winamp/Src/external_dependencies/openmpt-trunk/include/r8brain/CDSPRealFFT.h
2024-09-24 14:54:57 +02:00

820 lines
18 KiB
C++
Vendored

//$ nobt
//$ nocpp
/**
* @file CDSPRealFFT.h
*
* @brief Real-valued FFT transform class.
*
* This file includes FFT object implementation. All created FFT objects are
* kept in a global list after use, for a future reusal. Such approach
* minimizes time necessary to initialize the FFT object of the required
* length.
*
* r8brain-free-src Copyright (c) 2013-2022 Aleksey Vaneev
* See the "LICENSE" file for license.
*/
#ifndef R8B_CDSPREALFFT_INCLUDED
#define R8B_CDSPREALFFT_INCLUDED
#include "r8bbase.h"
#if !R8B_IPP && !R8B_PFFFT && !R8B_PFFFT_DOUBLE
#include "fft4g.h"
#endif // !R8B_IPP && !R8B_PFFFT && !R8B_PFFFT_DOUBLE
#if R8B_PFFFT
#include "pffft.h"
#endif // R8B_PFFFT
#if R8B_PFFFT_DOUBLE
#include "pffft_double/pffft_double.h"
#endif // R8B_PFFFT_DOUBLE
namespace r8b {
/**
* @brief Real-valued FFT transform class.
*
* Class implements a wrapper for real-valued discrete fast Fourier transform
* functions. The object of this class can only be obtained via the
* CDSPRealFFTKeeper class.
*
* Uses functions from the FFT package by: Copyright(C) 1996-2001 Takuya OOURA
* http://www.kurims.kyoto-u.ac.jp/~ooura/fft.html
*
* Also uses Intel IPP library functions if available (if the R8B_IPP=1 macro
* was defined). Note that IPP library's FFT functions are 2-3 times more
* efficient on the modern Intel Core i7-3770K processor than Ooura's
* functions. It may be worthwhile investing in IPP. Note, that FFT functions
* take less than 20% of the overall sample rate conversion time. However,
* when the "power of 2" resampling is used the performance of FFT functions
* becomes "everything".
*/
class CDSPRealFFT : public R8B_BASECLASS
{
R8BNOCTOR( CDSPRealFFT );
friend class CDSPRealFFTKeeper;
public:
/**
* @return A multiplication constant that should be used after inverse
* transform to obtain a correct value scale.
*/
double getInvMulConst() const
{
return( InvMulConst );
}
/**
* @return The length (the number of real values in a transform) of *this
* FFT object, expressed as Nth power of 2.
*/
int getLenBits() const
{
return( LenBits );
}
/**
* @return The length (the number of real values in a transform) of *this
* FFT object.
*/
int getLen() const
{
return( Len );
}
/**
* Function performs in-place forward FFT.
*
* @param[in,out] p Pointer to data block to transform, length should be
* equal to *this object's getLen().
*/
void forward( double* const p ) const
{
#if R8B_FLOATFFT
float* const op = (float*) p;
int i;
for( i = 0; i < Len; i++ )
{
op[ i ] = (float) p[ i ];
}
#endif // R8B_FLOATFFT
#if R8B_IPP
ippsFFTFwd_RToPerm_64f( p, p, SPtr, WorkBuffer );
#elif R8B_PFFFT
pffft_transform_ordered( setup, op, op, work, PFFFT_FORWARD );
#elif R8B_PFFFT_DOUBLE
pffftd_transform_ordered( setup, p, p, work, PFFFT_FORWARD );
#else // R8B_PFFFT_DOUBLE
ooura_fft :: rdft( Len, 1, p, wi.getPtr(), wd.getPtr() );
#endif // R8B_IPP
}
/**
* Function performs in-place inverse FFT.
*
* @param[in,out] p Pointer to data block to transform, length should be
* equal to *this object's getLen().
*/
void inverse( double* const p ) const
{
#if R8B_IPP
ippsFFTInv_PermToR_64f( p, p, SPtr, WorkBuffer );
#elif R8B_PFFFT
pffft_transform_ordered( setup, (float*) p, (float*) p, work,
PFFFT_BACKWARD );
#elif R8B_PFFFT_DOUBLE
pffftd_transform_ordered( setup, p, p, work, PFFFT_BACKWARD );
#else // R8B_PFFFT_DOUBLE
ooura_fft :: rdft( Len, -1, p, wi.getPtr(), wd.getPtr() );
#endif // R8B_IPP
#if R8B_FLOATFFT
const float* const ip = (const float*) p;
int i;
for( i = Len - 1; i >= 0; i-- )
{
p[ i ] = ip[ i ];
}
#endif // R8B_FLOATFFT
}
/**
* Function multiplies two complex-valued data blocks and places result in
* a new data block. Length of all data blocks should be equal to *this
* object's block length. Input blocks should have been produced with the
* forward() function of *this object.
