Home | History | Annotate | Download | only in libaudioprocessing
      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 #define LOG_TAG "AudioResamplerDyn"
     18 //#define LOG_NDEBUG 0
     19 
     20 #include <malloc.h>
     21 #include <string.h>
     22 #include <stdlib.h>
     23 #include <dlfcn.h>
     24 #include <math.h>
     25 
     26 #include <cutils/compiler.h>
     27 #include <cutils/properties.h>
     28 #include <utils/Debug.h>
     29 #include <utils/Log.h>
     30 #include <audio_utils/primitives.h>
     31 
     32 #include "AudioResamplerFirOps.h" // USE_NEON, USE_SSE and USE_INLINE_ASSEMBLY defined here
     33 #include "AudioResamplerFirProcess.h"
     34 #include "AudioResamplerFirProcessNeon.h"
     35 #include "AudioResamplerFirProcessSSE.h"
     36 #include "AudioResamplerFirGen.h" // requires math.h
     37 #include "AudioResamplerDyn.h"
     38 
     39 //#define DEBUG_RESAMPLER
     40 
     41 // use this for our buffer alignment.  Should be at least 32 bytes.
     42 constexpr size_t CACHE_LINE_SIZE = 64;
     43 
     44 namespace android {
     45 
     46 /*
     47  * InBuffer is a type agnostic input buffer.
     48  *
     49  * Layout of the state buffer for halfNumCoefs=8.
     50  *
     51  * [rrrrrrppppppppnnnnnnnnrrrrrrrrrrrrrrrrrrr.... rrrrrrr]
     52  *  S            I                                R
     53  *
     54  * S = mState
     55  * I = mImpulse
     56  * R = mRingFull
     57  * p = past samples, convoluted with the (p)ositive side of sinc()
     58  * n = future samples, convoluted with the (n)egative side of sinc()
     59  * r = extra space for implementing the ring buffer
     60  */
     61 
     62 template<typename TC, typename TI, typename TO>
     63 AudioResamplerDyn<TC, TI, TO>::InBuffer::InBuffer()
     64     : mState(NULL), mImpulse(NULL), mRingFull(NULL), mStateCount(0)
     65 {
     66 }
     67 
     68 template<typename TC, typename TI, typename TO>
     69 AudioResamplerDyn<TC, TI, TO>::InBuffer::~InBuffer()
     70 {
     71     init();
     72 }
     73 
     74 template<typename TC, typename TI, typename TO>
     75 void AudioResamplerDyn<TC, TI, TO>::InBuffer::init()
     76 {
     77     free(mState);
     78     mState = NULL;
     79     mImpulse = NULL;
     80     mRingFull = NULL;
     81     mStateCount = 0;
     82 }
     83 
     84 // resizes the state buffer to accommodate the appropriate filter length
     85 template<typename TC, typename TI, typename TO>
     86 void AudioResamplerDyn<TC, TI, TO>::InBuffer::resize(int CHANNELS, int halfNumCoefs)
     87 {
     88     // calculate desired state size
     89     size_t stateCount = halfNumCoefs * CHANNELS * 2 * kStateSizeMultipleOfFilterLength;
     90 
     91     // check if buffer needs resizing
     92     if (mState
     93             && stateCount == mStateCount
     94             && mRingFull-mState == (ssize_t) (mStateCount-halfNumCoefs*CHANNELS)) {
     95         return;
     96     }
     97 
     98     // create new buffer
     99     TI* state = NULL;
    100     (void)posix_memalign(
    101             reinterpret_cast<void **>(&state),
    102             CACHE_LINE_SIZE /* alignment */,
    103             stateCount * sizeof(*state));
    104     memset(state, 0, stateCount*sizeof(*state));
    105 
    106     // attempt to preserve state
    107     if (mState) {
    108         TI* srcLo = mImpulse - halfNumCoefs*CHANNELS;
    109         TI* srcHi = mImpulse + halfNumCoefs*CHANNELS;
    110         TI* dst = state;
    111 
    112         if (srcLo < mState) {
    113             dst += mState-srcLo;
    114             srcLo = mState;
    115         }
    116         if (srcHi > mState + mStateCount) {
    117             srcHi = mState + mStateCount;
    118         }
    119         memcpy(dst, srcLo, (srcHi - srcLo) * sizeof(*srcLo));
    120         free(mState);
    121     }
    122 
    123     // set class member vars
    124     mState = state;
    125     mStateCount = stateCount;
    126     mImpulse = state + halfNumCoefs*CHANNELS; // actually one sample greater than needed
    127     mRingFull = state + mStateCount - halfNumCoefs*CHANNELS;
    128 }
    129 
    130 // copy in the input data into the head (impulse+halfNumCoefs) of the buffer.
