Home | History | Annotate | Download | only in audioflinger
      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_GEN_H
     18 #define ANDROID_AUDIO_RESAMPLER_FIR_GEN_H
     19 
     20 namespace android {
     21 
     22 /*
     23  * generates a sine wave at equal steps.
     24  *
     25  * As most of our functions use sine or cosine at equal steps,
     26  * it is very efficient to compute them that way (single multiply and subtract),
     27  * rather than invoking the math library sin() or cos() each time.
     28  *
     29  * SineGen uses Goertzel's Algorithm (as a generator not a filter)
     30  * to calculate sine(wstart + n * wstep) or cosine(wstart + n * wstep)
     31  * by stepping through 0, 1, ... n.
     32  *
     33  * e^i(wstart+wstep) = 2cos(wstep) * e^i(wstart) - e^i(wstart-wstep)
     34  *
     35  * or looking at just the imaginary sine term, as the cosine follows identically:
     36  *
     37  * sin(wstart+wstep) = 2cos(wstep) * sin(wstart) - sin(wstart-wstep)
     38  *
     39  * Goertzel's algorithm is more efficient than the angle addition formula,
     40  * e^i(wstart+wstep) = e^i(wstart) * e^i(wstep), which takes up to
     41  * 4 multiplies and 2 adds (or 3* and 3+) and requires both sine and
     42  * cosine generation due to the complex * complex multiply (full rotation).
     43  *
     44  * See: http://en.wikipedia.org/wiki/Goertzel_algorithm
     45  *
     46  */
     47 
     48 class SineGen {
     49 public:
     50     SineGen(double wstart, double wstep, bool cosine = false) {
     51         if (cosine) {
     52             mCurrent = cos(wstart);
     53             mPrevious = cos(wstart - wstep);
     54         } else {
     55             mCurrent = sin(wstart);
     56             mPrevious = sin(wstart - wstep);
     57         }
     58         mTwoCos = 2.*cos(wstep);
     59     }
     60     SineGen(double expNow, double expPrev, double twoCosStep) {
     61         mCurrent = expNow;
     62         mPrevious = expPrev;
     63         mTwoCos = twoCosStep;
     64     }
     65     inline double value() const {
     66         return mCurrent;
     67     }
     68     inline void advance() {
     69         double tmp = mCurrent;
     70         mCurrent = mCurrent*mTwoCos - mPrevious;
     71         mPrevious = tmp;
     72     }
     73     inline double valueAdvance() {
     74         double tmp = mCurrent;
     75         mCurrent = mCurrent*mTwoCos - mPrevious;
     76         mPrevious = tmp;
     77         return tmp;
     78     }
     79 
     80 private:
     81     double mCurrent; // current value of sine/cosine
     82     double mPrevious; // previous value of sine/cosine
     83     double mTwoCos; // stepping factor
     84 };
     85 
     86 /*
     87  * generates a series of sine generators, phase offset by fixed steps.
     88  *
     89  * This is used to generate polyphase sine generators, one per polyphase
     90  * in the filter code below.
     91  *
     92  * The SineGen returned by value() starts at innerStart = outerStart + n*outerStep;
     93  * increments by innerStep.
     94  *
     95  */
     96 
     97 class SineGenGen {
     98 public:
     99     SineGenGen(double outerStart, double outerStep, double innerStep, bool cosine = false)
    100             : mSineInnerCur(outerStart, outerStep, cosine),
    101               mSineInnerPrev(outerStart-innerStep, outerStep, cosine)
    102     {
    103         mTwoCos = 2.*cos(innerStep);
    104     }
    105     inline SineGen value() {
    106         return SineGen(mSineInnerCur.value(), mSineInnerPrev.value(), mTwoCos);
    107     }
    108     inline void advance() {
    109         mSineInnerCur.advance();
    110         mSineInnerPrev.advance();
    111     }
    112     inline SineGen valueAdvance() {
    113         return SineGen(mSineInnerCur.valueAdvance(), mSineInnerPrev.valueAdvance(), mTwoCos);
    114     }
    115 
    116 private:
    117     SineGen mSineInnerCur; // generate the inner sine values (stepped by outerStep).
    118     SineGen mSineInnerPrev; // generate the inner sine previous values
    119                             // (behind by innerStep, stepped by outerStep).
