IntHalfbandFilterST: implemented intrinsics class

This commit is contained in:
f4exb 2016-11-08 17:29:46 +01:00
parent 995fd04543
commit 2504d199fd
4 changed files with 124 additions and 10 deletions

View File

@ -214,6 +214,7 @@ set(sdrbase_HEADERS
sdrbase/dsp/inthalfbandfiltereo2.h
sdrbase/dsp/inthalfbandfiltereo2i.h
sdrbase/dsp/inthalfbandfilterst.h
sdrbase/dsp/inthalfbandfiltersti.h
sdrbase/dsp/kissfft.h
sdrbase/dsp/kissengine.h
sdrbase/dsp/lowpass.h

View File

@ -3,7 +3,7 @@
// written by Edouard Griffiths //
// //
// Integer half-band FIR based interpolator and decimator //
// This is the double buffer variant //
// This is the even/odd and I/Q stride with double buffering variant //
// //
// 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 //
@ -24,6 +24,7 @@
#include <stdint.h>
#include "dsp/dsptypes.h"
#include "dsp/hbfiltertraits.h"
#include "dsp/inthalfbandfiltersti.h"
#include "util/export.h"
template<uint32_t HBFilterOrder>
@ -403,7 +404,8 @@ public:
}
protected:
qint32 m_samplesDB[2*HBFIRFilterTraits<HBFilterOrder>::hbOrder][2]; // double buffer technique with even/odd amnd I/Q stride
qint32 m_samplesDB[2*HBFilterOrder][2]; // double buffer technique with even/odd amnd I/Q stride
qint32 m_samplesAligned[HBFilterOrder][2] __attribute__ ((aligned (16)));
int m_ptr;
int m_size;
int m_state;
@ -431,9 +433,6 @@ protected:
void doFIR(Sample* sample)
{
int a = m_ptr + m_size; // tip pointer - odd
int b = m_ptr + 1; // tail pointer - aven
// calculate on odd values
if ((m_ptr % 2) == 1)
@ -442,6 +441,17 @@ protected:
m_qEvenAcc = 0;
m_iOddAcc = 0;
m_qOddAcc = 0;
#ifdef USE_SSE4_1
memcpy((void *) m_samplesAligned, (const void *) &(m_samplesDB[ m_ptr + 1][0]), HBFilterOrder*2*sizeof(qint32));
IntHalfbandFilterSTIntrinsics<HBFilterOrder>::work(
m_samplesAligned,
m_iEvenAcc,
m_qEvenAcc,
m_iOddAcc,
m_qOddAcc);
#else
int a = m_ptr + m_size; // tip pointer - odd
int b = m_ptr + 1; // tail pointer - aven
for (int i = 0; i < HBFIRFilterTraits<HBFilterOrder>::hbOrder / 4; i++)
{
@ -452,7 +462,7 @@ protected:
a -= 2;
b += 2;
}
#endif
m_iEvenAcc += ((qint32)m_samplesDB[m_ptr + m_size/2][0]) << (HBFIRFilterTraits<HBFilterOrder>::hbShift - 1);
m_qEvenAcc += ((qint32)m_samplesDB[m_ptr + m_size/2][1]) << (HBFIRFilterTraits<HBFilterOrder>::hbShift - 1);
m_iOddAcc += ((qint32)m_samplesDB[m_ptr + m_size/2 + 1][0]) << (HBFIRFilterTraits<HBFilterOrder>::hbShift - 1);
@ -470,9 +480,6 @@ protected:
void doFIR(qint32 *x, qint32 *y)
{
int a = m_ptr + m_size; // tip pointer - odd
int b = m_ptr + 1; // tail pointer - aven
// calculate on odd values
if ((m_ptr % 2) == 1)
@ -482,6 +489,18 @@ protected:
m_iOddAcc = 0;
m_qOddAcc = 0;
#ifdef USE_SSE4_1
memcpy((void *) m_samplesAligned, (const void *) &(m_samplesDB[ m_ptr + 1][0]), HBFilterOrder*2*sizeof(qint32));
IntHalfbandFilterSTIntrinsics<HBFilterOrder>::work(
m_samplesAligned,
m_iEvenAcc,
m_qEvenAcc,
m_iOddAcc,
m_qOddAcc);
#else
int a = m_ptr + m_size; // tip pointer - odd
int b = m_ptr + 1; // tail pointer - aven
for (int i = 0; i < HBFIRFilterTraits<HBFilterOrder>::hbOrder / 4; i++)
{
m_iEvenAcc += (m_samplesDB[a-1][0] + m_samplesDB[b][0]) * HBFIRFilterTraits<HBFilterOrder>::hbCoeffs[i];
@ -491,7 +510,7 @@ protected:
a -= 2;
b += 2;
}
#endif
m_iEvenAcc += ((qint32)m_samplesDB[m_ptr + m_size/2][0]) << (HBFIRFilterTraits<HBFilterOrder>::hbShift - 1);
m_qEvenAcc += ((qint32)m_samplesDB[m_ptr + m_size/2][1]) << (HBFIRFilterTraits<HBFilterOrder>::hbShift - 1);
m_iOddAcc += ((qint32)m_samplesDB[m_ptr + m_size/2 + 1][0]) << (HBFIRFilterTraits<HBFilterOrder>::hbShift - 1);