*
* @param aip1 Input data block 1.
* @param aip2 Input data block 2.
* @param[out] aop Output data block, should not be equal to aip1 nor
* aip2.
*/
void multiplyBlocks( const double* const aip1, const double* const aip2,
double* const aop ) const
{
#if R8B_FLOATFFT
const float* const ip1 = (const float*) aip1;
const float* const ip2 = (const float*) aip2;
float* const op = (float*) aop;
#else // R8B_FLOATFFT
const double* const ip1 = aip1;
const double* const ip2 = aip2;
double* const op = aop;
#endif // R8B_FLOATFFT
#if R8B_IPP
ippsMulPerm_64f( (Ipp64f*) ip1, (Ipp64f*) ip2, (Ipp64f*) op, Len );
#else // R8B_IPP
op[ 0 ] = ip1[ 0 ] * ip2[ 0 ];
op[ 1 ] = ip1[ 1 ] * ip2[ 1 ];
int i = 2;
while( i < Len )
{
op[ i ] = ip1[ i ] * ip2[ i ] - ip1[ i + 1 ] * ip2[ i + 1 ];
op[ i + 1 ] = ip1[ i ] * ip2[ i + 1 ] + ip1[ i + 1 ] * ip2[ i ];
i += 2;
}
#endif // R8B_IPP
}
/**
* Function multiplies two complex-valued data blocks in-place. Length of
* both data blocks should be equal to *this object's block length. Blocks
* should have been produced with the forward() function of *this object.
*
* @param aip Input data block 1.
* @param[in,out] aop Output/input data block 2.
*/
void multiplyBlocks( const double* const aip, double* const aop ) const
{
#if R8B_FLOATFFT
const float* const ip = (const float*) aip;
float* const op = (float*) aop;
float t;
#else // R8B_FLOATFFT
const double* const ip = aip;
double* const op = aop;
#if !R8B_IPP
double t;
#endif // !R8B_IPP
#endif // R8B_FLOATFFT
#if R8B_IPP
ippsMulPerm_64f( (Ipp64f*) op, (Ipp64f*) ip, (Ipp64f*) op, Len );
#else // R8B_IPP
op[ 0 ] *= ip[ 0 ];
op[ 1 ] *= ip[ 1 ];
int i = 2;
while( i < Len )
{
t = op[ i ] * ip[ i ] - op[ i + 1 ] * ip[ i + 1 ];
op[ i + 1 ] = op[ i ] * ip[ i + 1 ] + op[ i + 1 ] * ip[ i ];
op[ i ] = t;
i += 2;
}
#endif // R8B_IPP
}
/**
* Function multiplies two complex-valued data blocks in-place,
* considering that the "ip" block contains "zero-phase" response. Length
* of both data blocks should be equal to *this object's block length.
* Blocks should have been produced with the forward() function of *this
* object.
*
* @param aip Input data block 1, "zero-phase" response. This block should
* be first transformed via the convertToZP() function.
* @param[in,out] aop Output/input data block 2.
*/
void multiplyBlocksZP( const double* const aip, double* const aop ) const
{
#if R8B_FLOATFFT
const float* const ip = (const float*) aip;
float* const op = (float*) aop;
#else // R8B_FLOATFFT
const double* ip = aip;
double* op = aop;
#endif // R8B_FLOATFFT
// SIMD implementations assume that pointers are address-aligned.