    120     double mTwoCos; // the inner stepping factor for the returned SineGen.
    121 };
    122 
    123 static inline double sqr(double x) {
    124     return x * x;
    125 }
    126 
    127 /*
    128  * rounds a double to the nearest integer for FIR coefficients.
    129  *
    130  * One variant uses noise shaping, which must keep error history
    131  * to work (the err parameter, initialized to 0).
    132  * The other variant is a non-noise shaped version for
    133  * S32 coefficients (noise shaping doesn't gain much).
    134  *
    135  * Caution: No bounds saturation is applied, but isn't needed in this case.
    136  *
    137  * @param x is the value to round.
    138  *
    139  * @param maxval is the maximum integer scale factor expressed as an int64 (for headroom).
    140  * Typically this may be the maximum positive integer+1 (using the fact that double precision
    141  * FIR coefficients generated here are never that close to 1.0 to pose an overflow condition).
    142  *
    143  * @param err is the previous error (actual - rounded) for the previous rounding op.
    144  * For 16b coefficients this can improve stopband dB performance by up to 2dB.
    145  *
    146  * Many variants exist for the noise shaping: http://en.wikipedia.org/wiki/Noise_shaping
    147  *
    148  */
    149 
    150 static inline int64_t toint(double x, int64_t maxval, double& err) {
    151     double val = x * maxval;
    152     double ival = floor(val + 0.5 + err*0.2);
    153     err = val - ival;
    154     return static_cast<int64_t>(ival);
    155 }
    156 
    157 static inline int64_t toint(double x, int64_t maxval) {
    158     return static_cast<int64_t>(floor(x * maxval + 0.5));
    159 }
    160 
    161 /*
    162  * Modified Bessel function of the first kind
    163  * http://en.wikipedia.org/wiki/Bessel_function
    164  *
    165  * The formulas are taken from Abramowitz and Stegun,
    166  * _Handbook of Mathematical Functions_ (links below):
    167  *
    168  * http://people.math.sfu.ca/~cbm/aands/page_375.htm
    169  * http://people.math.sfu.ca/~cbm/aands/page_378.htm
    170  *
    171  * http://dlmf.nist.gov/10.25
    172  * http://dlmf.nist.gov/10.40
    173  *
    174  * Note we assume x is nonnegative (the function is symmetric,
    175  * pass in the absolute value as needed).
    176  *
    177  * Constants are compile time derived with templates I0Term<> and
    178  * I0ATerm<> to the precision of the compiler.  The series can be expanded
    179  * to any precision needed, but currently set around 24b precision.
    180  *
    181  * We use a bit of template math here, constexpr would probably be
    182  * more appropriate for a C++11 compiler.
    183  *
    184  * For the intermediate range 3.75 < x < 15, we use minimax polynomial fit.
    185  *
    186  */
    187 
    188 template <int N>
    189 struct I0Term {
    190     static const double value = I0Term<N-1>::value / (4. * N * N);
    191 };
    192 
    193 template <>
    194 struct I0Term<0> {
    195     static const double value = 1.;
    196 };
    197 
    198 template <int N>
    199 struct I0ATerm {
    200     static const double value = I0ATerm<N-1>::value * (2.*N-1.) * (2.*N-1.) / (8. * N);
    201 };
    202 
    203 template <>
    204 struct I0ATerm<0> { // 1/sqrt(2*PI);
    205     static const double value = 0.398942280401432677939946059934381868475858631164934657665925;
    206 };
    207 
    208 #if USE_HORNERS_METHOD
    209 /* Polynomial evaluation of A + Bx + Cx^2 + Dx^3 + ...
