VE7UWU/sdrangel
1
0
Fork 0
mirror of https://github.com/f4exb/sdrangel.git synced 2024-11-17 13:51:47 -05:00
Code Issues Projects Releases Wiki Activity

IntHalfBandFilterEO2: use dual forward and backward buffers to avoid byte shuffling in SIMD instructions. Implemented in the up channelizer

Browse Source
This commit is contained in:
f4exb 2016-11-06 01:07:13 +01:00
parent de8640caae
commit bc3dfb19cd
4 changed files with 676 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
CMakeLists.txt
View File

@ -214,6 +214,7 @@ set(sdrbase_HEADERS
sdrbase/dsp/inthalfbandfilter.h
sdrbase/dsp/inthalfbandfilterdb.h
sdrbase/dsp/inthalfbandfiltereo1.h
sdrbase/dsp/inthalfbandfiltereo2.h
sdrbase/dsp/kissfft.h
sdrbase/dsp/kissengine.h
sdrbase/dsp/lowpass.h

668
sdrbase/dsp/inthalfbandfiltereo2.h Normal file
View File

@ -0,0 +1,668 @@
///////////////////////////////////////////////////////////////////////////////////
// Copyright (C) 2016 F4EXB //
// written by Edouard Griffiths //
// //
// Integer half-band FIR based interpolator and decimator //
// This is the even/odd double buffer variant. Really useful only when SIMD is //
// used //
// //
// This program is free software; you can redistribute it and/or modify //
// it under the terms of the GNU General Public License as published by //
// the Free Software Foundation as version 3 of the License, or //
// //
// This program is distributed in the hope that it will be useful, //
// but WITHOUT ANY WARRANTY; without even the implied warranty of //
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the //
// GNU General Public License V3 for more details. //
// //
// You should have received a copy of the GNU General Public License //
// along with this program. If not, see <http://www.gnu.org/licenses/>. //
///////////////////////////////////////////////////////////////////////////////////
#ifndef SDRBASE_DSP_INTHALFBANDFILTEREO2_H_
#define SDRBASE_DSP_INTHALFBANDFILTEREO2_H_
#ifdef USE_SIMD
#include <smmintrin.h>
#endif
#include <stdint.h>
#include "dsp/dsptypes.h"
#include "dsp/hbfiltertraits.h"
#include "util/export.h"
template<uint32_t HBFilterOrder>
class SDRANGEL_API IntHalfbandFilterEO2 {
public:
IntHalfbandFilterEO2();
// downsample by 2, return center part of original spectrum
bool workDecimateCenter(Sample* sample)
{
// insert sample into ring-buffer
storeSample((FixReal) sample->real(), (FixReal) sample->imag());
switch(m_state)
{
case 0:
// advance write-pointer
advancePointer();
// next state
m_state = 1;
// tell caller we don't have a new sample
return false;
default:
// save result
doFIR(sample);
// advance write-pointer
advancePointer();
// next state
m_state = 0;
// tell caller we have a new sample
return true;
}
}
// upsample by 2, return center part of original spectrum - double buffer variant
bool workInterpolateCenter(Sample* sampleIn, Sample *SampleOut)
{
switch(m_state)
{
case 0:
// insert sample into ring-buffer
storeSample(0, 0);
// save result
doFIR(SampleOut);
// advance write-pointer
advancePointer();
// next state
