Update of /cvsroot/audacity/lib-src/soundtouch/source/SoundTouch
In directory sc8-pr-cvs11.sourceforge.net:/tmp/cvs-serv29546/source/SoundTouch
Added Files:
SoundTouch.dsp SoundTouch.sln mmx_optimized.cpp
sse_optimized.cpp
Log Message:
Commiting the new bits of the updated soundtouch library
--- NEW FILE: sse_optimized.cpp ---
////////////////////////////////////////////////////////////////////////////////
///
/// SSE optimized routines for Pentium-III, Athlon-XP and later CPUs. All SSE
/// optimized functions have been gathered into this single source
/// code file, regardless to their class or original source code file, in order
/// to ease porting the library to other compiler and processor platforms.
///
/// The SSE-optimizations are programmed using SSE compiler intrinsics that
/// are supported both by Microsoft Visual C++ and GCC compilers, so this file
/// should compile with both toolsets.
///
/// NOTICE: If using Visual Studio 6.0, you'll need to install the "Visual C++
/// 6.0 processor pack" update to support SSE instruction set. The update is
/// available for download at Microsoft Developers Network, see here:
/// http://msdn.microsoft.com/vstudio/downloads/tools/ppack/default.aspx
///
/// If the above URL is expired or removed, go to "http://msdn.microsoft.com";
and
/// perform a search with keywords "processor pack".
///
/// Author : Copyright (c) Olli Parviainen
/// Author e-mail : oparviai 'at' iki.fi
/// SoundTouch WWW: http://www.surina.net/soundtouch
///
////////////////////////////////////////////////////////////////////////////////
//
// Last changed : $Date: 2006/09/18 07:39:15 $
// File revision : $Revision: 1.2 $
//
// $Id: sse_optimized.cpp,v 1.2 2006/09/18 07:39:15 richardash1981 Exp $
//
////////////////////////////////////////////////////////////////////////////////
//
// License :
//
// SoundTouch audio processing library
// Copyright (c) Olli Parviainen
//
// This library is free software; you can redistribute it and/or
// modify it under the terms of the GNU Lesser General Public
// License as published by the Free Software Foundation; either
// version 2.1 of the License, or (at your option) any later version.
//
// This library 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
// Lesser General Public License for more details.
//
// You should have received a copy of the GNU Lesser General Public
// License along with this library; if not, write to the Free Software
// Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
//
////////////////////////////////////////////////////////////////////////////////
#include "cpu_detect.h"
#include "STTypes.h"
using namespace soundtouch;
#ifdef ALLOW_SSE
// SSE routines available only with float sample type
//////////////////////////////////////////////////////////////////////////////
//
// implementation of SSE optimized functions of class 'TDStretchSSE'
//
//////////////////////////////////////////////////////////////////////////////
#include "TDStretch.h"
#include <xmmintrin.h>
// Calculates cross correlation of two buffers
double TDStretchSSE::calcCrossCorrStereo(const float *pV1, const float *pV2)
const
{
uint i;
__m128 vSum, *pVec2;
// Note. It means a major slow-down if the routine needs to tolerate
// unaligned __m128 memory accesses. It's way faster if we can skip
// unaligned slots and use _mm_load_ps instruction instead of _mm_loadu_ps.
// This can mean up to ~ 10-fold difference (incl. part of which is
// due to skipping every second round for stereo sound though).
//
// Compile-time define ALLOW_NONEXACT_SIMD_OPTIMIZATION is provided
// for choosing if this little cheating is allowed.
#ifdef ALLOW_NONEXACT_SIMD_OPTIMIZATION
// Little cheating allowed, return valid correlation only for
// aligned locations, meaning every second round for stereo sound.
#define _MM_LOAD _mm_load_ps
if (((ulong)pV1) & 15) return -1e50; // skip unaligned locations
#else
// No cheating allowed, use unaligned load & take the resulting
// performance hit.
#define _MM_LOAD _mm_loadu_ps
#endif
// ensure overlapLength is divisible by 8
assert((overlapLength % 8) == 0);
// Calculates the cross-correlation value between 'pV1' and 'pV2' vectors
// Note: pV2 _must_ be aligned to 16-bit boundary, pV1 need not.
