Autres/audacity
1
0
Fork 0
mirror of https://github.com/cookiengineer/audacity synced 2025-10-22 22:43:01 +02:00
Code Issues Releases Wiki Activity

Update soundtouch to 1.7.1.

Browse Source
This commit is contained in:
lllucius
2013-10-24 06:36:42 +00:00
parent 8d97adecb1
commit f740766f37
107 changed files with 6813 additions and 60819 deletions
Download Patch File Download Diff File Expand all files Collapse all files

289
lib-src/soundtouch/source/SoundTouch/sse_optimized.cpp
View File

@@ -1,20 +1,20 @@
////////////////////////////////////////////////////////////////////////////////
///
/// 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
/// 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
/// 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
/// http://msdn.microsoft.com/en-us/vstudio/aa718349.aspx
///
/// If the above URL is expired or removed, go to "http://msdn.microsoft.com" and
/// 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
@@ -23,10 +23,10 @@
///
////////////////////////////////////////////////////////////////////////////////
//
// Last changed : $Date: 2006-09-18 22:29:23 $
// File revision : $Revision: 1.3 $
// Last changed : $Date: 2012-11-08 20:53:01 +0200 (Thu, 08 Nov 2012) $
// File revision : $Revision: 4 $
//
// $Id: sse_optimized.cpp,v 1.3 2006-09-18 22:29:23 martynshaw Exp $
// $Id: sse_optimized.cpp 160 2012-11-08 18:53:01Z oparviai $
//
////////////////////////////////////////////////////////////////////////////////
//
@@ -56,9 +56,9 @@
using namespace soundtouch;
#ifdef ALLOW_SSE
#ifdef SOUNDTOUCH_ALLOW_SSE
// SSE routines available only with float sample type
// SSE routines available only with float sample type
//////////////////////////////////////////////////////////////////////////////
//
@@ -68,74 +68,92 @@ using namespace soundtouch;
#include "TDStretch.h"
#include <xmmintrin.h>
#include <math.h>
// Calculates cross correlation of two buffers
double TDStretchSSE::calcCrossCorrStereo(const float *pV1, const float *pV2) const
double TDStretchSSE::calcCrossCorr(const float *pV1, const float *pV2) const
{
uint i;
__m128 vSum, *pVec2;
int i;
const float *pVec1;
const __m128 *pVec2;
__m128 vSum, vNorm;
// 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
// 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
// Compile-time define SOUNDTOUCH_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
#ifdef SOUNDTOUCH_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
if (((ulongptr)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
#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();
pVec1 = (const float*)pV1;
pVec2 = (const __m128*)pV2;
vSum = vNorm = _mm_setzero_ps();
// Unroll the loop by factor of 4 * 4 operations
for (i = 0; i < overlapLength / 8; i ++)
// Unroll the loop by factor of 4 * 4 operations. Use same routine for
// stereo & mono, for mono it just means twice the amount of unrolling.
for (i = 0; i < channels * overlapLength / 16; i ++)
{
__m128 vTemp;
// vSum += pV1[0..3] * pV2[0..3]
vSum = _mm_add_ps(vSum, _mm_mul_ps(_MM_LOAD(pV1),pVec2[0]));
vTemp = _MM_LOAD(pVec1);
vSum = _mm_add_ps(vSum, _mm_mul_ps(vTemp ,pVec2[0]));
vNorm = _mm_add_ps(vNorm, _mm_mul_ps(vTemp ,vTemp));
// vSum += pV1[4..7] * pV2[4..7]
vSum = _mm_add_ps(vSum, _mm_mul_ps(_MM_LOAD(pV1 + 4), pVec2[1]));
vTemp = _MM_LOAD(pVec1 + 4);
vSum = _mm_add_ps(vSum, _mm_mul_ps(vTemp, pVec2[1]));
vNorm = _mm_add_ps(vNorm, _mm_mul_ps(vTemp ,vTemp));
// vSum += pV1[8..11] * pV2[8..11]
vSum = _mm_add_ps(vSum, _mm_mul_ps(_MM_LOAD(pV1 + 8), pVec2[2]));
vTemp = _MM_LOAD(pVec1 + 8);
vSum = _mm_add_ps(vSum, _mm_mul_ps(vTemp, pVec2[2]));
vNorm = _mm_add_ps(vNorm, _mm_mul_ps(vTemp ,vTemp));
// vSum += pV1[12..15] * pV2[12..15]
vSum = _mm_add_ps(vSum, _mm_mul_ps(_MM_LOAD(pV1 + 12), pVec2[3]));
vTemp = _MM_LOAD(pVec1 + 12);
vSum = _mm_add_ps(vSum, _mm_mul_ps(vTemp, pVec2[3]));
vNorm = _mm_add_ps(vNorm, _mm_mul_ps(vTemp ,vTemp));
pV1 += 16;
pVec1 += 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]);
float *pvNorm = (float*)&vNorm;
double norm = sqrt(pvNorm[0] + pvNorm[1] + pvNorm[2] + pvNorm[3]);
if (norm < 1e-9) norm = 1.0; // to avoid div by zero
