1 /* 2 * Copyright (C) 2013 The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17 #ifndef ANDROID_AUDIO_RESAMPLER_FIR_PROCESS_H 18 #define ANDROID_AUDIO_RESAMPLER_FIR_PROCESS_H 19 20 namespace android { 21 22 // depends on AudioResamplerFirOps.h 23 24 /* variant for input type TI = int16_t input samples */ 25 template<typename TC> 26 static inline 27 void mac(int32_t& l, int32_t& r, TC coef, const int16_t* samples) 28 { 29 uint32_t rl = *reinterpret_cast<const uint32_t*>(samples); 30 l = mulAddRL(1, rl, coef, l); 31 r = mulAddRL(0, rl, coef, r); 32 } 33 34 template<typename TC> 35 static inline 36 void mac(int32_t& l, TC coef, const int16_t* samples) 37 { 38 l = mulAdd(samples[0], coef, l); 39 } 40 41 /* variant for input type TI = float input samples */ 42 template<typename TC> 43 static inline 44 void mac(float& l, float& r, TC coef, const float* samples) 45 { 46 l += *samples++ * coef; 47 r += *samples * coef; 48 } 49 50 template<typename TC> 51 static inline 52 void mac(float& l, TC coef, const float* samples) 53 { 54 l += *samples * coef; 55 } 56 57 /* variant for output type TO = int32_t output samples */ 58 static inline 59 int32_t volumeAdjust(int32_t value, int32_t volume) 60 { 61 return 2 * mulRL(0, value, volume); // Note: only use top 16b 62 } 63 64 /* variant for output type TO = float output samples */ 65 static inline 66 float volumeAdjust(float value, float volume) 67 { 68 return value * volume; 69 } 70 71 /* 72 * Helper template functions for loop unrolling accumulator operations. 73 * 74 * Unrolling the loops achieves about 2x gain. 75 * Using a recursive template rather than an array of TO[] for the accumulator 76 * values is an additional 10-20% gain. 77 */ 78 79 template<int CHANNELS, typename TO> 80 class Accumulator : public Accumulator<CHANNELS-1, TO> // recursive 81 { 82 public: 83 inline void clear() { 84 value = 0; 85 Accumulator<CHANNELS-1, TO>::clear(); 86 } 87 template<typename TC, typename TI> 88 inline void acc(TC coef, const TI*& data) { 89 mac(value, coef, data++); 90 Accumulator<CHANNELS-1, TO>::acc(coef, data); 91 } 92 inline void volume(TO*& out, TO gain) { 93 *out++ = volumeAdjust(value, gain); 94 Accumulator<CHANNELS-1, TO>::volume(out, gain); 95 } 96 97 TO value; // one per recursive inherited base class 98 }; 99 100 template<typename TO> 101 class Accumulator<0, TO> { 102 public: 103 inline void clear() { 104 } 105 template<typename TC, typename TI> 106 inline void acc(TC coef __unused, const TI*& data __unused) { 107 } 108 inline void volume(TO*& out __unused, TO gain __unused) { 109 } 110 }; 111 112 template<typename TC, typename TINTERP> 113 inline 114 TC interpolate(TC coef_0, TC coef_1, TINTERP lerp) 115 { 116 return lerp * (coef_1 - coef_0) + coef_0; 117 } 118 119 template<> 120 inline 121 int16_t interpolate<int16_t, uint32_t>(int16_t coef_0, int16_t coef_1, uint32_t lerp) 122 { // in some CPU architectures 16b x 16b multiplies are faster. 