    210  * using Horner's Method: http://en.wikipedia.org/wiki/Horner's_method
    211  *
    212  * This has fewer multiplications than Estrin's method below, but has back to back
    213  * floating point dependencies.
    214  *
    215  * On ARM this appears to work slower, so USE_HORNERS_METHOD is not default enabled.
    216  */
    217 
    218 inline double Poly2(double A, double B, double x) {
    219     return A + x * B;
    220 }
    221 
    222 inline double Poly4(double A, double B, double C, double D, double x) {
    223     return A + x * (B + x * (C + x * (D)));
    224 }
    225 
    226 inline double Poly7(double A, double B, double C, double D, double E, double F, double G,
    227         double x) {
    228     return A + x * (B + x * (C + x * (D + x * (E + x * (F + x * (G))))));
    229 }
    230 
    231 inline double Poly9(double A, double B, double C, double D, double E, double F, double G,
    232         double H, double I, double x) {
    233     return A + x * (B + x * (C + x * (D + x * (E + x * (F + x * (G + x * (H + x * (I))))))));
    234 }
    235 
    236 #else
    237 /* Polynomial evaluation of A + Bx + Cx^2 + Dx^3 + ...
    238  * using Estrin's Method: http://en.wikipedia.org/wiki/Estrin's_scheme
    239  *
    240  * This is typically faster, perhaps gains about 5-10% overall on ARM processors
    241  * over Horner's method above.
    242  */
    243 
    244 inline double Poly2(double A, double B, double x) {
    245     return A + B * x;
    246 }
    247 
    248 inline double Poly3(double A, double B, double C, double x, double x2) {
    249     return Poly2(A, B, x) + C * x2;
    250 }
    251 
    252 inline double Poly3(double A, double B, double C, double x) {
    253     return Poly2(A, B, x) + C * x * x;
    254 }
    255 
    256 inline double Poly4(double A, double B, double C, double D, double x, double x2) {
    257     return Poly2(A, B, x) + Poly2(C, D, x) * x2; // same as poly2(poly2, poly2, x2);
    258 }
    259 
    260 inline double Poly4(double A, double B, double C, double D, double x) {
    261     return Poly4(A, B, C, D, x, x * x);
    262 }
    263 
    264 inline double Poly7(double A, double B, double C, double D, double E, double F, double G,
    265         double x) {
    266     double x2 = x * x;
    267     return Poly4(A, B, C, D, x, x2) + Poly3(E, F, G, x, x2) * (x2 * x2);
    268 }
    269 
    270 inline double Poly8(double A, double B, double C, double D, double E, double F, double G,
    271         double H, double x, double x2, double x4) {
    272     return Poly4(A, B, C, D, x, x2) + Poly4(E, F, G, H, x, x2) * x4;
    273 }
    274 
    275 inline double Poly9(double A, double B, double C, double D, double E, double F, double G,
    276         double H, double I, double x) {
    277     double x2 = x * x;
    278 #if 1
    279     // It does not seem faster to explicitly decompose Poly8 into Poly4, but
    280     // could depend on compiler floating point scheduling.
    281     double x4 = x2 * x2;
    282     return Poly8(A, B, C, D, E, F, G, H, x, x2, x4) + I * (x4 * x4);
    283 #else
    284     double val = Poly4(A, B, C, D, x, x2);
    285     double x4 = x2 * x2;
    286     return val + Poly4(E, F, G, H, x, x2) * x4 + I * (x4 * x4);
    287 #endif
    288 }
    289 #endif
    290 
    291 static inline double I0(double x) {
    292     if (x < 3.75) {
    293         x *= x;
    294         return Poly7(I0Term<0>::value, I0Term<1>::value,
    295                 I0Term<2>::value, I0Term<3>::value,
    296                 I0Term<4>::value, I0Term<5>::value,
    297                 I0Term<6>::value, x); // e < 1.6e-7
    298     }
    299     if (1) {
    300         /*
    301          * Series expansion coefs are easy to calculate, but are expanded around 0,
    302          * so error is unequal over the interval 0 < x < 3.75, the error being
    303          * significantly better near 0.