#if !R8B_FLOATFFT && defined( R8B_SSE2 )
int c8 = Len >> 3;
while( c8 != 0 )
{
const __m128d iv1 = _mm_load_pd( ip );
const __m128d iv2 = _mm_load_pd( ip + 2 );
const __m128d ov1 = _mm_load_pd( op );
const __m128d ov2 = _mm_load_pd( op + 2 );
_mm_store_pd( op, _mm_mul_pd( iv1, ov1 ));
_mm_store_pd( op + 2, _mm_mul_pd( iv2, ov2 ));
const __m128d iv3 = _mm_load_pd( ip + 4 );
const __m128d ov3 = _mm_load_pd( op + 4 );
const __m128d iv4 = _mm_load_pd( ip + 6 );
const __m128d ov4 = _mm_load_pd( op + 6 );
_mm_store_pd( op + 4, _mm_mul_pd( iv3, ov3 ));
_mm_store_pd( op + 6, _mm_mul_pd( iv4, ov4 ));
ip += 8;
op += 8;
c8--;
}
int c = Len & 7;
while( c != 0 )
{
*op *= *ip;
ip++;
op++;
c--;
}
#elif !R8B_FLOATFFT && defined( R8B_NEON )
int c8 = Len >> 3;
while( c8 != 0 )
{
const float64x2_t iv1 = vld1q_f64( ip );
const float64x2_t iv2 = vld1q_f64( ip + 2 );
const float64x2_t ov1 = vld1q_f64( op );
const float64x2_t ov2 = vld1q_f64( op + 2 );
vst1q_f64( op, vmulq_f64( iv1, ov1 ));
vst1q_f64( op + 2, vmulq_f64( iv2, ov2 ));
const float64x2_t iv3 = vld1q_f64( ip + 4 );
const float64x2_t iv4 = vld1q_f64( ip + 6 );
const float64x2_t ov3 = vld1q_f64( op + 4 );
const float64x2_t ov4 = vld1q_f64( op + 6 );
vst1q_f64( op + 4, vmulq_f64( iv3, ov3 ));
vst1q_f64( op + 6, vmulq_f64( iv4, ov4 ));
ip += 8;
op += 8;
c8--;
}
int c = Len & 7;
while( c != 0 )
{
*op *= *ip;
ip++;
op++;
c--;
}
#else // SIMD
int i;
for( i = 0; i < Len; i++ )
{
op[ i ] *= ip[ i ];
}
#endif // SIMD
}
/**
* Function converts the specified forward-transformed block into
* "zero-phase" form suitable for use with the multiplyBlocksZ() function.
*
* @param[in,out] ap Block to transform.
*/
void convertToZP( double* const ap ) const
{
#if R8B_FLOATFFT
float* const p = (float*) ap;
#else // R8B_FLOATFFT
double* const p = ap;
#endif // R8B_FLOATFFT
int i = 2;
while( i < Len )
{
p[ i + 1 ] = p[ i ];
i += 2;
}
}
private:
int LenBits; ///< Length of FFT block (expressed as Nth power of 2).
///<
int Len; ///< Length of FFT block (number of real values).
///<
double InvMulConst; ///< Inverse FFT multiply constant.
///<
CDSPRealFFT* Next; ///< Next object in a singly-linked list.
///<
#if R8B_IPP
IppsFFTSpec_R_64f* SPtr; ///< Pointer to initialized data buffer
///< to be passed to IPP's FFT functions.
///<
CFixedBuffer< unsigned char > SpecBuffer; ///< Working buffer.
///<
CFixedBuffer< unsigned char > WorkBuffer; ///< Working buffer.
///<
#elif R8B_PFFFT
PFFFT_Setup* setup; ///< PFFFT setup object.
///<
CFixedBuffer< float > work; ///< Working buffer.
///<
#elif R8B_PFFFT_DOUBLE
PFFFTD_Setup* setup; ///< PFFFTD setup object.
///<
CFixedBuffer< double > work; ///< Working buffer.
///<
#else // R8B_PFFFT_DOUBLE
CFixedBuffer< int > wi; ///< Working buffer (ints).
///<
CFixedBuffer< double > wd; ///< Working buffer (doubles).
///<
#endif // R8B_IPP
/**
* A simple class that keeps the pointer to the object and deletes it
* automatically.
*/
class CObjKeeper
{
R8BNOCTOR( CObjKeeper );
public:
CObjKeeper()
: Object( NULL )
{
}
~CObjKeeper()
{
delete Object;
}
CObjKeeper& operator = ( CDSPRealFFT* const aObject )
{
Object = aObject;
return( *this );
}
operator CDSPRealFFT* () const
{
return( Object );
}
private:
CDSPRealFFT* Object; ///< FFT object being kept.
///<
};
CDSPRealFFT()
{
}
/**
* Constructor initializes FFT object.
*
* @param aLenBits The length of FFT block (Nth power of 2), specifies the
* number of real values in a block. Values from 1 to 30 inclusive are
* supported.