    131 template<typename TC, typename TI, typename TO>
    132 template<int CHANNELS>
    133 void AudioResamplerDyn<TC, TI, TO>::InBuffer::readAgain(TI*& impulse, const int halfNumCoefs,
    134         const TI* const in, const size_t inputIndex)
    135 {
    136     TI* head = impulse + halfNumCoefs*CHANNELS;
    137     for (size_t i=0 ; i<CHANNELS ; i++) {
    138         head[i] = in[inputIndex*CHANNELS + i];
    139     }
    140 }
    141 
    142 // advance the impulse pointer, and load in data into the head (impulse+halfNumCoefs)
    143 template<typename TC, typename TI, typename TO>
    144 template<int CHANNELS>
    145 void AudioResamplerDyn<TC, TI, TO>::InBuffer::readAdvance(TI*& impulse, const int halfNumCoefs,
    146         const TI* const in, const size_t inputIndex)
    147 {
    148     impulse += CHANNELS;
    149 
    150     if (CC_UNLIKELY(impulse >= mRingFull)) {
    151         const size_t shiftDown = mRingFull - mState - halfNumCoefs*CHANNELS;
    152         memcpy(mState, mState+shiftDown, halfNumCoefs*CHANNELS*2*sizeof(TI));
    153         impulse -= shiftDown;
    154     }
    155     readAgain<CHANNELS>(impulse, halfNumCoefs, in, inputIndex);
    156 }
    157 
    158 template<typename TC, typename TI, typename TO>
    159 void AudioResamplerDyn<TC, TI, TO>::InBuffer::reset()
    160 {
    161     // clear resampler state
    162     if (mState != nullptr) {
    163         memset(mState, 0, mStateCount * sizeof(TI));
    164     }
    165 }
    166 
    167 template<typename TC, typename TI, typename TO>
    168 void AudioResamplerDyn<TC, TI, TO>::Constants::set(
    169         int L, int halfNumCoefs, int inSampleRate, int outSampleRate)
    170 {
    171     int bits = 0;
    172     int lscale = inSampleRate/outSampleRate < 2 ? L - 1 :
    173             static_cast<int>(static_cast<uint64_t>(L)*inSampleRate/outSampleRate);
    174     for (int i=lscale; i; ++bits, i>>=1)
    175         ;
    176     mL = L;
    177     mShift = kNumPhaseBits - bits;
    178     mHalfNumCoefs = halfNumCoefs;
    179 }
    180 
    181 template<typename TC, typename TI, typename TO>
    182 AudioResamplerDyn<TC, TI, TO>::AudioResamplerDyn(
    183         int inChannelCount, int32_t sampleRate, src_quality quality)
    184     : AudioResampler(inChannelCount, sampleRate, quality),
    185       mResampleFunc(0), mFilterSampleRate(0), mFilterQuality(DEFAULT_QUALITY),
    186     mCoefBuffer(NULL)
    187 {
    188     mVolumeSimd[0] = mVolumeSimd[1] = 0;
    189     // The AudioResampler base class assumes we are always ready for 1:1 resampling.
    190     // We reset mInSampleRate to 0, so setSampleRate() will calculate filters for
    191     // setSampleRate() for 1:1. (May be removed if precalculated filters are used.)
    192     mInSampleRate = 0;
    193     mConstants.set(128, 8, mSampleRate, mSampleRate); // TODO: set better
    194 
    195     // fetch property based resampling parameters
    196     mPropertyEnableAtSampleRate = property_get_int32(
    197             "ro.audio.resampler.psd.enable_at_samplerate", mPropertyEnableAtSampleRate);
    198     mPropertyHalfFilterLength = property_get_int32(
    199             "ro.audio.resampler.psd.halflength", mPropertyHalfFilterLength);
    200     mPropertyStopbandAttenuation = property_get_int32(
    201             "ro.audio.resampler.psd.stopband", mPropertyStopbandAttenuation);
    202     mPropertyCutoffPercent = property_get_int32(
    203             "ro.audio.resampler.psd.cutoff_percent", mPropertyCutoffPercent);
    204     mPropertyTransitionBandwidthCheat = property_get_int32(
    205             "ro.audio.resampler.psd.tbwcheat", mPropertyTransitionBandwidthCheat);
    206 }
    207 
    208 template<typename TC, typename TI, typename TO>
    209 AudioResamplerDyn<TC, TI, TO>::~AudioResamplerDyn()
    210 {
    211     free(mCoefBuffer);
    212 }
    213 
    214 template<typename TC, typename TI, typename TO>
    215 void AudioResamplerDyn<TC, TI, TO>::init()
    216 {
    217     mFilterSampleRate = 0; // always trigger new filter generation
    218     mInBuffer.init();
    219 }
    220 
    221 template<typename TC, typename TI, typename TO>
    222 void AudioResamplerDyn<TC, TI, TO>::setVolume(float left, float right)
    223 {
    224     AudioResampler::setVolume(left, right);
    225     if (is_same<TO, float>::value || is_same<TO, double>::value) {
    226         mVolumeSimd[0] = static_cast<TO>(left);
    227         mVolumeSimd[1] = static_cast<TO>(right);
    228     } else {  // integer requires scaling to U4_28 (rounding down)
    229         // integer volumes are clamped to 0 to UNITY_GAIN so there
    230         // are no issues with signed overflow.