    304          *
    305          * A better solution is to use precise minimax polynomial fits.
    306          *
    307          * We use a slightly more complicated solution for 3.75 < x < 15, based on
    308          * the tables in Blair and Edwards, "Stable Rational Minimax Approximations
    309          * to the Modified Bessel Functions I0(x) and I1(x)", Chalk Hill Nuclear Laboratory,
    310          * AECL-4928.
    311          *
    312          * http://www.iaea.org/inis/collection/NCLCollectionStore/_Public/06/178/6178667.pdf
    313          *
    314          * See Table 11 for 0 < x < 15; e < 10^(-7.13).
    315          *
    316          * Note: Beta cannot exceed 15 (hence Stopband cannot exceed 144dB = 24b).
    317          *
    318          * This speeds up overall computation by about 40% over using the else clause below,
    319          * which requires sqrt and exp.
    320          *
    321          */
    322 
    323         x *= x;
    324         double num = Poly9(-0.13544938430e9, -0.33153754512e8,
    325                 -0.19406631946e7, -0.48058318783e5,
    326                 -0.63269783360e3, -0.49520779070e1,
    327                 -0.24970910370e-1, -0.74741159550e-4,
    328                 -0.18257612460e-6, x);
    329         double y = x - 225.; // reflection around 15 (squared)
    330         double den = Poly4(-0.34598737196e8, 0.23852643181e6,
    331                 -0.70699387620e3, 0.10000000000e1, y);
    332         return num / den;
    333 
    334 #if IO_EXTENDED_BETA
    335         /* Table 42 for x > 15; e < 10^(-8.11).
    336          * This is used for Beta>15, but is disabled here as
    337          * we never use Beta that high.
    338          *
    339          * NOTE: This should be enabled only for x > 15.
    340          */
    341 
    342         double y = 1./x;
    343         double z = y - (1./15);
    344         double num = Poly2(0.415079861746e1, -0.5149092496e1, z);
    345         double den = Poly3(0.103150763823e2, -0.14181687413e2,
    346                 0.1000000000e1, z);
    347         return exp(x) * sqrt(y) * num / den;
    348 #endif
    349     } else {
    350         /*
    351          * NOT USED, but reference for large Beta.
    352          *
    353          * Abramowitz and Stegun asymptotic formula.
    354          * works for x > 3.75.
    355          */
    356         double y = 1./x;
    357         return exp(x) * sqrt(y) *
    358                 // note: reciprocal squareroot may be easier!
    359                 // http://en.wikipedia.org/wiki/Fast_inverse_square_root
    360                 Poly9(I0ATerm<0>::value, I0ATerm<1>::value,
    361                         I0ATerm<2>::value, I0ATerm<3>::value,
    362                         I0ATerm<4>::value, I0ATerm<5>::value,
    363                         I0ATerm<6>::value, I0ATerm<7>::value,
    364                         I0ATerm<8>::value, y); // (... e) < 1.9e-7
    365     }
    366 }
    367 
    368 /* A speed optimized version of the Modified Bessel I0() which incorporates
    369  * the sqrt and numerator multiply and denominator divide into the computation.
    370  * This speeds up filter computation by about 10-15%.
    371  */
    372 static inline double I0SqrRat(double x2, double num, double den) {
    373     if (x2 < (3.75 * 3.75)) {
    374         return Poly7(I0Term<0>::value, I0Term<1>::value,
    375                 I0Term<2>::value, I0Term<3>::value,
    376                 I0Term<4>::value, I0Term<5>::value,
    377                 I0Term<6>::value, x2) * num / den; // e < 1.6e-7
    378     }
    379     num *= Poly9(-0.13544938430e9, -0.33153754512e8,
    380             -0.19406631946e7, -0.48058318783e5,
    381             -0.63269783360e3, -0.49520779070e1,
    382             -0.24970910370e-1, -0.74741159550e-4,
    383             -0.18257612460e-6, x2); // e < 10^(-7.13).