pVec2 = (__m128*)pV2;
vSum = _mm_setzero_ps();
// Unroll the loop by factor of 4 * 4 operations
for (i = 0; i < overlapLength / 8; i ++)
{
// vSum += pV1[0..3] * pV2[0..3]
vSum = _mm_add_ps(vSum, _mm_mul_ps(_MM_LOAD(pV1),pVec2[0]));
// vSum += pV1[4..7] * pV2[4..7]
vSum = _mm_add_ps(vSum, _mm_mul_ps(_MM_LOAD(pV1 + 4), pVec2[1]));
// vSum += pV1[8..11] * pV2[8..11]
vSum = _mm_add_ps(vSum, _mm_mul_ps(_MM_LOAD(pV1 + 8), pVec2[2]));
// vSum += pV1[12..15] * pV2[12..15]
vSum = _mm_add_ps(vSum, _mm_mul_ps(_MM_LOAD(pV1 + 12), pVec2[3]));
pV1 += 16;
pVec2 += 4;
}
// return value = vSum[0] + vSum[1] + vSum[2] + vSum[3]
float *pvSum = (float*)&vSum;
return (double)(pvSum[0] + pvSum[1] + pvSum[2] + pvSum[3]);
/* This is approximately corresponding routine in C-language:
double corr;
uint i;
// Calculates the cross-correlation value between 'pV1' and 'pV2' vectors
corr = 0.0;
for (i = 0; i < overlapLength / 8; i ++)
{
corr += pV1[0] * pV2[0] +
pV1[1] * pV2[1] +
pV1[2] * pV2[2] +
pV1[3] * pV2[3] +
pV1[4] * pV2[4] +
pV1[5] * pV2[5] +
pV1[6] * pV2[6] +
pV1[7] * pV2[7] +
pV1[8] * pV2[8] +
pV1[9] * pV2[9] +
pV1[10] * pV2[10] +
pV1[11] * pV2[11] +
pV1[12] * pV2[12] +
pV1[13] * pV2[13] +
pV1[14] * pV2[14] +
pV1[15] * pV2[15];
pV1 += 16;
pV2 += 16;
}
*/
/* This is corresponding routine in assembler. This may be teeny-weeny bit
faster
than intrinsic version, but more difficult to maintain & get compiled on
multiple
platforms.
uint overlapLengthLocal = overlapLength;
float corr;
_asm
{
// Very important note: data in 'pV2' _must_ be aligned to
// 16-byte boundary!
// give prefetch hints to CPU of what data are to be needed soonish
// give more aggressive hints on pV1 as that changes while pV2 stays
// same between runs
prefetcht0 [pV1]
prefetcht0 [pV2]
prefetcht0 [pV1 + 32]
mov eax, dword ptr pV1
mov ebx, dword ptr pV2
xorps xmm0, xmm0
mov ecx, overlapLengthLocal
shr ecx, 3 // div by eight
loop1:
prefetcht0 [eax + 64] // give a prefetch hint to CPU what data are
to be needed soonish
prefetcht0 [ebx + 32] // give a prefetch hint to CPU what data are
to be needed soonish
movups xmm1, [eax]
mulps xmm1, [ebx]
addps xmm0, xmm1
movups xmm2, [eax + 16]
mulps xmm2, [ebx + 16]
addps xmm0, xmm2
prefetcht0 [eax + 96] // give a prefetch hint to CPU what data are
to be needed soonish
prefetcht0 [ebx + 64] // give a prefetch hint to CPU what data are
to be needed soonish
movups xmm3, [eax + 32]
mulps xmm3, [ebx + 32]
addps xmm0, xmm3
movups xmm4, [eax + 48]
mulps xmm4, [ebx + 48]
addps xmm0, xmm4
add eax, 64
add ebx, 64
dec ecx
jnz loop1
// add the four floats of xmm0 together and return the result.
movhlps xmm1, xmm0 // move 3 & 4 of xmm0 to 1 & 2 of xmm1
addps xmm1, xmm0
movaps xmm2, xmm1
shufps xmm2, xmm2, 0x01 // move 2 of xmm2 as 1 of xmm2
addss xmm2, xmm1
movss corr, xmm2
}
return (double)corr;
*/
}
//////////////////////////////////////////////////////////////////////////////
//
// implementation of SSE optimized functions of class 'FIRFilter'
//
//////////////////////////////////////////////////////////////////////////////
#include "FIRFilter.h"
FIRFilterSSE::FIRFilterSSE() : FIRFilter()
{
filterCoeffsUnalign = NULL;
}
FIRFilterSSE::~FIRFilterSSE()
{
delete[] filterCoeffsUnalign;
}
// (overloaded) Calculates filter coefficients for SSE routine
void FIRFilterSSE::setCoefficients(const float *coeffs, uint newLength, uint
uResultDivFactor)
{
uint i;
float fDivider;
FIRFilter::setCoefficients(coeffs, newLength, uResultDivFactor);
// Scale the filter coefficients so that it won't be necessary to scale the
filtering result
// also rearrange coefficients suitably for 3DNow!