    231         mVolumeSimd[0] = u4_28_from_float(clampFloatVol(left));
    232         mVolumeSimd[1] = u4_28_from_float(clampFloatVol(right));
    233     }
    234 }
    235 
    236 // TODO: update to C++11
    237 
    238 template<typename T> T max(T a, T b) {return a > b ? a : b;}
    239 
    240 template<typename T> T absdiff(T a, T b) {return a > b ? a - b : b - a;}
    241 
    242 template<typename TC, typename TI, typename TO>
    243 void AudioResamplerDyn<TC, TI, TO>::createKaiserFir(Constants &c,
    244         double stopBandAtten, int inSampleRate, int outSampleRate, double tbwCheat)
    245 {
    246     // compute the normalized transition bandwidth
    247     const double tbw = firKaiserTbw(c.mHalfNumCoefs, stopBandAtten);
    248     const double halfbw = tbw * 0.5;
    249 
    250     double fcr; // compute fcr, the 3 dB amplitude cut-off.
    251     if (inSampleRate < outSampleRate) { // upsample
    252         fcr = max(0.5 * tbwCheat - halfbw, halfbw);
    253     } else { // downsample
    254         fcr = max(0.5 * tbwCheat * outSampleRate / inSampleRate - halfbw, halfbw);
    255     }
    256     createKaiserFir(c, stopBandAtten, fcr);
    257 }
    258 
    259 template<typename TC, typename TI, typename TO>
    260 void AudioResamplerDyn<TC, TI, TO>::createKaiserFir(Constants &c,
    261         double stopBandAtten, double fcr) {
    262     // compute the normalized transition bandwidth
    263     const double tbw = firKaiserTbw(c.mHalfNumCoefs, stopBandAtten);
    264     const int phases = c.mL;
    265     const int halfLength = c.mHalfNumCoefs;
    266 
    267     // create buffer
    268     TC *coefs = nullptr;
    269     int ret = posix_memalign(
    270             reinterpret_cast<void **>(&coefs),
    271             CACHE_LINE_SIZE /* alignment */,
    272             (phases + 1) * halfLength * sizeof(TC));
    273     LOG_ALWAYS_FATAL_IF(ret != 0, "Cannot allocate buffer memory, ret %d", ret);
    274     c.mFirCoefs = coefs;
    275     free(mCoefBuffer);
    276     mCoefBuffer = coefs;
    277 
    278     // square the computed minimum passband value (extra safety).
    279     double attenuation =
    280             computeWindowedSincMinimumPassbandValue(stopBandAtten);
    281     attenuation *= attenuation;
    282 
    283     // design filter
    284     firKaiserGen(coefs, phases, halfLength, stopBandAtten, fcr, attenuation);
    285 
    286     // update the design criteria
    287     mNormalizedCutoffFrequency = fcr;
    288     mNormalizedTransitionBandwidth = tbw;
    289     mFilterAttenuation = attenuation;
    290     mStopbandAttenuationDb = stopBandAtten;
    291     mPassbandRippleDb = computeWindowedSincPassbandRippleDb(stopBandAtten);
    292 
    293 #if 0
    294     // Keep this debug code in case an app causes resampler design issues.
    295     const double halfbw = tbw * 0.5;
    296     // print basic filter stats
    297     ALOGD("L:%d  hnc:%d  stopBandAtten:%lf  fcr:%lf  atten:%lf  tbw:%lf\n",
    298             c.mL, c.mHalfNumCoefs, stopBandAtten, fcr, attenuation, tbw);
    299 
    300     // test the filter and report results.
    301     // Since this is a polyphase filter, normalized fp and fs must be scaled.