m_state = 1;
// tell caller we didn't consume the sample
return false;
default:
// insert sample into ring-buffer
storeSample((FixReal) sampleIn->real(), (FixReal) sampleIn->imag());
// save result
doFIR(SampleOut);
// advance write-pointer
advancePointer();
// next state
m_state = 0;
// tell caller we consumed the sample
return true;
}
}
bool workDecimateCenter(qint32 *x, qint32 *y)
{
// insert sample into ring-buffer
storeSample(*x, *y);
switch(m_state)
{
case 0:
// advance write-pointer
advancePointer();
// next state
m_state = 1;
// tell caller we don't have a new sample
return false;
default:
// save result
doFIR(x, y);
// advance write-pointer
advancePointer();
// next state
m_state = 0;
// tell caller we have a new sample
return true;
}
}
// downsample by 2, return lower half of original spectrum
bool workDecimateLowerHalf(Sample* sample)
{
switch(m_state)
{
case 0:
// insert sample into ring-buffer
storeSample((FixReal) -sample->imag(), (FixReal) sample->real());
// advance write-pointer
advancePointer();
// next state
m_state = 1;
// tell caller we don't have a new sample
return false;
case 1:
// insert sample into ring-buffer
storeSample((FixReal) -sample->real(), (FixReal) -sample->imag());
// save result
doFIR(sample);
// advance write-pointer
advancePointer();
// next state
m_state = 2;
// tell caller we have a new sample
return true;
case 2:
// insert sample into ring-buffer
storeSample((FixReal) sample->imag(), (FixReal) -sample->real());
// advance write-pointer
advancePointer();
// next state
m_state = 3;
// tell caller we don't have a new sample
return false;
default:
// insert sample into ring-buffer
storeSample((FixReal) sample->real(), (FixReal) sample->imag());
// save result
doFIR(sample);
// advance write-pointer
advancePointer();
// next state
m_state = 0;
// tell caller we have a new sample
return true;
}
}
// upsample by 2, from lower half of original spectrum - double buffer variant
bool workInterpolateLowerHalf(Sample* sampleIn, Sample *sampleOut)
{
Sample s;
switch(m_state)
{
case 0:
// insert sample into ring-buffer
storeSample(0, 0);
// save result
doFIR(&s);
sampleOut->setReal(s.imag());
sampleOut->setImag(-s.real());
// advance write-pointer
advancePointer();
// next state
m_state = 1;
// tell caller we didn't consume the sample
return false;
case 1:
// insert sample into ring-buffer
storeSample((FixReal) sampleIn->real(), (FixReal) sampleIn->imag());
// save result
doFIR(&s);
sampleOut->setReal(-s.real());
sampleOut->setImag(-s.imag());
// advance write-pointer
advancePointer();
// next state
m_state = 2;
// tell caller we consumed the sample
return true;
case 2:
// insert sample into ring-buffer
storeSample(0, 0);
// save result
doFIR(&s);
sampleOut->setReal(-s.imag());
sampleOut->setImag(s.real());
// advance write-pointer
advancePointer();
// next state
m_state = 3;
// tell caller we didn't consume the sample
return false;
default:
// insert sample into ring-buffer
storeSample((FixReal) sampleIn->real(), (FixReal) sampleIn->imag());
// save result
doFIR(&s);