// Ensure that filter coeffs array is aligned to 16-byte boundary
delete[] filterCoeffsUnalign;
filterCoeffsUnalign = new float[2 * newLength + 4];
filterCoeffsAlign = (float *)(((unsigned long)filterCoeffsUnalign + 15) &
-16);
fDivider = (float)resultDivider;
// rearrange the filter coefficients for mmx routines
for (i = 0; i < newLength; i ++)
{
filterCoeffsAlign[2 * i + 0] =
filterCoeffsAlign[2 * i + 1] = coeffs[i + 0] / fDivider;
}
}
// SSE-optimized version of the filter routine for stereo sound
uint FIRFilterSSE::evaluateFilterStereo(float *dest, const float *source, uint
numSamples) const
{
int count = (numSamples - length) & -2;
int j;
assert(count % 2 == 0);
if (count < 2) return 0;
assert((length % 8) == 0);
assert(((unsigned long)filterCoeffsAlign) % 16 == 0);
// filter is evaluated for two stereo samples with each iteration, thus use
of 'j += 2'
for (j = 0; j < count; j += 2)
{
const float *pSrc;
const __m128 *pFil;
__m128 sum1, sum2;
uint i;
pSrc = source; // source audio data
pFil = (__m128*)filterCoeffsAlign; // filter coefficients. NOTE:
Assumes coefficients
// are aligned to 16-byte boundary
sum1 = sum2 = _mm_setzero_ps();
for (i = 0; i < length / 8; i ++)
{
// Unroll loop for efficiency & calculate filter for 2*2 stereo
samples
// at each pass
// sum1 is accu for 2*2 filtered stereo sound data at the primary
sound data offset
// sum2 is accu for 2*2 filtered stereo sound data for the next
sound sample offset.
sum1 = _mm_add_ps(sum1, _mm_mul_ps(_mm_loadu_ps(pSrc) ,
pFil[0]));
sum2 = _mm_add_ps(sum2, _mm_mul_ps(_mm_loadu_ps(pSrc + 2),
pFil[0]));
sum1 = _mm_add_ps(sum1, _mm_mul_ps(_mm_loadu_ps(pSrc + 4),
pFil[1]));
sum2 = _mm_add_ps(sum2, _mm_mul_ps(_mm_loadu_ps(pSrc + 6),
pFil[1]));
sum1 = _mm_add_ps(sum1, _mm_mul_ps(_mm_loadu_ps(pSrc + 8) ,
pFil[2]));
sum2 = _mm_add_ps(sum2, _mm_mul_ps(_mm_loadu_ps(pSrc + 10),
pFil[2]));
sum1 = _mm_add_ps(sum1, _mm_mul_ps(_mm_loadu_ps(pSrc + 12),
pFil[3]));
sum2 = _mm_add_ps(sum2, _mm_mul_ps(_mm_loadu_ps(pSrc + 14),
pFil[3]));
pSrc += 16;
pFil += 4;
}
// Now sum1 and sum2 both have a filtered 2-channel sample each, but we
still need
// to sum the two hi- and lo-floats of these registers together.
// post-shuffle & add the filtered values and store to dest.
_mm_storeu_ps(dest, _mm_add_ps(
_mm_shuffle_ps(sum1, sum2, _MM_SHUFFLE(1,0,3,2)), // s2_1
s2_0 s1_3 s1_2
_mm_shuffle_ps(sum1, sum2, _MM_SHUFFLE(3,2,1,0)) // s2_3
s2_2 s1_1 s1_0
));
source += 4;
dest += 4;
}
// Ideas for further improvement:
// 1. If it could be guaranteed that 'source' were always aligned to
16-byte
// boundary, a faster aligned '_mm_load_ps' instruction could be used.
// 2. If it could be guaranteed that 'dest' were always aligned to 16-byte
// boundary, a faster '_mm_store_ps' instruction could be used.
return (uint)count;
/* original routine in C-language. please notice the C-version has
differently
organized coefficients though.
double suml1, suml2;
double sumr1, sumr2;
uint i, j;
for (j = 0; j < count; j += 2)
{
const float *ptr;
const float *pFil;
suml1 = sumr1 = 0.0;
suml2 = sumr2 = 0.0;
ptr = src;
pFil = filterCoeffs;
for (i = 0; i < lengthLocal; i ++)
{
// unroll loop for efficiency.