    384     double y = x2 - 225.; // reflection around 15 (squared)
    385     den *= Poly4(-0.34598737196e8, 0.23852643181e6,
    386             -0.70699387620e3, 0.10000000000e1, y);
    387     return num / den;
    388 }
    389 
    390 /*
    391  * calculates the transition bandwidth for a Kaiser filter
    392  *
    393  * Formula 3.2.8, Vaidyanathan, _Multirate Systems and Filter Banks_, p. 48
    394  * Formula 7.76, Oppenheim and Schafer, _Discrete-time Signal Processing, 3e_, p. 542
    395  *
    396  * @param halfNumCoef is half the number of coefficients per filter phase.
    397  *
    398  * @param stopBandAtten is the stop band attenuation desired.
    399  *
    400  * @return the transition bandwidth in normalized frequency (0 <= f <= 0.5)
    401  */
    402 static inline double firKaiserTbw(int halfNumCoef, double stopBandAtten) {
    403     return (stopBandAtten - 7.95)/((2.*14.36)*halfNumCoef);
    404 }
    405 
    406 /*
    407  * calculates the fir transfer response of the overall polyphase filter at w.
    408  *
    409  * Calculates the DTFT transfer coefficient H(w) for 0 <= w <= PI, utilizing the
    410  * fact that h[n] is symmetric (cosines only, no complex arithmetic).
    411  *
    412  * We use Goertzel's algorithm to accelerate the computation to essentially
    413  * a single multiply and 2 adds per filter coefficient h[].
    414  *
    415  * Be careful be careful to consider that h[n] is the overall polyphase filter,
    416  * with L phases, so rescaling H(w)/L is probably what you expect for "unity gain",
    417  * as you only use one of the polyphases at a time.
    418  */
    419 template <typename T>
    420 static inline double firTransfer(const T* coef, int L, int halfNumCoef, double w) {
    421     double accum = static_cast<double>(coef[0])*0.5;  // "center coefficient" from first bank
    422     coef += halfNumCoef;    // skip first filterbank (picked up by the last filterbank).
    423 #if SLOW_FIRTRANSFER
    424     /* Original code for reference.  This is equivalent to the code below, but slower. */
    425     for (int i=1 ; i<=L ; ++i) {
    426         for (int j=0, ix=i ; j<halfNumCoef ; ++j, ix+=L) {
    427             accum += cos(ix*w)*static_cast<double>(*coef++);
    428         }
    429     }
    430 #else
    431     /*
    432      * Our overall filter is stored striped by polyphases, not a contiguous h[n].
    433      * We could fetch coefficients in a non-contiguous fashion
    434      * but that will not scale to vector processing.
    435      *
    436      * We apply Goertzel's algorithm directly to each polyphase filter bank instead of
    437      * using cosine generation/multiplication, thereby saving one multiply per inner loop.
    438      *
    439      * See: http://en.wikipedia.org/wiki/Goertzel_algorithm
    440      * Also: Oppenheim and Schafer, _Discrete Time Signal Processing, 3e_, p. 720.
    441      *
    442      * We use the basic recursion to incorporate the cosine steps into real sequence x[n]:
    443      * s[n] = x[n] + (2cosw)*s[n-1] + s[n-2]
    444      *
    445      * y[n] = s[n] - e^(iw)s[n-1]
    446      *      = sum_{k=-\infty}^{n} x[k]e^(-iw(n-k))
    447      *      = e^(-iwn) sum_{k=0}^{n} x[k]e^(iwk)
    448      *
    449      * The summation contains the frequency steps we want multiplied by the source
    450      * (similar to a DTFT).