*/
CDSPRealFFT( const int aLenBits )
: LenBits( aLenBits )
, Len( 1 << aLenBits )
#if R8B_IPP
, InvMulConst( 1.0 / Len )
#elif R8B_PFFFT
, InvMulConst( 1.0 / Len )
#elif R8B_PFFFT_DOUBLE
, InvMulConst( 1.0 / Len )
#else // R8B_PFFFT_DOUBLE
, InvMulConst( 2.0 / Len )
#endif // R8B_IPP
{
#if R8B_IPP
int SpecSize;
int SpecBufferSize;
int BufferSize;
ippsFFTGetSize_R_64f( LenBits, IPP_FFT_NODIV_BY_ANY,
ippAlgHintFast, &SpecSize, &SpecBufferSize, &BufferSize );
CFixedBuffer< unsigned char > InitBuffer( SpecBufferSize );
SpecBuffer.alloc( SpecSize );
WorkBuffer.alloc( BufferSize );
ippsFFTInit_R_64f( &SPtr, LenBits, IPP_FFT_NODIV_BY_ANY,
ippAlgHintFast, SpecBuffer, InitBuffer );
#elif R8B_PFFFT
setup = pffft_new_setup( Len, PFFFT_REAL );
work.alloc( Len );
#elif R8B_PFFFT_DOUBLE
setup = pffftd_new_setup( Len, PFFFT_REAL );
work.alloc( Len );
#else // R8B_PFFFT_DOUBLE
wi.alloc( (int) ceil( 2.0 + sqrt( (double) ( Len >> 1 ))));
wi[ 0 ] = 0;
wd.alloc( Len >> 1 );
#endif // R8B_IPP
}
~CDSPRealFFT()
{
#if R8B_PFFFT
pffft_destroy_setup( setup );
#elif R8B_PFFFT_DOUBLE
pffftd_destroy_setup( setup );
#endif // R8B_PFFFT_DOUBLE
delete Next;
}
};
/**
* @brief A "keeper" class for real-valued FFT transform objects.
*
* Class implements "keeper" functionality for handling CDSPRealFFT objects.
* The allocated FFT objects are placed on the global static list of objects
* for future reuse instead of deallocation.
*/
class CDSPRealFFTKeeper : public R8B_BASECLASS
{
R8BNOCTOR( CDSPRealFFTKeeper );
public:
CDSPRealFFTKeeper()
: Object( NULL )
{
}
/**
* Function acquires FFT object with the specified block length.
*
* @param LenBits The length of FFT block (Nth power of 2), in the range
* [1; 30] inclusive, specifies the number of real values in a FFT block.
*/
CDSPRealFFTKeeper( const int LenBits )
{
Object = acquire( LenBits );
}
~CDSPRealFFTKeeper()
{
if( Object != NULL )
{
release( Object );
}
}
/**
* @return Pointer to the acquired FFT object.
*/
const CDSPRealFFT* operator -> () const
{
R8BASSERT( Object != NULL );
return( Object );
}
/**
* Function acquires FFT object with the specified block length. This
* function can be called any number of times.
*
* @param LenBits The length of FFT block (Nth power of 2), in the range
* [1; 30] inclusive, specifies the number of real values in a FFT block.
*/
void init( const int LenBits )
{
if( Object != NULL )
{
if( Object -> LenBits == LenBits )
{
return;
}
release( Object );
}
Object = acquire( LenBits );
}
/**
* Function releases a previously acquired FFT object.
*/
void reset()
{
if( Object != NULL )
{
release( Object );
Object = NULL;
}
}
private:
CDSPRealFFT* Object; ///< FFT object.
///<
static CSyncObject StateSync; ///< FFTObjects synchronizer.
///<
static CDSPRealFFT :: CObjKeeper FFTObjects[]; ///< Pool of FFT objects of
///< various lengths.
///<
/**
* Function acquires FFT object from the global pool.
*
* @param LenBits FFT block length (expressed as Nth power of 2).
*/
CDSPRealFFT* acquire( const int LenBits )
{
R8BASSERT( LenBits > 0 && LenBits <= 30 );
R8BSYNC( StateSync );
if( FFTObjects[ LenBits ] == NULL )
{
return( new CDSPRealFFT( LenBits ));
}
CDSPRealFFT* ffto = FFTObjects[ LenBits ];
FFTObjects[ LenBits ] = ffto -> Next;
return( ffto );
}
/**
* Function releases a previously acquired FFT object.
*
* @param ffto FFT object to release.
*/
void release( CDSPRealFFT* const ffto )
{
R8BSYNC( StateSync );
ffto -> Next = FFTObjects[ ffto -> LenBits ];
FFTObjects[ ffto -> LenBits ] = ffto;
}
};
/**
* Function calculates the minimum-phase transform of the filter kernel, using
* a discrete Hilbert transform in cepstrum domain.