suml1 += ptr[0] * pFil[0] +
ptr[2] * pFil[2] +
ptr[4] * pFil[4] +
ptr[6] * pFil[6];
sumr1 += ptr[1] * pFil[1] +
ptr[3] * pFil[3] +
ptr[5] * pFil[5] +
ptr[7] * pFil[7];
suml2 += ptr[8] * pFil[0] +
ptr[10] * pFil[2] +
ptr[12] * pFil[4] +
ptr[14] * pFil[6];
sumr2 += ptr[9] * pFil[1] +
ptr[11] * pFil[3] +
ptr[13] * pFil[5] +
ptr[15] * pFil[7];
ptr += 16;
pFil += 8;
}
dest[0] = (float)suml1;
dest[1] = (float)sumr1;
dest[2] = (float)suml2;
dest[3] = (float)sumr2;
src += 4;
dest += 4;
}
*/
/* Similar routine in assembly, again obsoleted due to maintainability
_asm
{
// Very important note: data in 'src' _must_ be aligned to
// 16-byte boundary!
mov edx, count
mov ebx, dword ptr src
mov eax, dword ptr dest
shr edx, 1
loop1:
// "outer loop" : during each round 2*2 output samples are calculated
// give prefetch hints to CPU of what data are to be needed soonish
prefetcht0 [ebx]
prefetcht0 [filterCoeffsLocal]
mov esi, ebx
mov edi, filterCoeffsLocal
xorps xmm0, xmm0
xorps xmm1, xmm1
mov ecx, lengthLocal
loop2:
// "inner loop" : during each round eight FIR filter taps are evaluated
for 2*2 samples
prefetcht0 [esi + 32] // give a prefetch hint to CPU what data are
to be needed soonish
prefetcht0 [edi + 32] // give a prefetch hint to CPU what data are
to be needed soonish
movups xmm2, [esi] // possibly unaligned load
movups xmm3, [esi + 8] // possibly unaligned load
mulps xmm2, [edi]
mulps xmm3, [edi]
addps xmm0, xmm2
addps xmm1, xmm3
movups xmm4, [esi + 16] // possibly unaligned load
movups xmm5, [esi + 24] // possibly unaligned load
mulps xmm4, [edi + 16]
mulps xmm5, [edi + 16]
addps xmm0, xmm4
addps xmm1, xmm5
prefetcht0 [esi + 64] // give a prefetch hint to CPU what data are
to be needed soonish
prefetcht0 [edi + 64] // give a prefetch hint to CPU what data are
to be needed soonish
movups xmm6, [esi + 32] // possibly unaligned load
movups xmm7, [esi + 40] // possibly unaligned load
mulps xmm6, [edi + 32]
mulps xmm7, [edi + 32]
addps xmm0, xmm6
addps xmm1, xmm7
movups xmm4, [esi + 48] // possibly unaligned load
movups xmm5, [esi + 56] // possibly unaligned load
mulps xmm4, [edi + 48]
mulps xmm5, [edi + 48]
addps xmm0, xmm4
addps xmm1, xmm5
add esi, 64
add edi, 64
dec ecx
jnz loop2
// Now xmm0 and xmm1 both have a filtered 2-channel sample each, but we
still need
// to sum the two hi- and lo-floats of these registers together.
movhlps xmm2, xmm0 // xmm2 = xmm2_3 xmm2_2 xmm0_3 xmm0_2
movlhps xmm2, xmm1 // xmm2 = xmm1_1 xmm1_0 xmm0_3 xmm0_2
shufps xmm0, xmm1, 0xe4 // xmm0 = xmm1_3 xmm1_2 xmm0_1 xmm0_0
addps xmm0, xmm2
movaps [eax], xmm0
add ebx, 16
add eax, 16
dec edx
jnz loop1
}
*/
}
#endif // ALLOW_SSE
--- NEW FILE: SoundTouch.dsp ---
# Microsoft Developer Studio Project File - Name="SoundTouch" - Package
Owner=<4>
# Microsoft Developer Studio Generated Build File, Format Version 6.00
# ** DO NOT EDIT **
# TARGTYPE "Win32 (x86) Static Library" 0x0104
CFG=SoundTouch - Win32 Debug
!MESSAGE This is not a valid makefile. To build this project using NMAKE,
!MESSAGE use the Export Makefile command and run
!MESSAGE
!MESSAGE NMAKE /f "SoundTouch.mak".