    451      *
    452      * Using symmetry, and just the real part (be careful, this must happen
    453      * after any internal complex multiplications), the polyphase filterbank
    454      * transfer function is:
    455      *
    456      * Hpp[n, w, w_0] = sum_{k=0}^{n} x[k] * cos(wk + w_0)
    457      *                = Re{ e^(iwn + iw_0) y[n]}
    458      *                = cos(wn+w_0) * s[n] - cos(w(n+1)+w_0) * s[n-1]
    459      *
    460      * using the fact that s[n] of real x[n] is real.
    461      *
    462      */
    463     double dcos = 2. * cos(L*w);
    464     int start = ((halfNumCoef)*L + 1);
    465     SineGen cc((start - L) * w, w, true); // cosine
    466     SineGen cp(start * w, w, true); // cosine
    467     for (int i=1 ; i<=L ; ++i) {
    468         double sc = 0;
    469         double sp = 0;
    470         for (int j=0 ; j<halfNumCoef ; ++j) {
    471             double tmp = sc;
    472             sc  = static_cast<double>(*coef++) + dcos*sc - sp;
    473             sp = tmp;
    474         }
    475         // If we are awfully clever, we can apply Goertzel's algorithm
    476         // again on the sc and sp sequences returned here.
    477         accum += cc.valueAdvance() * sc - cp.valueAdvance() * sp;
    478     }
    479 #endif
    480     return accum*2.;
    481 }
    482 
    483 /*
    484  * evaluates the minimum and maximum |H(f)| bound in a band region.
    485  *
    486  * This is usually done with equally spaced increments in the target band in question.
    487  * The passband is often very small, and sampled that way. The stopband is often much
    488  * larger.
    489  *
    490  * We use the fact that the overall polyphase filter has an additional bank at the end
    491  * for interpolation; hence it is overspecified for the H(f) computation.  Thus the
    492  * first polyphase is never actually checked, excepting its first term.
    493  *
    494  * In this code we use the firTransfer() evaluator above, which uses Goertzel's
    495  * algorithm to calculate the transfer function at each point.
    496  *
    497  * TODO: An alternative with equal spacing is the FFT/DFT.  An alternative with unequal
    498  * spacing is a chirp transform.
    499  *
    500  * @param coef is the designed polyphase filter banks
    501  *
    502  * @param L is the number of phases (for interpolation)
    503  *
    504  * @param halfNumCoef should be half the number of coefficients for a single
    505  * polyphase.
    506  *
    507  * @param fstart is the normalized frequency start.
    508  *
    509  * @param fend is the normalized frequency end.
    510  *
    511  * @param steps is the number of steps to take (sampling) between frequency start and end
    512  *
    513  * @param firMin returns the minimum transfer |H(f)| found
    514  *
    515  * @param firMax returns the maximum transfer |H(f)| found
    516  *
    517  * 0 <= f <= 0.5.
    518  * This is used to test passband and stopband performance.
    519  */
    520 template <typename T>
    521 static void testFir(const T* coef, int L, int halfNumCoef,
    522         double fstart, double fend, int steps, double &firMin, double &firMax) {
    523     double wstart = fstart*(2.*M_PI);
    524     double wend = fend*(2.*M_PI);
    525     double wstep = (wend - wstart)/steps;
    526     double fmax, fmin;
    527     double trf = firTransfer(coef, L, halfNumCoef, wstart);
    528     if (trf<0) {
    529         trf = -trf;
    530     }
    531     fmin = fmax = trf;
    532     wstart += wstep;
    533     for (int i=1; i<steps; ++i) {
    534         trf = firTransfer(coef, L, halfNumCoef, wstart);
    535         if (trf<0) {
    536             trf = -trf;
    537         }
    538         if (trf>fmax) {
    539             fmax = trf;
    540         }
    541         else if (trf<fmin) {
    542             fmin = trf;
    543         }
    544         wstart += wstep;
    545     }
    546     // renormalize - this is only needed for integer filter types
    547     double norm = 1./((1ULL<<(sizeof(T)*8-1))*L);
    548 
    549     firMin = fmin * norm;
    550     firMax = fmax * norm;
    551 }
    552 
    553 /*
    554  * evaluates the |H(f)| lowpass band characteristics.