    302     const double fp = (fcr - halfbw) / phases;
    303     const double fs = (fcr + halfbw) / phases;
    304 
    305     double passMin, passMax, passRipple;
    306     double stopMax, stopRipple;
    307 
    308     const int32_t passSteps = 1000;
    309 
    310     testFir(coefs, c.mL, c.mHalfNumCoefs, fp, fs, passSteps, passSteps * c.mL /*stopSteps*/,
    311             passMin, passMax, passRipple, stopMax, stopRipple);
    312     ALOGD("passband(%lf, %lf): %.8lf %.8lf %.8lf\n", 0., fp, passMin, passMax, passRipple);
    313     ALOGD("stopband(%lf, %lf): %.8lf %.3lf\n", fs, 0.5, stopMax, stopRipple);
    314 #endif
    315 }
    316 
    317 // recursive gcd. Using objdump, it appears the tail recursion is converted to a while loop.
    318 static int gcd(int n, int m)
    319 {
    320     if (m == 0) {
    321         return n;
    322     }
    323     return gcd(m, n % m);
    324 }
    325 
    326 static bool isClose(int32_t newSampleRate, int32_t prevSampleRate,
    327         int32_t filterSampleRate, int32_t outSampleRate)
    328 {
    329 
    330     // different upsampling ratios do not need a filter change.
    331     if (filterSampleRate != 0
    332             && filterSampleRate < outSampleRate
    333             && newSampleRate < outSampleRate)
    334         return true;
    335 
    336     // check design criteria again if downsampling is detected.
    337     int pdiff = absdiff(newSampleRate, prevSampleRate);
    338     int adiff = absdiff(newSampleRate, filterSampleRate);
    339 
    340     // allow up to 6% relative change increments.
    341     // allow up to 12% absolute change increments (from filter design)
    342     return pdiff < prevSampleRate>>4 && adiff < filterSampleRate>>3;
    343 }
    344 
    345 template<typename TC, typename TI, typename TO>
    346 void AudioResamplerDyn<TC, TI, TO>::setSampleRate(int32_t inSampleRate)
    347 {
    348     if (mInSampleRate == inSampleRate) {
    349         return;
    350     }
    351     int32_t oldSampleRate = mInSampleRate;
    352     uint32_t oldPhaseWrapLimit = mConstants.mL << mConstants.mShift;
    353     bool useS32 = false;
    354 
    355     mInSampleRate = inSampleRate;
    356 
    357     // TODO: Add precalculated Equiripple filters
    358 
    359     if (mFilterQuality != getQuality() ||
    360             !isClose(inSampleRate, oldSampleRate, mFilterSampleRate, mSampleRate)) {
    361         mFilterSampleRate = inSampleRate;
    362         mFilterQuality = getQuality();
    363 
    364         double stopBandAtten;
    365         double tbwCheat = 1.; // how much we "cheat" into aliasing
    366         int halfLength;
    367         double fcr = 0.;
    368 
    369         // Begin Kaiser Filter computation
    370         //
    371         // The quantization floor for S16 is about 96db - 10*log_10(#length) + 3dB.
    372         // Keep the stop band attenuation no greater than 84-85dB for 32 length S16 filters
    373         //
    374         // For s32 we keep the stop band attenuation at the same as 16b resolution, about
    375         // 96-98dB
    376         //
    377 
    378         if (mPropertyEnableAtSampleRate >= 0 && mSampleRate >= mPropertyEnableAtSampleRate) {
    379             // An alternative method which allows allows a greater fcr
    380             // at the expense of potential aliasing.
    381             halfLength = mPropertyHalfFilterLength;
    382             stopBandAtten = mPropertyStopbandAttenuation;
    383             useS32 = true;
    384 
    385             // Use either the stopband location for design (tbwCheat)
    386             // or use the 3dB cutoff location for design (fcr).
    387             // This choice is exclusive and based on whether fcr > 0.
    388             if (mPropertyTransitionBandwidthCheat != 0) {
    389                 tbwCheat = mPropertyTransitionBandwidthCheat / 100.;
    390             } else {
    391                 fcr = mInSampleRate <= mSampleRate
    392                         ? 0.5 : 0.5 * mSampleRate / mInSampleRate;
    393                 fcr *= mPropertyCutoffPercent / 100.;
    394             }
    395         } else {
    396             // Voice quality devices have lower sampling rates
    397             // (and may be a consequence of downstream AMR-WB / G.722 codecs).
    398             // For these devices, we ensure a wider resampler passband
    399             // at the expense of aliasing noise (stopband attenuation
    400             // and stopband frequency).