!MESSAGE
!MESSAGE You can specify a configuration when running NMAKE
!MESSAGE by defining the macro CFG on the command line. For example:
!MESSAGE
!MESSAGE NMAKE /f "SoundTouch.mak" CFG="SoundTouch - Win32 Debug"
!MESSAGE
!MESSAGE Possible choices for configuration are:
!MESSAGE
!MESSAGE "SoundTouch - Win32 Release" (based on "Win32 (x86) Static Library")
!MESSAGE "SoundTouch - Win32 Debug" (based on "Win32 (x86) Static Library")
!MESSAGE
# Begin Project
# PROP AllowPerConfigDependencies 0
# PROP Scc_ProjName ""
# PROP Scc_LocalPath ""
CPP=cl.exe
RSC=rc.exe
!IF "$(CFG)" == "SoundTouch - Win32 Release"
# PROP BASE Use_MFC 0
# PROP BASE Use_Debug_Libraries 0
# PROP BASE Output_Dir "Release"
# PROP BASE Intermediate_Dir "Release"
# PROP BASE Target_Dir ""
# PROP Use_MFC 0
# PROP Use_Debug_Libraries 0
# PROP Output_Dir "Release"
# PROP Intermediate_Dir "Release"
# PROP Target_Dir ""
# ADD BASE CPP /nologo /W3 /GX /O2 /D "WIN32" /D "NDEBUG" /D "_MBCS" /D "_LIB"
/YX /FD /c
# ADD CPP /nologo /W3 /GX /Zi /O2 /I "..\..\include" /D "WIN32" /D "NDEBUG" /D
"_MBCS" /D "_LIB" /YX /FD /c
# ADD BASE RSC /l 0x40b /d "NDEBUG"
# ADD RSC /l 0x40b /d "NDEBUG"
BSC32=bscmake.exe
# ADD BASE BSC32 /nologo
# ADD BSC32 /nologo
LIB32=link.exe -lib
# ADD BASE LIB32 /nologo
# ADD LIB32 /nologo
# Begin Special Build Tool
SOURCE="$(InputPath)"
PostBuild_Cmds=copy .\Release\SoundTouch.lib ..\..\lib\
# End Special Build Tool
!ELSEIF "$(CFG)" == "SoundTouch - Win32 Debug"
# PROP BASE Use_MFC 0
# PROP BASE Use_Debug_Libraries 1
# PROP BASE Output_Dir "Debug"
# PROP BASE Intermediate_Dir "Debug"
# PROP BASE Target_Dir ""
# PROP Use_MFC 0
# PROP Use_Debug_Libraries 1
# PROP Output_Dir "Debug"
# PROP Intermediate_Dir "Debug"
# PROP Target_Dir ""
# ADD BASE CPP /nologo /W3 /Gm /GX /ZI /Od /D "WIN32" /D "_DEBUG" /D "_MBCS" /D
"_LIB" /YX /FD /GZ /c
# ADD CPP /nologo /W3 /Gm /GX /ZI /Od /I "..\..\include" /D "WIN32" /D "_DEBUG"
/D "_MBCS" /D "_LIB" /FR /YX /FD /GZ /c
# ADD BASE RSC /l 0x40b /d "_DEBUG"
# ADD RSC /l 0x40b /d "_DEBUG"
BSC32=bscmake.exe
# ADD BASE BSC32 /nologo
# ADD BSC32 /nologo
LIB32=link.exe -lib
# ADD BASE LIB32 /nologo
# ADD LIB32 /nologo /out:"Debug\SoundTouchD.lib"
# Begin Special Build Tool
SOURCE="$(InputPath)"
PostBuild_Cmds=copy .\Debug\SoundTouchD.lib ..\..\lib\
# End Special Build Tool
!ENDIF
# Begin Target
# Name "SoundTouch - Win32 Release"
# Name "SoundTouch - Win32 Debug"
# Begin Group "Source Files"
# PROP Default_Filter "cpp;c;cxx;rc;def;r;odl;idl;hpj;bat"
# Begin Source File
SOURCE=.\3dnow_win.cpp
# End Source File
# Begin Source File
SOURCE=.\AAFilter.cpp
# End Source File
# Begin Source File
SOURCE=.\cpu_detect_x86_win.cpp
# End Source File
# Begin Source File
SOURCE=.\FIFOSampleBuffer.cpp
# End Source File
# Begin Source File
SOURCE=.\FIRFilter.cpp
# End Source File
# Begin Source File
SOURCE=.\mmx_optimized.cpp
# End Source File
# Begin Source File
SOURCE=.\RateTransposer.cpp
# End Source File
# Begin Source File
SOURCE=.\SoundTouch.cpp
# End Source File
# Begin Source File
SOURCE=.\sse_optimized.cpp
# End Source File
# Begin Source File
SOURCE=.\TDStretch.cpp
# End Source File
# End Group
# Begin Group "Header Files"
# PROP Default_Filter "h;hpp;hxx;hm;inl"
# Begin Source File
SOURCE=.\AAFilter.h
# End Source File
# Begin Source File
SOURCE=.\cpu_detect.h
# End Source File
# Begin Source File
SOURCE=..\..\include\FIFOSampleBuffer.h
# End Source File
# Begin Source File
SOURCE=..\..\include\FIFOSamplePipe.h
# End Source File
# Begin Source File
SOURCE=.\FIRFilter.h
# End Source File
# Begin Source File
SOURCE=.\RateTransposer.h
# End Source File
# Begin Source File
SOURCE=..\..\include\SoundTouch.h
# End Source File
# Begin Source File
SOURCE=..\..\include\STTypes.h
# End Source File
# Begin Source File
SOURCE=.\TDStretch.h
# End Source File
# End Group
# End Target
# End Project
--- NEW FILE: mmx_optimized.cpp ---
////////////////////////////////////////////////////////////////////////////////
///
/// MMX optimized routines. All MMX optimized functions have been gathered into
/// this single source code file, regardless to their class or original source
/// code file, in order to ease porting the library to other compiler and
/// processor platforms.