View File

@ -0,0 +1,92 @@
///////////////////////////////////////////////////////////////////////////////////
// Copyright (C) 2016 F4EXB //
// written by Edouard Griffiths //
// //
// Integer half-band FIR based interpolator and decimator //
// This is the even/odd and I/Q stride with double buffering variant //
// This is the SIMD intrinsics code //
// //
// 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_INTHALFBANDFILTERSTI_H_
#define SDRBASE_DSP_INTHALFBANDFILTERSTI_H_
#include <stdint.h>
#if defined(USE_SSE4_1)
#include <smmintrin.h>
#endif
#include "hbfiltertraits.h"
template<uint32_t HBFilterOrder>
class IntHalfbandFilterSTIntrinsics
{
public:
static void work(
int32_t samples[HBFilterOrder][2],
int32_t& iEvenAcc, int32_t& qEvenAcc,
int32_t& iOddAcc, int32_t& qOddAcc)
{
#if defined(USE_SSE4_1)
int a = HBFIRFilterTraits<HBFilterOrder>::hbOrder - 2; // tip
int b = 0; // tail
const __m128i* h = (const __m128i*) HBFIRFilterTraits<HBFilterOrder>::hbCoeffs;
__m128i sum = _mm_setzero_si128();
__m128i sh, shh, sa, sb;
int32_t sums[4] __attribute__ ((aligned (16)));
for (int i = 0; i < HBFIRFilterTraits<HBFilterOrder>::hbOrder / 16; i++)
{
sh = _mm_load_si128(h);
shh = _mm_shuffle_epi32(sh, _MM_SHUFFLE(0,0,0,0));
sa = _mm_load_si128((__m128i*) &(samples[a][0])); // Ei,Eq,Oi,Oq
sb = _mm_load_si128((__m128i*) &(samples[b][0]));
sum = _mm_add_epi32(sum, _mm_mullo_epi32(_mm_add_epi32(sa, sb), shh));
a -= 2;
b += 2;
shh = _mm_shuffle_epi32(sh, _MM_SHUFFLE(1,1,1,1));
sa = _mm_load_si128((__m128i*) &(samples[a][0])); // Ei,Eq,Oi,Oq
sb = _mm_load_si128((__m128i*) &(samples[b][0]));
sum = _mm_add_epi32(sum, _mm_mullo_epi32(_mm_add_epi32(sa, sb), shh));
a -= 2;
b += 2;
shh = _mm_shuffle_epi32(sh, _MM_SHUFFLE(2,2,2,2));
sa = _mm_load_si128((__m128i*) &(samples[a][0])); // Ei,Eq,Oi,Oq
sb = _mm_load_si128((__m128i*) &(samples[b][0]));
sum = _mm_add_epi32(sum, _mm_mullo_epi32(_mm_add_epi32(sa, sb), shh));
a -= 2;
b += 2;
shh = _mm_shuffle_epi32(sh, _MM_SHUFFLE(3,3,3,3));
sa = _mm_load_si128((__m128i*) &(samples[a][0])); // Ei,Eq,Oi,Oq
sb = _mm_load_si128((__m128i*) &(samples[b][0]));
sum = _mm_add_epi32(sum, _mm_mullo_epi32(_mm_add_epi32(sa, sb), shh));
a -= 2;
b += 2;
++h;
}
// Extract values from sum vector
_mm_store_si128((__m128i*) sums, sum);
iEvenAcc = sums[0];
qEvenAcc = sums[1];
iOddAcc = sums[2];
qOddAcc = sums[3];
#endif
}
};
#endif /* SDRBASE_DSP_INTHALFBANDFILTERSTI_H_ */

View File

@ -147,6 +147,8 @@ HEADERS += mainwindow.h\
dsp/inthalfbandfiltereo1i.h\
dsp/inthalfbandfiltereo2.h\
dsp/inthalfbandfiltereo2i.h\
dsp/inthalfbandfilterst.h\
dsp/inthalfbandfiltersti.h\
dsp/kissfft.h\
dsp/kissengine.h\
dsp/lowpass.h\