sampleOut->setReal(s.real());
sampleOut->setImag(s.imag());
// advance write-pointer
advancePointer();
// next state
m_state = 0;
// tell caller we consumed the sample
return true;
}
}
// downsample by 2, return upper half of original spectrum
bool workDecimateUpperHalf(Sample* sample)
{
switch(m_state)
{
case 0:
// insert sample into ring-buffer
storeSample((FixReal) sample->imag(), (FixReal) -sample->real());
// advance write-pointer
advancePointer();
// next state
m_state = 1;
// tell caller we don't have a new sample
return false;
case 1:
// insert sample into ring-buffer
storeSample((FixReal) -sample->real(), (FixReal) -sample->imag());
// save result
doFIR(sample);
// advance write-pointer
advancePointer();
// next state
m_state = 2;
// tell caller we have a new sample
return true;
case 2:
// insert sample into ring-buffer
storeSample((FixReal) -sample->imag(), (FixReal) sample->real());
// advance write-pointer
advancePointer();
// next state
m_state = 3;
// tell caller we don't have a new sample
return false;
default:
// insert sample into ring-buffer
storeSample((FixReal) sample->real(), (FixReal) sample->imag());
// save result
doFIR(sample);
// advance write-pointer
advancePointer();
// next state
m_state = 0;
// tell caller we have a new sample
return true;
}
}
// upsample by 2, move original spectrum to upper half - double buffer variant
bool workInterpolateUpperHalf(Sample* sampleIn, Sample *sampleOut)
{
Sample s;
switch(m_state)
{
case 0:
// insert sample into ring-buffer
storeSample(0, 0);
// save result
doFIR(&s);
sampleOut->setReal(-s.imag());
sampleOut->setImag(s.real());
// advance write-pointer
advancePointer();
// next state
m_state = 1;
// tell caller we didn't consume the sample
return false;
case 1:
// insert sample into ring-buffer
storeSample((FixReal) sampleIn->real(), (FixReal) sampleIn->imag());
// save result
doFIR(&s);
sampleOut->setReal(-s.real());
sampleOut->setImag(-s.imag());
// advance write-pointer
advancePointer();
// next state
m_state = 2;
// tell caller we consumed the sample
return true;
case 2:
// insert sample into ring-buffer
storeSample(0, 0);
// save result
doFIR(&s);
sampleOut->setReal(s.imag());
sampleOut->setImag(-s.real());
// advance write-pointer
advancePointer();
// next state
m_state = 3;
// tell caller we didn't consume the sample
return false;
default:
// insert sample into ring-buffer
storeSample((FixReal) sampleIn->real(), (FixReal) sampleIn->imag());
// save result
doFIR(&s);
sampleOut->setReal(s.real());
sampleOut->setImag(s.imag());
// advance write-pointer
advancePointer();
// next state
m_state = 0;
// tell caller we consumed the sample
return true;
}
}
void myDecimate(const Sample* sample1, Sample* sample2)
{
storeSample((FixReal) sample1->real(), (FixReal) sample1->imag());
advancePointer();
storeSample((FixReal) sample2->real(), (FixReal) sample2->imag());
doFIR(sample2);
advancePointer();
}
void myDecimate(qint32 x1, qint32 y1, qint32 *x2, qint32 *y2)
{
storeSample(x1, y1);
advancePointer();
storeSample(*x2, *y2);
doFIR(x2, y2);