    401             //
    402             constexpr uint32_t kVoiceDeviceSampleRate = 16000;
    403 
    404             if (mFilterQuality == DYN_HIGH_QUALITY) {
    405                 // float or 32b coefficients
    406                 useS32 = true;
    407                 stopBandAtten = 98.;
    408                 if (inSampleRate >= mSampleRate * 4) {
    409                     halfLength = 48;
    410                 } else if (inSampleRate >= mSampleRate * 2) {
    411                     halfLength = 40;
    412                 } else {
    413                     halfLength = 32;
    414                 }
    415 
    416                 if (mSampleRate <= kVoiceDeviceSampleRate) {
    417                     if (inSampleRate >= mSampleRate * 2) {
    418                         halfLength += 16;
    419                     } else {
    420                         halfLength += 8;
    421                     }
    422                     stopBandAtten = 84.;
    423                     tbwCheat = 1.05;
    424                 }
    425             } else if (mFilterQuality == DYN_LOW_QUALITY) {
    426                 // float or 16b coefficients
    427                 useS32 = false;
    428                 stopBandAtten = 80.;
    429                 if (inSampleRate >= mSampleRate * 4) {
    430                     halfLength = 24;
    431                 } else if (inSampleRate >= mSampleRate * 2) {
    432                     halfLength = 16;
    433                 } else {
    434                     halfLength = 8;
    435                 }
    436                 if (mSampleRate <= kVoiceDeviceSampleRate) {
    437                     if (inSampleRate >= mSampleRate * 2) {
    438                         halfLength += 8;
    439                     }
    440                     tbwCheat = 1.05;
    441                 } else if (inSampleRate <= mSampleRate) {
    442                     tbwCheat = 1.05;
    443                 } else {
    444                     tbwCheat = 1.03;
    445                 }
    446             } else { // DYN_MED_QUALITY
    447                 // float or 16b coefficients
    448                 // note: > 64 length filters with 16b coefs can have quantization noise problems
    449                 useS32 = false;
    450                 stopBandAtten = 84.;
    451                 if (inSampleRate >= mSampleRate * 4) {
    452                     halfLength = 32;
    453                 } else if (inSampleRate >= mSampleRate * 2) {
    454                     halfLength = 24;
    455                 } else {
    456                     halfLength = 16;
    457                 }
    458 
    459                 if (mSampleRate <= kVoiceDeviceSampleRate) {
    460                     if (inSampleRate >= mSampleRate * 2) {
    461                         halfLength += 16;
    462                     } else {
    463                         halfLength += 8;
    464                     }
    465                     tbwCheat = 1.05;
    466                 } else if (inSampleRate <= mSampleRate) {
    467                     tbwCheat = 1.03;
    468                 } else {
    469                     tbwCheat = 1.01;
    470                 }
    471             }
    472         }
    473 
    474         if (fcr > 0.) {
    475             ALOGV("%s: mFilterQuality:%d inSampleRate:%d mSampleRate:%d halfLength:%d "
    476                     "stopBandAtten:%lf fcr:%lf",
    477                     __func__, mFilterQuality, inSampleRate, mSampleRate, halfLength,
    478                     stopBandAtten, fcr);
    479         } else {
    480             ALOGV("%s: mFilterQuality:%d inSampleRate:%d mSampleRate:%d halfLength:%d "
    481                     "stopBandAtten:%lf tbwCheat:%lf",
    482                     __func__, mFilterQuality, inSampleRate, mSampleRate, halfLength,
    483                     stopBandAtten, tbwCheat);
    484         }
    485 
    486 
    487         // determine the number of polyphases in the filterbank.
    488         // for 16b, it is desirable to have 2^(16/2) = 256 phases.
    489         // https://ccrma.stanford.edu/~jos/resample/Relation_Interpolation_Error_Quantization.html
    490         //
    491         // We are a bit more lax on this.
    492 
    493         int phases = mSampleRate / gcd(mSampleRate, inSampleRate);
    494 
    495         // TODO: Once dynamic sample rate change is an option, the code below
    496         // should be modified to execute only when dynamic sample rate change is enabled.
    497         //
    498         // as above, #phases less than 63 is too few phases for accurate linear interpolation.
    499         // we increase the phases to compensate, but more phases means more memory per
    500         // filter and more time to compute the filter.
    501         //
    502         // if we know that the filter will be used for dynamic sample rate changes,
    503         // that would allow us skip this part for fixed sample rate resamplers.