///
/// The MMX-optimizations are programmed using MMX compiler intrinsics that
/// are supported both by Microsoft Visual C++ and GCC compilers, so this file
/// should compile with both toolsets.
///
/// NOTICE: If using Visual Studio 6.0, you'll need to install the "Visual C++
/// 6.0 processor pack" update to support compiler intrinsic syntax. The update
/// is available for download at Microsoft Developers Network, see here:
/// http://msdn.microsoft.com/vstudio/downloads/tools/ppack/default.aspx
///
/// Author : Copyright (c) Olli Parviainen
/// Author e-mail : oparviai 'at' iki.fi
/// SoundTouch WWW: http://www.surina.net/soundtouch
///
////////////////////////////////////////////////////////////////////////////////
//
// Last changed : $Date: 2006/09/18 07:39:15 $
// File revision : $Revision: 1.2 $
//
// $Id: mmx_optimized.cpp,v 1.2 2006/09/18 07:39:15 richardash1981 Exp $
//
////////////////////////////////////////////////////////////////////////////////
//
// License :
//
// SoundTouch audio processing library
// Copyright (c) Olli Parviainen
//
// This library is free software; you can redistribute it and/or
// modify it under the terms of the GNU Lesser General Public
// License as published by the Free Software Foundation; either
// version 2.1 of the License, or (at your option) any later version.
//
// This library 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
// Lesser General Public License for more details.
//
// You should have received a copy of the GNU Lesser General Public
// License along with this library; if not, write to the Free Software
// Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
//
////////////////////////////////////////////////////////////////////////////////
#include "STTypes.h"
#ifdef ALLOW_MMX
// MMX routines available only with integer sample type
#if !(WIN32 || __i386__ || __x86_64__)
#error "wrong platform - this source code file is exclusively for x86 platforms"
#endif
using namespace soundtouch;
//////////////////////////////////////////////////////////////////////////////
//
// implementation of MMX optimized functions of class 'TDStretchMMX'
//
//////////////////////////////////////////////////////////////////////////////
#include "TDStretch.h"
#include <mmintrin.h>
#include <limits.h>
// Calculates cross correlation of two buffers
long TDStretchMMX::calcCrossCorrStereo(const short *pV1, const short *pV2) const
{
const __m64 *pVec1, *pVec2;
__m64 shifter;
__m64 accu;
long corr;
uint i;
pVec1 = (__m64*)pV1;
pVec2 = (__m64*)pV2;
shifter = _m_from_int(overlapDividerBits);
accu = _mm_setzero_si64();
// Process 4 parallel sets of 2 * stereo samples each during each
// round to improve CPU-level parallellization.
for (i = 0; i < overlapLength / 8; i ++)
{
__m64 temp;
// dictionary of instructions:
// _m_pmaddwd : 4*16bit multiply-add, resulting two 32bits =
[a0*b0+a1*b1 ; a2*b2+a3*b3]
// _mm_add_pi32 : 2*32bit add
// _m_psrad : 32bit right-shift
temp = _mm_add_pi32(_mm_madd_pi16(pVec1[0], pVec2[0]),
_mm_madd_pi16(pVec1[1], pVec2[1]));
accu = _mm_add_pi32(accu, _mm_sra_pi32(temp, shifter));
temp = _mm_add_pi32(_mm_madd_pi16(pVec1[2], pVec2[2]),
_mm_madd_pi16(pVec1[3], pVec2[3]));
accu = _mm_add_pi32(accu, _mm_sra_pi32(temp, shifter));
pVec1 += 4;
pVec2 += 4;
}
// copy hi-dword of mm0 to lo-dword of mm1, then sum mmo+mm1
// and finally store the result into the variable "corr"
accu = _mm_add_pi32(accu, _mm_srli_si64(accu, 32));
corr = _m_to_int(accu);
// Clear MMS state
_m_empty();
return corr;
// Note: Warning about the missing EMMS instruction is harmless
// as it'll be called elsewhere.