/* This is approximately corresponding routine in C-language:
double corr;
float *pvSum = (float*)&vSum;
return (double)(pvSum[0] + pvSum[1] + pvSum[2] + pvSum[3]) / norm;
/* This is approximately corresponding routine in C-language yet without normalization:
double corr, norm;
uint i;
// Calculates the cross-correlation value between 'pV1' and 'pV2' vectors
corr = 0.0;
for (i = 0; i < overlapLength / 8; i ++)
corr = norm = 0.0;
for (i = 0; i < channels * overlapLength / 16; i ++)
{
corr += pV1[0] * pV2[0] +
pV1[1] * pV2[1] +
@@ -154,77 +172,12 @@ double TDStretchSSE::calcCrossCorrStereo(const float *pV1, const float *pV2) con
pV1[14] * pV2[14] +
pV1[15] * pV2[15];
for (j = 0; j < 15; j ++) norm += pV1[j] * pV1[j];
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;
return corr / sqrt(norm);
*/
}
@@ -239,6 +192,7 @@ double TDStretchSSE::calcCrossCorrStereo(const float *pV1, const float *pV2) con
FIRFilterSSE::FIRFilterSSE() : FIRFilter()
{
filterCoeffsAlign = NULL;
filterCoeffsUnalign = NULL;
}
@@ -246,6 +200,8 @@ FIRFilterSSE::FIRFilterSSE() : FIRFilter()
FIRFilterSSE::~FIRFilterSSE()
{
delete[] filterCoeffsUnalign;
filterCoeffsAlign = NULL;
filterCoeffsUnalign = NULL;
}
@@ -258,15 +214,15 @@ void FIRFilterSSE::setCoefficients(const float *coeffs, uint newLength, uint uRe
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!
// also rearrange coefficients suitably for SSE
// 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);
filterCoeffsAlign = (float *)SOUNDTOUCH_ALIGN_POINTER_16(filterCoeffsUnalign);
fDivider = (float)resultDivider;
// rearrange the filter coefficients for mmx routines
// rearrange the filter coefficients for mmx routines
for (i = 0; i < newLength; i ++)
{
filterCoeffsAlign[2 * i + 0] =
@@ -279,15 +235,18 @@ void FIRFilterSSE::setCoefficients(const float *coeffs, uint newLength, uint uRe
// 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 count = (int)((numSamples - length) & (uint)-2);
int j;
assert(count % 2 == 0);
if (count < 2) return 0;
assert(source != NULL);
assert(dest != NULL);
assert((length % 8) == 0);
assert(((unsigned long)filterCoeffsAlign) % 16 == 0);
assert(filterCoeffsAlign != NULL);
assert(((ulongptr)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)
@@ -297,14 +256,14 @@ uint FIRFilterSSE::evaluateFilterStereo(float *dest, const float *source, uint n
__m128 sum1, sum2;
uint i;
pSrc = source; // source audio data
pFil = (__m128*)filterCoeffsAlign; // filter coefficients. NOTE: Assumes coefficients
// are aligned to 16-byte boundary
pSrc = (const float*)source; // source audio data
pFil = (const __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 ++)
for (i = 0; i < length / 8; i ++)
{
// Unroll loop for efficiency & calculate filter for 2*2 stereo samples
// 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
@@ -339,14 +298,14 @@ uint FIRFilterSSE::evaluateFilterStereo(float *dest, const float *source, uint n
}
// Ideas for further improvement:
// 1. If it could be guaranteed that 'source' were always aligned to 16-byte
// 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
// 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
/* original routine in C-language. please notice the C-version has differently
organized coefficients though.
double suml1, suml2;
double sumr1, sumr2;
@@ -361,26 +320,26 @@ uint FIRFilterSSE::evaluateFilterStereo(float *dest, const float *source, uint n
suml2 = sumr2 = 0.0;
ptr = src;
pFil = filterCoeffs;
for (i = 0; i < lengthLocal; i ++)
for (i = 0; i < lengthLocal; i ++)
{
// unroll loop for efficiency.
suml1 += ptr[0] * pFil[0] +
suml1 += ptr[0] * pFil[0] +
ptr[2] * pFil[2] +
ptr[4] * pFil[4] +
ptr[6] * pFil[6];
sumr1 += ptr[1] * pFil[1] +
sumr1 += ptr[1] * pFil[1] +
ptr[3] * pFil[3] +
ptr[5] * pFil[5] +
ptr[7] * pFil[7];
suml2 += ptr[8] * pFil[0] +
suml2 += ptr[8] * pFil[0] +
ptr[10] * pFil[2] +
ptr[12] * pFil[4] +
ptr[14] * pFil[6];
sumr2 += ptr[9] * pFil[1] +
sumr2 += ptr[9] * pFil[1] +
ptr[11] * pFil[3] +
ptr[13] * pFil[5] +
ptr[15] * pFil[7];
@@ -397,88 +356,6 @@ uint FIRFilterSSE::evaluateFilterStereo(float *dest, const float *source, uint n
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
#endif // SOUNDTOUCH_ALLOW_SSE