    555  *
    556  * This function tests the lowpass characteristics for the overall polyphase filter,
    557  * and is used to verify the design.  For this case, fp should be set to the
    558  * passband normalized frequency from 0 to 0.5 for the overall filter (thus it
    559  * is the designed polyphase bank value / L).  Likewise for fs.
    560  *
    561  * @param coef is the designed polyphase filter banks
    562  *
    563  * @param L is the number of phases (for interpolation)
    564  *
    565  * @param halfNumCoef should be half the number of coefficients for a single
    566  * polyphase.
    567  *
    568  * @param fp is the passband normalized frequency, 0 < fp < fs < 0.5.
    569  *
    570  * @param fs is the stopband normalized frequency, 0 < fp < fs < 0.5.
    571  *
    572  * @param passSteps is the number of passband sampling steps.
    573  *
    574  * @param stopSteps is the number of stopband sampling steps.
    575  *
    576  * @param passMin is the minimum value in the passband
    577  *
    578  * @param passMax is the maximum value in the passband (useful for scaling).  This should
    579  * be less than 1., to avoid sine wave test overflow.
    580  *
    581  * @param passRipple is the passband ripple.  Typically this should be less than 0.1 for
    582  * an audio filter.  Generally speaker/headphone device characteristics will dominate
    583  * the passband term.
    584  *
    585  * @param stopMax is the maximum value in the stopband.
    586  *
    587  * @param stopRipple is the stopband ripple, also known as stopband attenuation.
    588  * Typically this should be greater than ~80dB for low quality, and greater than
    589  * ~100dB for full 16b quality, otherwise aliasing may become noticeable.
    590  *
    591  */
    592 template <typename T>
    593 static void testFir(const T* coef, int L, int halfNumCoef,
    594         double fp, double fs, int passSteps, int stopSteps,
    595         double &passMin, double &passMax, double &passRipple,
    596         double &stopMax, double &stopRipple) {
    597     double fmin, fmax;
    598     testFir(coef, L, halfNumCoef, 0., fp, passSteps, fmin, fmax);
    599     double d1 = (fmax - fmin)/2.;
    600     passMin = fmin;
    601     passMax = fmax;
    602     passRipple = -20.*log10(1. - d1); // passband ripple
    603     testFir(coef, L, halfNumCoef, fs, 0.5, stopSteps, fmin, fmax);
    604     // fmin is really not important for the stopband.
    605     stopMax = fmax;
    606     stopRipple = -20.*log10(fmax); // stopband ripple/attenuation
    607 }
    608 
    609 /*
    610  * Calculates the overall polyphase filter based on a windowed sinc function.
    611  *
    612  * The windowed sinc is an odd length symmetric filter of exactly L*halfNumCoef*2+1
    613  * taps for the entire kernel.  This is then decomposed into L+1 polyphase filterbanks.
    614  * The last filterbank is used for interpolation purposes (and is mostly composed
    615  * of the first bank shifted by one sample), and is unnecessary if one does
    616  * not do interpolation.
    617  *
    618  * We use the last filterbank for some transfer function calculation purposes,
    619  * so it needs to be generated anyways.
    620  *
    621  * @param coef is the caller allocated space for coefficients.  This should be
    622  * exactly (L+1)*halfNumCoef in size.
    623  *
    624  * @param L is the number of phases (for interpolation)
    625  *
    626  * @param halfNumCoef should be half the number of coefficients for a single
    627  * polyphase.
    628  *
    629  * @param stopBandAtten is the stopband value, should be >50dB.