}
void TDStretchMMX::clearCrossCorrState()
{
// Clear MMS state
_m_empty();
//_asm EMMS;
}
// MMX-optimized version of the function overlapStereo
void TDStretchMMX::overlapStereo(short *output, const short *input) const
{
const __m64 *pVinput, *pVMidBuf;
__m64 *pVdest;
__m64 mix1, mix2, adder, shifter;
uint i;
pVinput = (const __m64*)input;
pVMidBuf = (const __m64*)pMidBuffer;
pVdest = (__m64*)output;
// mix1 = mixer values for 1st stereo sample
// mix1 = mixer values for 2nd stereo sample
// adder = adder for updating mixer values after each round
mix1 = _mm_set_pi16(0, overlapLength, 0, overlapLength);
adder = _mm_set_pi16(1, -1, 1, -1);
mix2 = _mm_add_pi16(mix1, adder);
adder = _mm_add_pi16(adder, adder);
shifter = _m_from_int(overlapDividerBits);
for (i = 0; i < overlapLength / 4; i ++)
{
__m64 temp1, temp2;
// load & shuffle data so that input & mixbuffer data samples are paired
temp1 = _mm_unpacklo_pi16(pVMidBuf[0], pVinput[0]); // = i0l m0l
i0r m0r
temp2 = _mm_unpackhi_pi16(pVMidBuf[0], pVinput[0]); // = i1l m1l
i1r m1r
// temp = (temp .* mix) >> shifter
temp1 = _mm_sra_pi32(_mm_madd_pi16(temp1, mix1), shifter);
temp2 = _mm_sra_pi32(_mm_madd_pi16(temp2, mix2), shifter);
pVdest[0] = _mm_packs_pi32(temp1, temp2); // pack 2*2*32bit => 4*16bit
// update mix += adder
mix1 = _mm_add_pi16(mix1, adder);
mix2 = _mm_add_pi16(mix2, adder);
// --- second round begins here ---
// load & shuffle data so that input & mixbuffer data samples are paired
temp1 = _mm_unpacklo_pi16(pVMidBuf[1], pVinput[1]); // = i2l m2l
i2r m2r
temp2 = _mm_unpackhi_pi16(pVMidBuf[1], pVinput[1]); // = i3l m3l
i3r m3r
// temp = (temp .* mix) >> shifter
temp1 = _mm_sra_pi32(_mm_madd_pi16(temp1, mix1), shifter);
temp2 = _mm_sra_pi32(_mm_madd_pi16(temp2, mix2), shifter);
pVdest[1] = _mm_packs_pi32(temp1, temp2); // pack 2*2*32bit => 4*16bit
// update mix += adder
mix1 = _mm_add_pi16(mix1, adder);
mix2 = _mm_add_pi16(mix2, adder);
pVinput += 2;
pVMidBuf += 2;
pVdest += 2;
}
_m_empty(); // clear MMS state
}
//////////////////////////////////////////////////////////////////////////////
//
// implementation of MMX optimized functions of class 'FIRFilter'
//
//////////////////////////////////////////////////////////////////////////////
#include "FIRFilter.h"
FIRFilterMMX::FIRFilterMMX() : FIRFilter()
{
filterCoeffsUnalign = NULL;
}
FIRFilterMMX::~FIRFilterMMX()
{
delete[] filterCoeffsUnalign;
}
// (overloaded) Calculates filter coefficients for MMX routine
void FIRFilterMMX::setCoefficients(const short *coeffs, uint newLength, uint
uResultDivFactor)
{
uint i;
FIRFilter::setCoefficients(coeffs, newLength, uResultDivFactor);
// Ensure that filter coeffs array is aligned to 16-byte boundary
delete[] filterCoeffsUnalign;
filterCoeffsUnalign = new short[2 * newLength + 8];
filterCoeffsAlign = (short *)(((uint)filterCoeffsUnalign + 15) & -16);
// rearrange the filter coefficients for mmx routines
for (i = 0;i < length; i += 4)
{
filterCoeffsAlign[2 * i + 0] = coeffs[i + 0];
filterCoeffsAlign[2 * i + 1] = coeffs[i + 2];
filterCoeffsAlign[2 * i + 2] = coeffs[i + 0];
filterCoeffsAlign[2 * i + 3] = coeffs[i + 2];
filterCoeffsAlign[2 * i + 4] = coeffs[i + 1];