123 return (static_cast<int16_t>(lerp) * static_cast<int16_t>(coef_1 - coef_0) >> 15) + coef_0; 124 } 125 126 template<> 127 inline 128 int32_t interpolate<int32_t, uint32_t>(int32_t coef_0, int32_t coef_1, uint32_t lerp) 129 { 130 return (lerp * static_cast<int64_t>(coef_1 - coef_0) >> 31) + coef_0; 131 } 132 133 /* class scope for passing in functions into templates */ 134 struct InterpCompute { 135 template<typename TC, typename TINTERP> 136 static inline 137 TC interpolatep(TC coef_0, TC coef_1, TINTERP lerp) { 138 return interpolate(coef_0, coef_1, lerp); 139 } 140 141 template<typename TC, typename TINTERP> 142 static inline 143 TC interpolaten(TC coef_0, TC coef_1, TINTERP lerp) { 144 return interpolate(coef_0, coef_1, lerp); 145 } 146 }; 147 148 struct InterpNull { 149 template<typename TC, typename TINTERP> 150 static inline 151 TC interpolatep(TC coef_0, TC coef_1 __unused, TINTERP lerp __unused) { 152 return coef_0; 153 } 154 155 template<typename TC, typename TINTERP> 156 static inline 157 TC interpolaten(TC coef_0 __unused, TC coef_1, TINTERP lerp __unused) { 158 return coef_1; 159 } 160 }; 161 162 /* 163 * Calculates a single output frame (two samples). 164 * 165 * The Process*() functions compute both the positive half FIR dot product and 166 * the negative half FIR dot product, accumulates, and then applies the volume. 167 * 168 * Use fir() to compute the proper coefficient pointers for a polyphase 169 * filter bank. 170 * 171 * ProcessBase() is the fundamental processing template function. 172 * 173 * ProcessL() calls ProcessBase() with TFUNC = InterpNull, for fixed/locked phase. 174 * Process() calls ProcessBase() with TFUNC = InterpCompute, for interpolated phase. 175 */ 176 177 template <int CHANNELS, int STRIDE, typename TFUNC, typename TC, typename TI, typename TO, typename TINTERP> 178 static inline 179 void ProcessBase(TO* const out, 180 size_t count, 181 const TC* coefsP, 182 const TC* coefsN, 183 const TI* sP, 184 const TI* sN, 185 TINTERP lerpP, 186 const TO* const volumeLR) 187 { 188 COMPILE_TIME_ASSERT_FUNCTION_SCOPE(CHANNELS > 0) 189 190 if (CHANNELS > 2) { 191 // TO accum[CHANNELS]; 192 Accumulator<CHANNELS, TO> accum; 193 194 // for (int j = 0; j < CHANNELS; ++j) accum[j] = 0; 195 accum.clear(); 196 for (size_t i = 0; i < count; ++i) { 197 TC c = TFUNC::interpolatep(coefsP[0], coefsP[count], lerpP); 198 199 // for (int j = 0; j < CHANNELS; ++j) mac(accum[j], c, sP + j); 200 const TI *tmp_data = sP; // tmp_ptr seems to work better 201 accum.acc(c, tmp_data); 202 203 coefsP++; 204 sP -= CHANNELS; 205 c = TFUNC::interpolaten(coefsN[count], coefsN[0], lerpP); 206 207 // for (int j = 0; j < CHANNELS; ++j) mac(accum[j], c, sN + j); 208 tmp_data = sN; // tmp_ptr seems faster than directly using sN 209 accum.acc(c, tmp_data); 210 211 coefsN++; 212 sN += CHANNELS; 213 } 214 // for (int j = 0; j < CHANNELS; ++j) out[j] += volumeAdjust(accum[j], volumeLR[0]); 215 TO *tmp_out = out; // may remove if const out definition changes. 216 accum.volume(tmp_out, volumeLR[0]); 217 } else if (CHANNELS == 2) { 218 TO l = 0; 219 TO r = 0; 220 for (size_t i = 0; i < count; ++i) { 221 mac(l, r, TFUNC::interpolatep(coefsP[0], coefsP[count], lerpP), sP); 222 coefsP++; 223 sP -= CHANNELS; 224 mac(l, r, TFUNC::interpolaten(coefsN[count], coefsN[0], lerpP), sN); 225 coefsN++; 226 sN += CHANNELS; 227 } 228 out[0] += volumeAdjust(l, volumeLR[0]); 229 out[1] += volumeAdjust(r, volumeLR[1]); 230 } else { /* CHANNELS == 1 */ 231 TO l = 0; 232 for (size_t i = 0; i < count; ++i) { 233 mac(l, TFUNC::interpolatep(coefsP[0], coefsP[count], lerpP), sP); 234 coefsP++; 235 sP -= CHANNELS; 236 mac(l, TFUNC::interpolaten(coefsN[count], coefsN[0], lerpP), sN); 237 coefsN++; 238 sN += CHANNELS; 239 } 240 out[0] += volumeAdjust(l, volumeLR[0]); 241 out[1] += volumeAdjust(l, volumeLR[1]); 242 } 243 } 244 245 template <int CHANNELS, int STRIDE, typename TC, typename