    504         //
    505         while (phases<63) {
    506             phases *= 2; // this code only needed to support dynamic rate changes
    507         }
    508 
    509         if (phases>=256) {  // too many phases, always interpolate
    510             phases = 127;
    511         }
    512 
    513         // create the filter
    514         mConstants.set(phases, halfLength, inSampleRate, mSampleRate);
    515         if (fcr > 0.) {
    516             createKaiserFir(mConstants, stopBandAtten, fcr);
    517         } else {
    518             createKaiserFir(mConstants, stopBandAtten,
    519                     inSampleRate, mSampleRate, tbwCheat);
    520         }
    521     } // End Kaiser filter
    522 
    523     // update phase and state based on the new filter.
    524     const Constants& c(mConstants);
    525     mInBuffer.resize(mChannelCount, c.mHalfNumCoefs);
    526     const uint32_t phaseWrapLimit = c.mL << c.mShift;
    527     // try to preserve as much of the phase fraction as possible for on-the-fly changes
    528     mPhaseFraction = static_cast<unsigned long long>(mPhaseFraction)
    529             * phaseWrapLimit / oldPhaseWrapLimit;
    530     mPhaseFraction %= phaseWrapLimit; // should not do anything, but just in case.
    531     mPhaseIncrement = static_cast<uint32_t>(static_cast<uint64_t>(phaseWrapLimit)
    532             * inSampleRate / mSampleRate);
    533 
    534     // determine which resampler to use
    535     // check if locked phase (works only if mPhaseIncrement has no "fractional phase bits")
    536     int locked = (mPhaseIncrement << (sizeof(mPhaseIncrement)*8 - c.mShift)) == 0;
    537     if (locked) {
    538         mPhaseFraction = mPhaseFraction >> c.mShift << c.mShift; // remove fractional phase
    539     }
    540 
    541     // stride is the minimum number of filter coefficients processed per loop iteration.
    542     // We currently only allow a stride of 16 to match with SIMD processing.
    543     // This means that the filter length must be a multiple of 16,
    544     // or half the filter length (mHalfNumCoefs) must be a multiple of 8.
    545     //
    546     // Note: A stride of 2 is achieved with non-SIMD processing.
    547     int stride = ((c.mHalfNumCoefs & 7) == 0) ? 16 : 2;
    548     LOG_ALWAYS_FATAL_IF(stride < 16, "Resampler stride must be 16 or more");
    549     LOG_ALWAYS_FATAL_IF(mChannelCount < 1 || mChannelCount > 8,
    550             "Resampler channels(%d) must be between 1 to 8", mChannelCount);
    551     // stride 16 (falls back to stride 2 for machines that do not support NEON)
    552     if (locked) {
    553         switch (mChannelCount) {
    554         case 1:
    555             mResampleFunc = &AudioResamplerDyn<TC, TI, TO>::resample<1, true, 16>;
    556             break;
    557         case 2:
    558             mResampleFunc = &AudioResamplerDyn<TC, TI, TO>::resample<2, true, 16>;
    559             break;
    560         case 3:
    561             mResampleFunc = &AudioResamplerDyn<TC, TI, TO>::resample<3, true, 16>;
    562             break;
    563         case 4:
    564             mResampleFunc = &AudioResamplerDyn<TC, TI, TO>::resample<4, true, 16>;
    565             break;
    566         case 5:
    567             mResampleFunc = &AudioResamplerDyn<TC, TI, TO>::resample<5, true, 16>;
    568             break;
    569         case 6:
    570             mResampleFunc = &AudioResamplerDyn<TC, TI, TO>::resample<6, true, 16>;
    571             break;
    572         case 7:
    573             mResampleFunc = &AudioResamplerDyn<TC, TI, TO>::resample<7, true, 16>;
    574             break;
    575         case 8:
    576             mResampleFunc = &AudioResamplerDyn<TC, TI, TO>::resample<8, true, 16>;
    577             break;
    578         }
    579     } else {
    580         switch (mChannelCount) {
    581         case 1:
    582             mResampleFunc = &AudioResamplerDyn<TC, TI, TO>::resample<1, false, 16>;
    583             break;
    584         case 2:
    585             mResampleFunc = &AudioResamplerDyn<TC, TI, TO>::resample<2, false, 16>;
    586             break;
    587         case 3:
    588             mResampleFunc = &AudioResamplerDyn<TC, TI, TO>::resample<3, false, 16>;
