27+sh-16) = Q(sh+11) (all 16 bits are value bits) 270 inv_gain32 = WebRtcSpl_DivW32W16((WebRtc_Word32)2147483647, den16); // 1/gain in Q31/Q(sh+11) = Q(20-sh) 271 272 //initial conditions 273 inv_gain16 = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(inv_gain32, 2); // 1/gain in Q(20-sh-2) = Q(18-sh) 274 275 for (i=0;i<HALF_SUBFRAMELEN;i++) 276 { 277 278 tmp32 = WEBRTC_SPL_LSHIFT_W32(lat_inQ25[i + temp1], 1); //Q25->Q26 279 tmp32 = WEBRTC_SPL_MUL_16_32_RSFT16(inv_gain16, tmp32); //lat_in[]*inv_gain in (Q(18-sh)*Q26)>>16 = Q(28-sh) 280 tmp32 = WEBRTC_SPL_SHIFT_W32(tmp32, -(28-sh)); // lat_in[]*inv_gain in Q0 281 282 ARfQ0vec[i] = (WebRtc_Word16)WebRtcSpl_SatW32ToW16(tmp32); // Q0 283 } 284 285 for (i=orderCoef-1;i>=0;i--) //get the state of f&g for the first input, for all orders 286 { 287 tmp32 = WEBRTC_SPL_RSHIFT_W32(((WEBRTC_SPL_MUL_16_16(cthQ15[i],ARfQ0vec[0])) - (WEBRTC_SPL_MUL_16_16(sthQ15[i],stateGQ0[i])) + 16384), 15); 288 tmpAR = (WebRtc_Word16)WebRtcSpl_SatW32ToW16(tmp32); // Q0 289 290 tmp32 = WEBRTC_SPL_RSHIFT_W32(((WEBRTC_SPL_MUL_16_16(sthQ15[i],ARfQ0vec[0])) + (WEBRTC_SPL_MUL_16_16(cthQ15[i], stateGQ0[i])) + 16384), 15); 291 ARgQ0vec[i+1] = (WebRtc_Word16)WebRtcSpl_SatW32ToW16(tmp32); // Q0 292 ARfQ0vec[0] = tmpAR; 293 } 294 ARgQ0vec[0] = ARfQ0vec[0]; 295 296 // Filter ARgQ0vec[] and ARfQ0vec[] through coefficients cthQ15[] and sthQ15[]. 297 WebRtcIsacfix_FilterArLoop(ARgQ0vec, ARfQ0vec, cthQ15, sthQ15, orderCoef); 298 299 for(n=0;n<HALF_SUBFRAMELEN;n++) 300 { 301 lat_outQ0[n + temp1] = ARfQ0vec[n]; 302 } 303 304 305 /* cannot use memcpy in the following */ 306 307 for (i=0;i<ord_1;i++) 308 { 309 stateGQ0[i] = ARgQ0vec[i]; 310 } 311 } 312 313 return; 314 } 315