advancePointer();
}
protected:
qint32 m_evenB[2][HBFIRFilterTraits<HBFilterOrder>::hbOrder]; // double buffer technique
qint32 m_oddB[2][HBFIRFilterTraits<HBFilterOrder>::hbOrder]; // double buffer technique
qint32 m_evenA[2][HBFIRFilterTraits<HBFilterOrder>::hbOrder]; // double buffer technique
qint32 m_oddA[2][HBFIRFilterTraits<HBFilterOrder>::hbOrder]; // double buffer technique
int m_ptrA;
int m_ptrB;
int m_size;
int m_state;
void storeSample(const FixReal& sampleI, const FixReal& sampleQ)
{
if ((m_ptrB % 2) == 0)
{
m_evenB[0][m_ptrB/2] = sampleI;
m_evenB[1][m_ptrB/2] = sampleQ;
m_evenB[0][m_ptrB/2 + m_size] = sampleI;
m_evenB[1][m_ptrB/2 + m_size] = sampleQ;
m_evenA[0][m_ptrA/2] = sampleI;
m_evenA[1][m_ptrA/2] = sampleQ;
m_evenA[0][m_ptrA/2 + m_size] = sampleI;
m_evenA[1][m_ptrA/2 + m_size] = sampleQ;
}
else
{
m_oddB[0][m_ptrB/2] = sampleI;
m_oddB[1][m_ptrB/2] = sampleQ;
m_oddB[0][m_ptrB/2 + m_size] = sampleI;
m_oddB[1][m_ptrB/2 + m_size] = sampleQ;
m_oddA[0][m_ptrA/2] = sampleI;
m_oddA[1][m_ptrA/2] = sampleQ;
m_oddA[0][m_ptrA/2 + m_size] = sampleI;
m_oddA[1][m_ptrA/2 + m_size] = sampleQ;
}
}
void storeSample(qint32 x, qint32 y)
{
if ((m_ptrB % 2) == 0)
{
m_evenB[0][m_ptrB/2] = x;
m_evenB[1][m_ptrB/2] = y;
m_evenB[0][m_ptrB/2 + m_size] = x;
m_evenB[1][m_ptrB/2 + m_size] = y;
m_evenA[0][m_ptrA/2] = x;
m_evenA[1][m_ptrA/2] = y;
m_evenA[0][m_ptrA/2 + m_size] = x;
m_evenA[1][m_ptrA/2 + m_size] = y;
}
else
{
m_oddB[0][m_ptrB/2] = x;
m_oddB[1][m_ptrB/2] = y;
m_oddB[0][m_ptrB/2 + m_size] = x;
m_oddB[1][m_ptrB/2 + m_size] = y;
m_oddA[0][m_ptrA/2] = x;
m_oddA[1][m_ptrA/2] = y;
m_oddA[0][m_ptrA/2 + m_size] = x;
m_oddA[1][m_ptrA/2 + m_size] = y;
}
}
void advancePointer()
{
m_ptrA = (m_ptrA - 1 + 2*m_size) % (2*m_size);
m_ptrB = (m_ptrB + 1) % (2*m_size);
}
void doFIR(Sample* sample)
{
int a = m_ptrA/2; // tip pointer
int b = m_ptrB/2 + 1; // tail pointer
qint32 iAcc = 0;
qint32 qAcc = 0;
#ifdef USE_SIMD
//#warning "IntHalfbandFiler SIMD"
const __m128i* h = (const __m128i*) HBFIRFilterTraits<HBFilterOrder>::hbCoeffs;
__m128i sumI = _mm_setzero_si128();
__m128i sumQ = _mm_setzero_si128();
__m128i sa, sb;
for (int i = 0; i < HBFIRFilterTraits<HBFilterOrder>::hbOrder / 16; i++)
{
if ((m_ptrB % 2) == 0)
{
sa = _mm_loadu_si128((__m128i*) &(m_evenA[0][a]));
sb = _mm_loadu_si128((__m128i*) &(m_evenB[0][b]));
sumI = _mm_add_epi32(sumI, _mm_mullo_epi32(_mm_add_epi32(sa, sb), *h));
sa = _mm_loadu_si128((__m128i*) &(m_evenA[1][a]));
sb = _mm_loadu_si128((__m128i*) &(m_evenB[1][b]));
sumQ = _mm_add_epi32(sumQ, _mm_mullo_epi32(_mm_add_epi32(sa, sb), *h));
}
else
{
sa = _mm_loadu_si128((__m128i*) &(m_oddA[0][a]));
sb = _mm_loadu_si128((__m128i*) &(m_oddB[0][b]));
sumI = _mm_add_epi32(sumI, _mm_mullo_epi32(_mm_add_epi32(sa, sb), *h));
sa = _mm_loadu_si128((__m128i*) &(m_oddA[1][a]));
sb = _mm_loadu_si128((__m128i*) &(m_oddB[1][b]));
sumQ = _mm_add_epi32(sumQ, _mm_mullo_epi32(_mm_add_epi32(sa, sb), *h));
}
a += 4;
b += 4;
++h;
}
// horizontal add of four 32 bit partial sums
sumI = _mm_add_epi32(sumI, _mm_srli_si128(sumI, 8));