    589             break;
    590         case 4:
    591             mResampleFunc = &AudioResamplerDyn<TC, TI, TO>::resample<4, false, 16>;
    592             break;
    593         case 5:
    594             mResampleFunc = &AudioResamplerDyn<TC, TI, TO>::resample<5, false, 16>;
    595             break;
    596         case 6:
    597             mResampleFunc = &AudioResamplerDyn<TC, TI, TO>::resample<6, false, 16>;
    598             break;
    599         case 7:
    600             mResampleFunc = &AudioResamplerDyn<TC, TI, TO>::resample<7, false, 16>;
    601             break;
    602         case 8:
    603             mResampleFunc = &AudioResamplerDyn<TC, TI, TO>::resample<8, false, 16>;
    604             break;
    605         }
    606     }
    607 #ifdef DEBUG_RESAMPLER
    608     printf("channels:%d  %s  stride:%d  %s  coef:%d  shift:%d\n",
    609             mChannelCount, locked ? "locked" : "interpolated",
    610             stride, useS32 ? "S32" : "S16", 2*c.mHalfNumCoefs, c.mShift);
    611 #endif
    612 }
    613 
    614 template<typename TC, typename TI, typename TO>
    615 size_t AudioResamplerDyn<TC, TI, TO>::resample(int32_t* out, size_t outFrameCount,
    616             AudioBufferProvider* provider)
    617 {
    618     return (this->*mResampleFunc)(reinterpret_cast<TO*>(out), outFrameCount, provider);
    619 }
    620 
    621 template<typename TC, typename TI, typename TO>
    622 template<int CHANNELS, bool LOCKED, int STRIDE>
    623 size_t AudioResamplerDyn<TC, TI, TO>::resample(TO* out, size_t outFrameCount,
    624         AudioBufferProvider* provider)
    625 {
    626     // TODO Mono -> Mono is not supported. OUTPUT_CHANNELS reflects minimum of stereo out.
    627     const int OUTPUT_CHANNELS = (CHANNELS < 2) ? 2 : CHANNELS;
    628     const Constants& c(mConstants);
    629     const TC* const coefs = mConstants.mFirCoefs;
    630     TI* impulse = mInBuffer.getImpulse();
    631     size_t inputIndex = 0;
    632     uint32_t phaseFraction = mPhaseFraction;
    633     const uint32_t phaseIncrement = mPhaseIncrement;
    634     size_t outputIndex = 0;
    635     size_t outputSampleCount = outFrameCount * OUTPUT_CHANNELS;
    636     const uint32_t phaseWrapLimit = c.mL << c.mShift;
    637     size_t inFrameCount = (phaseIncrement * (uint64_t)outFrameCount + phaseFraction)
    638             / phaseWrapLimit;
    639     // sanity check that inFrameCount is in signed 32 bit integer range.
    640     ALOG_ASSERT(0 <= inFrameCount && inFrameCount < (1U << 31));
    641 
    642     //ALOGV("inFrameCount:%d  outFrameCount:%d"
    643     //        "  phaseIncrement:%u  phaseFraction:%u  phaseWrapLimit:%u",
    644     //        inFrameCount, outFrameCount, phaseIncrement, phaseFraction, phaseWrapLimit);
    645 
    646     // NOTE: be very careful when modifying the code here. register
    647     // pressure is very high and a small change might cause the compiler
    648     // to generate far less efficient code.
    649     // Always sanity check the result with objdump or test-resample.
    650 
    651     // the following logic is a bit convoluted to keep the main processing loop
    652     // as tight as possible with register allocation.
    653     while (outputIndex < outputSampleCount) {
    654         //ALOGV("LOOP: inFrameCount:%d  outputIndex:%d  outFrameCount:%d"
    655         //        "  phaseFraction:%u  phaseWrapLimit:%u",
    656         //        inFrameCount, outputIndex, outFrameCount, phaseFraction, phaseWrapLimit);
    657 
    658         // check inputIndex overflow
    659         ALOG_ASSERT(inputIndex <= mBuffer.frameCount, "inputIndex%zu > frameCount%zu",
    660                 inputIndex, mBuffer.frameCount);
    661         // Buffer is empty, fetch a new one if necessary (inFrameCount > 0).
    662         // We may not fetch a new buffer if the existing data is sufficient.
    663         while (mBuffer.frameCount == 0 && inFrameCount > 0) {
    664             mBuffer.frameCount = inFrameCount;
    665             provider->getNextBuffer(&mBuffer);
    666             if (mBuffer.raw == NULL) {
    667                 // We are either at the end of playback or in an underrun situation.
    668                 // Reset buffer to prevent pop noise at the next buffer.