TI, typename TO> 246 static inline 247 void ProcessL(TO* const out, 248 int count, 249 const TC* coefsP, 250 const TC* coefsN, 251 const TI* sP, 252 const TI* sN, 253 const TO* const volumeLR) 254 { 255 ProcessBase<CHANNELS, STRIDE, InterpNull>(out, count, coefsP, coefsN, sP, sN, 0, volumeLR); 256 } 257 258 template <int CHANNELS, int STRIDE, typename TC, typename TI, typename TO, typename TINTERP> 259 static inline 260 void Process(TO* const out, 261 int count, 262 const TC* coefsP, 263 const TC* coefsN, 264 const TC* coefsP1 __unused, 265 const TC* coefsN1 __unused, 266 const TI* sP, 267 const TI* sN, 268 TINTERP lerpP, 269 const TO* const volumeLR) 270 { 271 ProcessBase<CHANNELS, STRIDE, InterpCompute>(out, count, coefsP, coefsN, sP, sN, lerpP, volumeLR); 272 } 273 274 /* 275 * Calculates a single output frame (two samples) from input sample pointer. 276 * 277 * This sets up the params for the accelerated Process() and ProcessL() 278 * functions to do the appropriate dot products. 279 * 280 * @param out should point to the output buffer with space for at least one output frame. 281 * 282 * @param phase is the fractional distance between input frames for interpolation: 283 * phase >= 0 && phase < phaseWrapLimit. It can be thought of as a rational fraction 284 * of phase/phaseWrapLimit. 285 * 286 * @param phaseWrapLimit is #polyphases<<coefShift, where #polyphases is the number of polyphases 287 * in the polyphase filter. Likewise, #polyphases can be obtained as (phaseWrapLimit>>coefShift). 288 * 289 * @param coefShift gives the bit alignment of the polyphase index in the phase parameter. 290 * 291 * @param halfNumCoefs is the half the number of coefficients per polyphase filter. Since the 292 * overall filterbank is odd-length symmetric, only halfNumCoefs need be stored. 293 * 294 * @param coefs is the polyphase filter bank, starting at from polyphase index 0, and ranging to 295 * and including the #polyphases. Each polyphase of the filter has half-length halfNumCoefs 296 * (due to symmetry). The total size of the filter bank in coefficients is 297 * (#polyphases+1)*halfNumCoefs. 298 * 299 * The filter bank coefs should be aligned to a minimum of 16 bytes (preferrably to cache line). 300 * 301 * The coefs should be attenuated (to compensate for passband ripple) 302 * if storing back into the native format. 303 * 304 * @param samples are unaligned input samples. The position is in the "middle" of the 305 * sample array with respect to the FIR filter: 306 * the negative half of the filter is dot product from samples+1 to samples+halfNumCoefs; 307 * the positive half of the filter is dot product from samples to samples-halfNumCoefs+1. 308 * 309 * @param volumeLR is a pointer to an array of two 32 bit volume values, one per stereo channel, 310 * expressed as a S32 integer. A negative value inverts the channel 180 degrees. 311 * The pointer volumeLR should be aligned to a minimum of 8 bytes. 312 * A typical value for volume is 0x1000 to align to a unity gain output of 20.12. 313 * 314 * In between calls to filterCoefficient, the phase is incremented by phaseIncrement, where 315 * phaseIncrement is calculated as inputSampling * phaseWrapLimit / outputSampling. 