sumI = _mm_add_epi32(sumI, _mm_srli_si128(sumI, 4));
iAcc = _mm_cvtsi128_si32(sumI);
sumQ = _mm_add_epi32(sumQ, _mm_srli_si128(sumQ, 8));
sumQ = _mm_add_epi32(sumQ, _mm_srli_si128(sumQ, 4));
qAcc = _mm_cvtsi128_si32(sumQ);
#else
for (int i = 0; i < HBFIRFilterTraits<HBFilterOrder>::hbOrder / 4; i++)
{
if ((m_ptrB % 2) == 0)
{
iAcc += (m_evenA[0][a] + m_evenB[0][b]) * HBFIRFilterTraits<HBFilterOrder>::hbCoeffs[i];
qAcc += (m_evenA[1][a] + m_evenB[1][b]) * HBFIRFilterTraits<HBFilterOrder>::hbCoeffs[i];
}
else
{
iAcc += (m_oddA[0][a] + m_oddB[0][b]) * HBFIRFilterTraits<HBFilterOrder>::hbCoeffs[i];
qAcc += (m_oddA[1][a] + m_oddB[1][b]) * HBFIRFilterTraits<HBFilterOrder>::hbCoeffs[i];
}
a += 1;
b += 1;
}
#endif
if ((m_ptrB % 2) == 0)
{
iAcc += ((qint32)m_oddB[0][m_ptrB/2 + m_size/2]) << (HBFIRFilterTraits<HBFilterOrder>::hbShift - 1);
qAcc += ((qint32)m_oddB[1][m_ptrB/2 + m_size/2]) << (HBFIRFilterTraits<HBFilterOrder>::hbShift - 1);
}
else
{
iAcc += ((qint32)m_evenB[0][m_ptrB/2 + m_size/2 + 1]) << (HBFIRFilterTraits<HBFilterOrder>::hbShift - 1);
qAcc += ((qint32)m_evenB[1][m_ptrB/2 + m_size/2 + 1]) << (HBFIRFilterTraits<HBFilterOrder>::hbShift - 1);
}
sample->setReal(iAcc >> HBFIRFilterTraits<HBFilterOrder>::hbShift -1);
sample->setImag(qAcc >> HBFIRFilterTraits<HBFilterOrder>::hbShift -1);
}
void doFIR(qint32 *x, qint32 *y)
{
int a = m_ptrA/2; // tip pointer
int b = m_ptrB/2 + 1; // tail pointer
qint32 iAcc = 0;
qint32 qAcc = 0;
#ifdef USE_SIMD
const __m128i* h = (const __m128i*) HBFIRFilterTraits<HBFilterOrder>::hbCoeffs;
__m128i sumI = _mm_setzero_si128();
__m128i sumQ = _mm_setzero_si128();
__m128i sa, sb;
for (int i = 0; i < HBFIRFilterTraits<HBFilterOrder>::hbOrder / 16; i++)
{
if ((m_ptrB % 2) == 0)
{
sa = _mm_loadu_si128((__m128i*) &(m_evenA[0][a]));
sb = _mm_loadu_si128((__m128i*) &(m_evenB[0][b]));
sumI = _mm_add_epi32(sumI, _mm_mullo_epi32(_mm_add_epi32(sa, sb), *h));
sa = _mm_loadu_si128((__m128i*) &(m_evenB[1][a]));
sb = _mm_loadu_si128((__m128i*) &(m_evenB[1][b]));
sumQ = _mm_add_epi32(sumQ, _mm_mullo_epi32(_mm_add_epi32(sa, sb), *h));
}
else
{
sa = _mm_loadu_si128((__m128i*) &(m_oddA[0][a]));
sb = _mm_loadu_si128((__m128i*) &(m_oddB[0][b]));
sumI = _mm_add_epi32(sumI, _mm_mullo_epi32(_mm_add_epi32(sa, sb), *h));
sa = _mm_loadu_si128((__m128i*) &(m_oddB[1][a]));
sb = _mm_loadu_si128((__m128i*) &(m_oddB[1][b]));
sumQ = _mm_add_epi32(sumQ, _mm_mullo_epi32(_mm_add_epi32(sa, sb), *h));
}
a += 4;
b += 4;
++h;
}
// horizontal add of four 32 bit partial sums
sumI = _mm_add_epi32(sumI, _mm_srli_si128(sumI, 8));
sumI = _mm_add_epi32(sumI, _mm_srli_si128(sumI, 4));
iAcc = _mm_cvtsi128_si32(sumI);
sumQ = _mm_add_epi32(sumQ, _mm_srli_si128(sumQ, 8));
sumQ = _mm_add_epi32(sumQ, _mm_srli_si128(sumQ, 4));
qAcc = _mm_cvtsi128_si32(sumQ);
#else
for (int i = 0; i < HBFIRFilterTraits<HBFilterOrder>::hbOrder / 4; i++)
{
if ((m_ptrB % 2) == 0)
{
iAcc += (m_evenA[0][a] + m_evenB[0][b]) * HBFIRFilterTraits<HBFilterOrder>::hbCoeffs[i];
qAcc += (m_evenA[1][a] + m_evenB[1][b]) * HBFIRFilterTraits<HBFilterOrder>::hbCoeffs[i];