filterCoeffsAlign[2 * i + 5] = coeffs[i + 3];
filterCoeffsAlign[2 * i + 6] = coeffs[i + 1];
filterCoeffsAlign[2 * i + 7] = coeffs[i + 3];
}
}
// mmx-optimized version of the filter routine for stereo sound
uint FIRFilterMMX::evaluateFilterStereo(short *dest, const short *src, const
uint numSamples) const
{
// Create stack copies of the needed member variables for asm routines :
uint i, j;
__m64 *pVdest = (__m64*)dest;
if (length < 2) return 0;
for (i = 0; i < numSamples / 2; i ++)
{
__m64 accu1;
__m64 accu2;
const __m64 *pVsrc = (const __m64*)src;
const __m64 *pVfilter = (const __m64*)filterCoeffsAlign;
accu1 = accu2 = _mm_setzero_si64();
for (j = 0; j < lengthDiv8 * 2; j ++)
{
__m64 temp1, temp2;
temp1 = _mm_unpacklo_pi16(pVsrc[0], pVsrc[1]); // = l2 l0 r2 r0
temp2 = _mm_unpackhi_pi16(pVsrc[0], pVsrc[1]); // = l3 l1 r3 r1
accu1 = _mm_add_pi32(accu1, _mm_madd_pi16(temp1, pVfilter[0])); //
+= l2*f2+l0*f0 r2*f2+r0*f0
accu1 = _mm_add_pi32(accu1, _mm_madd_pi16(temp2, pVfilter[1])); //
+= l3*f3+l1*f1 r3*f3+r1*f1
temp1 = _mm_unpacklo_pi16(pVsrc[1], pVsrc[2]); // = l4 l2 r4 r2
accu2 = _mm_add_pi32(accu2, _mm_madd_pi16(temp2, pVfilter[0])); //
+= l3*f2+l1*f0 r3*f2+r1*f0
accu2 = _mm_add_pi32(accu2, _mm_madd_pi16(temp1, pVfilter[1])); //
+= l4*f3+l2*f1 r4*f3+r2*f1
// accu1 += l2*f2+l0*f0 r2*f2+r0*f0
// += l3*f3+l1*f1 r3*f3+r1*f1
// accu2 += l3*f2+l1*f0 r3*f2+r1*f0
// l4*f3+l2*f1 r4*f3+r2*f1
pVfilter += 2;
pVsrc += 2;
}
// accu >>= resultDivFactor
accu1 = _mm_srai_pi32(accu1, resultDivFactor);
accu2 = _mm_srai_pi32(accu2, resultDivFactor);
// pack 2*2*32bits => 4*16 bits
pVdest[0] = _mm_packs_pi32(accu1, accu2);
src += 4;
pVdest ++;
}
_m_empty(); // clear emms state
return (numSamples & 0xfffffffe) - length;
}
#endif // ALLOW_MMX
--- NEW FILE: SoundTouch.sln ---
Microsoft Visual Studio Solution File, Format Version 8.00
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "SoundTouch",
"SoundTouch.vcproj", "{0A626C77-0515-4131-AA80-E0BFFC479FEB}"
ProjectSection(ProjectDependencies) = postProject
EndProjectSection
EndProject
Global
GlobalSection(SolutionConfiguration) = preSolution
Debug = Debug
Release = Release
EndGlobalSection
GlobalSection(ProjectConfiguration) = postSolution
{0A626C77-0515-4131-AA80-E0BFFC479FEB}.Debug.ActiveCfg =
Debug|Win32
{0A626C77-0515-4131-AA80-E0BFFC479FEB}.Debug.Build.0 =
Debug|Win32
{0A626C77-0515-4131-AA80-E0BFFC479FEB}.Release.ActiveCfg =
Release|Win32
{0A626C77-0515-4131-AA80-E0BFFC479FEB}.Release.Build.0 =
Release|Win32
EndGlobalSection
GlobalSection(ExtensibilityGlobals) = postSolution
EndGlobalSection
GlobalSection(ExtensibilityAddIns) = postSolution
EndGlobalSection
EndGlobal
-------------------------------------------------------------------------
Using Tomcat but need to do more? Need to support web services, security?
Get stuff done quickly with pre-integrated technology to make your job easier
Download IBM WebSphere Application Server v.1.0.1 based on Apache Geronimo
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=120709&bid=263057&dat=121642
_______________________________________________
Audacity-cvs mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/audacity-cvs