    669                 mInBuffer.reset();
    670                 goto resample_exit;
    671             }
    672             inFrameCount -= mBuffer.frameCount;
    673             if (phaseFraction >= phaseWrapLimit) { // read in data
    674                 mInBuffer.template readAdvance<CHANNELS>(
    675                         impulse, c.mHalfNumCoefs,
    676                         reinterpret_cast<TI*>(mBuffer.raw), inputIndex);
    677                 inputIndex++;
    678                 phaseFraction -= phaseWrapLimit;
    679                 while (phaseFraction >= phaseWrapLimit) {
    680                     if (inputIndex >= mBuffer.frameCount) {
    681                         inputIndex = 0;
    682                         provider->releaseBuffer(&mBuffer);
    683                         break;
    684                     }
    685                     mInBuffer.template readAdvance<CHANNELS>(
    686                             impulse, c.mHalfNumCoefs,
    687                             reinterpret_cast<TI*>(mBuffer.raw), inputIndex);
    688                     inputIndex++;
    689                     phaseFraction -= phaseWrapLimit;
    690                 }
    691             }
    692         }
    693         const TI* const in = reinterpret_cast<const TI*>(mBuffer.raw);
    694         const size_t frameCount = mBuffer.frameCount;
    695         const int coefShift = c.mShift;
    696         const int halfNumCoefs = c.mHalfNumCoefs;
    697         const TO* const volumeSimd = mVolumeSimd;
    698 
    699         // main processing loop
    700         while (CC_LIKELY(outputIndex < outputSampleCount)) {
    701             // caution: fir() is inlined and may be large.
    702             // output will be loaded with the appropriate values
    703             //
    704             // from the input samples in impulse[-halfNumCoefs+1]... impulse[halfNumCoefs]
    705             // from the polyphase filter of (phaseFraction / phaseWrapLimit) in coefs.
    706             //
    707             //ALOGV("LOOP2: inFrameCount:%d  outputIndex:%d  outFrameCount:%d"
    708             //        "  phaseFraction:%u  phaseWrapLimit:%u",
    709             //        inFrameCount, outputIndex, outFrameCount, phaseFraction, phaseWrapLimit);
    710             ALOG_ASSERT(phaseFraction < phaseWrapLimit);
    711             fir<CHANNELS, LOCKED, STRIDE>(
    712                     &out[outputIndex],
    713                     phaseFraction, phaseWrapLimit,
    714                     coefShift, halfNumCoefs, coefs,
    715                     impulse, volumeSimd);
    716 
    717             outputIndex += OUTPUT_CHANNELS;
    718 
    719             phaseFraction += phaseIncrement;
    720             while (phaseFraction >= phaseWrapLimit) {
    721                 if (inputIndex >= frameCount) {
    722                     goto done;  // need a new buffer
    723                 }
    724                 mInBuffer.template readAdvance<CHANNELS>(impulse, halfNumCoefs, in, inputIndex);
    725                 inputIndex++;
    726                 phaseFraction -= phaseWrapLimit;
    727             }
    728         }
    729 done:
    730         // We arrive here when we're finished or when the input buffer runs out.
    731         // Regardless we need to release the input buffer if we've acquired it.
    732         if (inputIndex > 0) {  // we've acquired a buffer (alternatively could check frameCount)
    733             ALOG_ASSERT(inputIndex == frameCount, "inputIndex(%zu) != frameCount(%zu)",
    734                     inputIndex, frameCount);  // must have been fully read.
    735             inputIndex = 0;
    736             provider->releaseBuffer(&mBuffer);
    737             ALOG_ASSERT(mBuffer.frameCount == 0);
    738         }
    739     }
    740 
    741 resample_exit:
    742     // inputIndex must be zero in all three cases:
    743     // (1) the buffer never was been acquired; (2) the buffer was
    744     // released at "done:"; or (3) getNextBuffer() failed.
    745     ALOG_ASSERT(inputIndex == 0, "Releasing: inputindex:%zu frameCount:%zu  phaseFraction:%u",
    746             inputIndex, mBuffer.frameCount, phaseFraction);
    747     ALOG_ASSERT(mBuffer.frameCount == 0); // there must be no frames in the buffer
    748     mInBuffer.setImpulse(impulse);
    749     mPhaseFraction = phaseFraction;
    750     return outputIndex / OUTPUT_CHANNELS;
    751 }
    752 
    753 /* instantiate templates used by AudioResampler::create */
    754 template class AudioResamplerDyn<float, float, float>;
    755 template class AudioResamplerDyn<int16_t, int16_t, int32_t>;
    756 template class AudioResamplerDyn<int32_t, int16_t, int32_t>;
    757 
    758 // ----------------------------------------------------------------------------
    759 } // namespace android
    760