}
else
{
iAcc += (m_oddA[0][a] + m_oddB[0][b]) * HBFIRFilterTraits<HBFilterOrder>::hbCoeffs[i];
qAcc += (m_oddA[1][a] + m_oddB[1][b]) * HBFIRFilterTraits<HBFilterOrder>::hbCoeffs[i];
}
a += 1;
b += 1;
}
#endif
if ((m_ptrB % 2) == 0)
{
iAcc += ((qint32)m_oddB[0][m_ptrB/2 + m_size/2]) << (HBFIRFilterTraits<HBFilterOrder>::hbShift - 1);
qAcc += ((qint32)m_oddB[1][m_ptrB/2 + m_size/2]) << (HBFIRFilterTraits<HBFilterOrder>::hbShift - 1);
}
else
{
iAcc += ((qint32)m_evenB[0][m_ptrB/2 + m_size/2 + 1]) << (HBFIRFilterTraits<HBFilterOrder>::hbShift - 1);
qAcc += ((qint32)m_evenB[1][m_ptrB/2 + m_size/2 + 1]) << (HBFIRFilterTraits<HBFilterOrder>::hbShift - 1);
}
*x = iAcc >> (HBFIRFilterTraits<HBFilterOrder>::hbShift -1); // HB_SHIFT incorrect do not loose the gained bit
*y = qAcc >> (HBFIRFilterTraits<HBFilterOrder>::hbShift -1);
}
};
template<uint32_t HBFilterOrder>
IntHalfbandFilterEO2<HBFilterOrder>::IntHalfbandFilterEO2()
{
m_size = HBFIRFilterTraits<HBFilterOrder>::hbOrder/2;
for (int i = 0; i < 2*m_size; i++)
{
m_evenB[0][i] = 0;
m_evenB[1][i] = 0;
m_oddB[0][i] = 0;
m_oddB[1][i] = 0;
}
m_ptrA = 0;
m_ptrB = 0;
m_state = 0;
}
#endif /* SDRBASE_DSP_INTHALFBANDFILTEREO2_H_ */

8
sdrbase/dsp/upchannelizer.cpp
View File

@ -203,20 +203,20 @@ void UpChannelizer::applyConfiguration()
#ifdef USE_SIMD
UpChannelizer::FilterStage::FilterStage(Mode mode) :
m_filter(new IntHalfbandFilterEO1<UPCHANNELIZER_HB_FILTER_ORDER>),
m_filter(new IntHalfbandFilterEO2<UPCHANNELIZER_HB_FILTER_ORDER>),
m_workFunction(0)
{
switch(mode) {
case ModeCenter:
m_workFunction = &IntHalfbandFilterEO1<UPCHANNELIZER_HB_FILTER_ORDER>::workInterpolateCenter;
m_workFunction = &IntHalfbandFilterEO2<UPCHANNELIZER_HB_FILTER_ORDER>::workInterpolateCenter;
break;
case ModeLowerHalf:
m_workFunction = &IntHalfbandFilterEO1<UPCHANNELIZER_HB_FILTER_ORDER>::workInterpolateLowerHalf;
m_workFunction = &IntHalfbandFilterEO2<UPCHANNELIZER_HB_FILTER_ORDER>::workInterpolateLowerHalf;
break;
case ModeUpperHalf:
m_workFunction = &IntHalfbandFilterEO1<UPCHANNELIZER_HB_FILTER_ORDER>::workInterpolateUpperHalf;
m_workFunction = &IntHalfbandFilterEO2<UPCHANNELIZER_HB_FILTER_ORDER>::workInterpolateUpperHalf;
break;
}
}

6
sdrbase/dsp/upchannelizer.h
View File

@ -24,7 +24,7 @@
#include "util/export.h"
#include "util/message.h"
#ifdef USE_SIMD
#include "dsp/inthalfbandfiltereo1.h"
#include "dsp/inthalfbandfiltereo2.h"
#else
#include "dsp/inthalfbandfilterdb.h"
#endif
@ -74,8 +74,8 @@ protected:
};
#ifdef USE_SIMD
typedef bool (IntHalfbandFilterEO1<UPCHANNELIZER_HB_FILTER_ORDER>::*WorkFunction)(Sample* sIn, Sample *sOut);
IntHalfbandFilterEO1<UPCHANNELIZER_HB_FILTER_ORDER>* m_filter;
typedef bool (IntHalfbandFilterEO2<UPCHANNELIZER_HB_FILTER_ORDER>::*WorkFunction)(Sample* sIn, Sample *sOut);
IntHalfbandFilterEO2<UPCHANNELIZER_HB_FILTER_ORDER>* m_filter;
#else
typedef bool (IntHalfbandFilterDB<UPCHANNELIZER_HB_FILTER_ORDER>::*WorkFunction)(Sample* sIn, Sample *sOut);
IntHalfbandFilterDB<UPCHANNELIZER_HB_FILTER_ORDER>* m_filter;