*
* For more details, see part III.B of
* http://www.hpl.hp.com/personal/Niranjan_Damera-Venkata/files/ComplexMinPhase.pdf
*
* @param[in,out] Kernel Filter kernel buffer.
* @param KernelLen Filter kernel's length, in samples.
* @param LenMult Kernel length multiplier. Used as a coefficient of the
* "oversampling" in the frequency domain. Such oversampling is needed to
* improve the precision of the minimum-phase transform. If the filter's
* attenuation is high, this multiplier should be increased or otherwise the
* required attenuation will not be reached due to "smoothing" effect of this
* transform.
* @param DoFinalMul "True" if the final multiplication after transform should
* be performed or not. Such multiplication returns the gain of the signal to
* its original value. This parameter can be set to "false" if normalization
* of the resulting filter kernel is planned to be used.
* @param[out] DCGroupDelay If not NULL, this variable receives group delay
* at DC offset, in samples (can be a non-integer value).
*/
inline void calcMinPhaseTransform( double* const Kernel, const int KernelLen,
const int LenMult = 2, const bool DoFinalMul = true,
double* const DCGroupDelay = NULL )
{
R8BASSERT( KernelLen > 0 );
R8BASSERT( LenMult >= 2 );
const int LenBits = getBitOccupancy(( KernelLen * LenMult ) - 1 );
const int Len = 1 << LenBits;
const int Len2 = Len >> 1;
int i;
CFixedBuffer< double > ip( Len );
CFixedBuffer< double > ip2( Len2 + 1 );
memcpy( &ip[ 0 ], Kernel, KernelLen * sizeof( ip[ 0 ]));
memset( &ip[ KernelLen ], 0, ( Len - KernelLen ) * sizeof( ip[ 0 ]));
CDSPRealFFTKeeper ffto( LenBits );
ffto -> forward( ip );
// Create the "log |c|" spectrum while saving the original power spectrum
// in the "ip2" buffer.
#if R8B_FLOATFFT
float* const aip = (float*) &ip[ 0 ];
float* const aip2 = (float*) &ip2[ 0 ];
const float nzbias = 1e-35;
#else // R8B_FLOATFFT
double* const aip = &ip[ 0 ];
double* const aip2 = &ip2[ 0 ];
const double nzbias = 1e-300;
#endif // R8B_FLOATFFT
aip2[ 0 ] = aip[ 0 ];
aip[ 0 ] = log( fabs( aip[ 0 ]) + nzbias );
aip2[ Len2 ] = aip[ 1 ];
aip[ 1 ] = log( fabs( aip[ 1 ]) + nzbias );
for( i = 1; i < Len2; i++ )
{
aip2[ i ] = sqrt( aip[ i * 2 ] * aip[ i * 2 ] +
aip[ i * 2 + 1 ] * aip[ i * 2 + 1 ]);
aip[ i * 2 ] = log( aip2[ i ] + nzbias );
aip[ i * 2 + 1 ] = 0.0;
}
// Convert to cepstrum and apply discrete Hilbert transform.
ffto -> inverse( ip );
const double m1 = ffto -> getInvMulConst();
const double m2 = -m1;
ip[ 0 ] = 0.0;
for( i = 1; i < Len2; i++ )
{
ip[ i ] *= m1;
}
ip[ Len2 ] = 0.0;
for( i = Len2 + 1; i < Len; i++ )
{
ip[ i ] *= m2;
}
// Convert Hilbert-transformed cepstrum back to the "log |c|" spectrum and
// perform its exponentiation, multiplied by the power spectrum previously
// saved in the "ip2" buffer.
ffto -> forward( ip );
aip[ 0 ] = aip2[ 0 ];
aip[ 1 ] = aip2[ Len2 ];
for( i = 1; i < Len2; i++ )
{
aip[ i * 2 + 0 ] = cos( aip[ i * 2 + 1 ]) * aip2[ i ];
aip[ i * 2 + 1 ] = sin( aip[ i * 2 + 1 ]) * aip2[ i ];
}
ffto -> inverse( ip );
if( DoFinalMul )
{
for( i = 0; i < KernelLen; i++ )
{
Kernel[ i ] = ip[ i ] * m1;
}
}
else
{
memcpy( &Kernel[ 0 ], &ip[ 0 ], KernelLen * sizeof( Kernel[ 0 ]));
}
if( DCGroupDelay != NULL )
{
double tmp;
calcFIRFilterResponseAndGroupDelay( Kernel, KernelLen, 0.0,
tmp, tmp, *DCGroupDelay );
}
}
} // namespace r8b
#endif // VOX_CDSPREALFFT_INCLUDED