316 * 317 * The filter polyphase index is given by indexP = phase >> coefShift. Due to 318 * odd length symmetric filter, the polyphase index of the negative half depends on 319 * whether interpolation is used. 320 * 321 * The fractional siting between the polyphase indices is given by the bits below coefShift: 322 * 323 * lerpP = phase << 32 - coefShift >> 1; // for 32 bit unsigned phase multiply 324 * lerpP = phase << 32 - coefShift >> 17; // for 16 bit unsigned phase multiply 325 * 326 * For integer types, this is expressed as: 327 * 328 * lerpP = phase << sizeof(phase)*8 - coefShift 329 * >> (sizeof(phase)-sizeof(*coefs))*8 + 1; 330 * 331 * For floating point, lerpP is the fractional phase scaled to [0.0, 1.0): 332 * 333 * lerpP = (phase << 32 - coefShift) / (1 << 32); // floating point equivalent 334 */ 335 336 template<int CHANNELS, bool LOCKED, int STRIDE, typename TC, typename TI, typename TO> 337 static inline 338 void fir(TO* const out, 339 const uint32_t phase, const uint32_t phaseWrapLimit, 340 const int coefShift, const int halfNumCoefs, const TC* const coefs, 341 const TI* const samples, const TO* const volumeLR) 342 { 343 // NOTE: be very careful when modifying the code here. register 344 // pressure is very high and a small change might cause the compiler 345 // to generate far less efficient code. 346 // Always sanity check the result with objdump or test-resample. 347 348 if (LOCKED) { 349 // locked polyphase (no interpolation) 350 // Compute the polyphase filter index on the positive and negative side. 351 uint32_t indexP = phase >> coefShift; 352 uint32_t indexN = (phaseWrapLimit - phase) >> coefShift; 353 const TC* coefsP = coefs + indexP*halfNumCoefs; 354 const TC* coefsN = coefs + indexN*halfNumCoefs; 355 const TI* sP = samples; 356 const TI* sN = samples + CHANNELS; 357 358 // dot product filter. 359 ProcessL<CHANNELS, STRIDE>(out, 360 halfNumCoefs, coefsP, coefsN, sP, sN, volumeLR); 361 } else { 362 // interpolated polyphase 363 // Compute the polyphase filter index on the positive and negative side. 364 uint32_t indexP = phase >> coefShift; 365 uint32_t indexN = (phaseWrapLimit - phase - 1) >> coefShift; // one's complement. 366 const TC* coefsP = coefs + indexP*halfNumCoefs; 367 const TC* coefsN = coefs + indexN*halfNumCoefs; 368 const TC* coefsP1 = coefsP + halfNumCoefs; 369 const TC* coefsN1 = coefsN + halfNumCoefs; 370 const TI* sP = samples; 371 const TI* sN = samples + CHANNELS; 372 373 // Interpolation fraction lerpP derived by shifting all the way up and down 374 // to clear the appropriate bits and align to the appropriate level 375 // for the integer multiply. The constants should resolve in compile time. 376 // 377 // The interpolated filter coefficient is derived as follows for the pos/neg half: 378 // 379 // interpolated[P] = index[P]*lerpP + index[P+1]*(1-lerpP) 380 // interpolated[N] = index[N+1]*lerpP + index[N]*(1-lerpP) 381 382 // on-the-fly interpolated dot product filter 383 if (is_same<TC, float>::value || is_same<TC, double>::value) { 384 static const TC scale = 1. / (65536. * 65536.); // scale phase bits to [0.0, 1.0) 385 TC lerpP = TC(phase << (sizeof(phase)*8 - coefShift)) * scale; 386 387 Process<CHANNELS, STRIDE>(out, 388 halfNumCoefs, coefsP, coefsN, coefsP1, coefsN1, sP, sN, lerpP, volumeLR); 389 } else { 390 uint32_t lerpP = phase << (sizeof(phase)*8 - coefShift) 391 >> ((sizeof(phase)-sizeof(*coefs))*8 + 1); 392 393 Process<CHANNELS, STRIDE>(out, 394 halfNumCoefs, coefsP, coefsN, coefsP1, coefsN1, sP, sN, lerpP, volumeLR); 395 } 396 } 397 } 398 399 }; // namespace android 400 401 #endif /*ANDROID_AUDIO_RESAMPLER_FIR_PROCESS_H*/ 402