    630  *
    631  * @param fcr is cutoff frequency/sampling rate (<0.5).  At this point, the energy
    632  * should be 6dB less. (fcr is where the amplitude drops by half).  Use the
    633  * firKaiserTbw() to calculate the transition bandwidth.  fcr is the midpoint
    634  * between the stop band and the pass band (fstop+fpass)/2.
    635  *
    636  * @param atten is the attenuation (generally slightly less than 1).
    637  */
    638 
    639 template <typename T>
    640 static inline void firKaiserGen(T* coef, int L, int halfNumCoef,
    641         double stopBandAtten, double fcr, double atten) {
    642     //
    643     // Formula 3.2.5, 3.2.7, Vaidyanathan, _Multirate Systems and Filter Banks_, p. 48
    644     // Formula 7.75, Oppenheim and Schafer, _Discrete-time Signal Processing, 3e_, p. 542
    645     //
    646     // See also: http://melodi.ee.washington.edu/courses/ee518/notes/lec17.pdf
    647     //
    648     // Kaiser window and beta parameter
    649     //
    650     //         | 0.1102*(A - 8.7)                         A > 50
    651     //  beta = | 0.5842*(A - 21)^0.4 + 0.07886*(A - 21)   21 <= A <= 50
    652     //         | 0.                                       A < 21
    653     //
    654     // with A is the desired stop-band attenuation in dBFS
    655     //
    656     //    30 dB    2.210
    657     //    40 dB    3.384
    658     //    50 dB    4.538
    659     //    60 dB    5.658
    660     //    70 dB    6.764
    661     //    80 dB    7.865
    662     //    90 dB    8.960
    663     //   100 dB   10.056
    664 
    665     const int N = L * halfNumCoef; // non-negative half
    666     const double beta = 0.1102 * (stopBandAtten - 8.7); // >= 50dB always
    667     const double xstep = (2. * M_PI) * fcr / L;
    668     const double xfrac = 1. / N;
    669     const double yscale = atten * L / (I0(beta) * M_PI);
    670     const double sqrbeta = sqr(beta);
    671 
    672     // We use sine generators, which computes sines on regular step intervals.
    673     // This speeds up overall computation about 40% from computing the sine directly.
    674 
    675     SineGenGen sgg(0., xstep, L*xstep); // generates sine generators (one per polyphase)
    676 
    677     for (int i=0 ; i<=L ; ++i) { // generate an extra set of coefs for interpolation
    678 
    679         // computation for a single polyphase of the overall filter.
    680         SineGen sg = sgg.valueAdvance(); // current sine generator for "j" inner loop.
    681         double err = 0; // for noise shaping on int16_t coefficients (over each polyphase)
    682 
    683         for (int j=0, ix=i ; j<halfNumCoef ; ++j, ix+=L) {
    684             double y;
    685             if (CC_LIKELY(ix)) {
    686                 double x = static_cast<double>(ix);
    687 
    688                 // sine generator: sg.valueAdvance() returns sin(ix*xstep);
    689                 // y = I0(beta * sqrt(1.0 - sqr(x * xfrac))) * yscale * sg.valueAdvance() / x;
    690                 y = I0SqrRat(sqrbeta * (1.0 - sqr(x * xfrac)), yscale * sg.valueAdvance(), x);
    691             } else {
    692                 y = 2. * atten * fcr; // center of filter, sinc(0) = 1.
    693                 sg.advance();
    694             }
    695 
    696             if (is_same<T, int16_t>::value) { // int16_t needs noise shaping
    697                 *coef++ = static_cast<T>(toint(y, 1ULL<<(sizeof(T)*8-1), err));
    698             } else if (is_same<T, int32_t>::value) {
    699                 *coef++ = static_cast<T>(toint(y, 1ULL<<(sizeof(T)*8-1)));
    700             } else { // assumed float or double
    701                 *coef++ = static_cast<T>(y);
    702             }
    703         }
    704     }
    705 }
    706 
    707 }; // namespace android
    708 
    709 #endif /*ANDROID_AUDIO_RESAMPLER_FIR_GEN_H*/
    710