michael@0: ////////////////////////////////////////////////////////////////////////////////
michael@0: ///
michael@0: /// SSE optimized routines for Pentium-III, Athlon-XP and later CPUs. All SSE 
michael@0: /// optimized functions have been gathered into this single source 
michael@0: /// code file, regardless to their class or original source code file, in order 
michael@0: /// to ease porting the library to other compiler and processor platforms.
michael@0: ///
michael@0: /// The SSE-optimizations are programmed using SSE compiler intrinsics that
michael@0: /// are supported both by Microsoft Visual C++ and GCC compilers, so this file
michael@0: /// should compile with both toolsets.
michael@0: ///
michael@0: /// NOTICE: If using Visual Studio 6.0, you'll need to install the "Visual C++ 
michael@0: /// 6.0 processor pack" update to support SSE instruction set. The update is 
michael@0: /// available for download at Microsoft Developers Network, see here:
michael@0: /// http://msdn.microsoft.com/en-us/vstudio/aa718349.aspx
michael@0: ///
michael@0: /// If the above URL is expired or removed, go to "http://msdn.microsoft.com" and 
michael@0: /// perform a search with keywords "processor pack".
michael@0: ///
michael@0: /// Author        : Copyright (c) Olli Parviainen
michael@0: /// Author e-mail : oparviai 'at' iki.fi
michael@0: /// SoundTouch WWW: http://www.surina.net/soundtouch
michael@0: ///
michael@0: ////////////////////////////////////////////////////////////////////////////////
michael@0: //
michael@0: // Last changed  : $Date: 2014-01-07 12:25:40 -0600 (Tue, 07 Jan 2014) $
michael@0: // File revision : $Revision: 4 $
michael@0: //
michael@0: // $Id: sse_optimized.cpp 184 2014-01-07 18:25:40Z oparviai $
michael@0: //
michael@0: ////////////////////////////////////////////////////////////////////////////////
michael@0: //
michael@0: // License :
michael@0: //
michael@0: //  SoundTouch audio processing library
michael@0: //  Copyright (c) Olli Parviainen
michael@0: //
michael@0: //  This library is free software; you can redistribute it and/or
michael@0: //  modify it under the terms of the GNU Lesser General Public
michael@0: //  License as published by the Free Software Foundation; either
michael@0: //  version 2.1 of the License, or (at your option) any later version.
michael@0: //
michael@0: //  This library is distributed in the hope that it will be useful,
michael@0: //  but WITHOUT ANY WARRANTY; without even the implied warranty of
michael@0: //  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
michael@0: //  Lesser General Public License for more details.
michael@0: //
michael@0: //  You should have received a copy of the GNU Lesser General Public
michael@0: //  License along with this library; if not, write to the Free Software
michael@0: //  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
michael@0: //
michael@0: ////////////////////////////////////////////////////////////////////////////////
michael@0: 
michael@0: #include "cpu_detect.h"
michael@0: #include "STTypes.h"
michael@0: 
michael@0: using namespace soundtouch;
michael@0: 
michael@0: #ifdef SOUNDTOUCH_ALLOW_SSE
michael@0: 
michael@0: // SSE routines available only with float sample type    
michael@0: 
michael@0: //////////////////////////////////////////////////////////////////////////////
michael@0: //
michael@0: // implementation of SSE optimized functions of class 'TDStretchSSE'
michael@0: //
michael@0: //////////////////////////////////////////////////////////////////////////////
michael@0: 
michael@0: #include "TDStretch.h"
michael@0: #include <xmmintrin.h>
michael@0: #include <math.h>
michael@0: 
michael@0: // Calculates cross correlation of two buffers
michael@0: double TDStretchSSE::calcCrossCorr(const float *pV1, const float *pV2, double &norm) const
michael@0: {
michael@0:     int i;
michael@0:     const float *pVec1;
michael@0:     const __m128 *pVec2;
michael@0:     __m128 vSum, vNorm;
michael@0: 
michael@0:     // Note. It means a major slow-down if the routine needs to tolerate 
michael@0:     // unaligned __m128 memory accesses. It's way faster if we can skip 
michael@0:     // unaligned slots and use _mm_load_ps instruction instead of _mm_loadu_ps.
michael@0:     // This can mean up to ~ 10-fold difference (incl. part of which is
michael@0:     // due to skipping every second round for stereo sound though).
michael@0:     //
michael@0:     // Compile-time define SOUNDTOUCH_ALLOW_NONEXACT_SIMD_OPTIMIZATION is provided
michael@0:     // for choosing if this little cheating is allowed.
michael@0: 
michael@0: #ifdef SOUNDTOUCH_ALLOW_NONEXACT_SIMD_OPTIMIZATION
michael@0:     // Little cheating allowed, return valid correlation only for 
michael@0:     // aligned locations, meaning every second round for stereo sound.
michael@0: 
michael@0:     #define _MM_LOAD    _mm_load_ps
michael@0: 
michael@0:     if (((ulongptr)pV1) & 15) return -1e50;    // skip unaligned locations
michael@0: 
michael@0: #else
michael@0:     // No cheating allowed, use unaligned load & take the resulting
michael@0:     // performance hit.
michael@0:     #define _MM_LOAD    _mm_loadu_ps
michael@0: #endif 
michael@0: 
michael@0:     // ensure overlapLength is divisible by 8
michael@0:     assert((overlapLength % 8) == 0);
michael@0: 
michael@0:     // Calculates the cross-correlation value between 'pV1' and 'pV2' vectors
michael@0:     // Note: pV2 _must_ be aligned to 16-bit boundary, pV1 need not.
michael@0:     pVec1 = (const float*)pV1;
michael@0:     pVec2 = (const __m128*)pV2;
michael@0:     vSum = vNorm = _mm_setzero_ps();
michael@0: 
michael@0:     // Unroll the loop by factor of 4 * 4 operations. Use same routine for
michael@0:     // stereo & mono, for mono it just means twice the amount of unrolling.
michael@0:     for (i = 0; i < channels * overlapLength / 16; i ++) 
michael@0:     {
michael@0:         __m128 vTemp;
michael@0:         // vSum += pV1[0..3] * pV2[0..3]
michael@0:         vTemp = _MM_LOAD(pVec1);
michael@0:         vSum  = _mm_add_ps(vSum,  _mm_mul_ps(vTemp ,pVec2[0]));
michael@0:         vNorm = _mm_add_ps(vNorm, _mm_mul_ps(vTemp ,vTemp));
michael@0: 
michael@0:         // vSum += pV1[4..7] * pV2[4..7]
michael@0:         vTemp = _MM_LOAD(pVec1 + 4);
michael@0:         vSum  = _mm_add_ps(vSum, _mm_mul_ps(vTemp, pVec2[1]));
michael@0:         vNorm = _mm_add_ps(vNorm, _mm_mul_ps(vTemp ,vTemp));
michael@0: 
michael@0:         // vSum += pV1[8..11] * pV2[8..11]
michael@0:         vTemp = _MM_LOAD(pVec1 + 8);
michael@0:         vSum  = _mm_add_ps(vSum, _mm_mul_ps(vTemp, pVec2[2]));
michael@0:         vNorm = _mm_add_ps(vNorm, _mm_mul_ps(vTemp ,vTemp));
michael@0: 
michael@0:         // vSum += pV1[12..15] * pV2[12..15]
michael@0:         vTemp = _MM_LOAD(pVec1 + 12);
michael@0:         vSum  = _mm_add_ps(vSum, _mm_mul_ps(vTemp, pVec2[3]));
michael@0:         vNorm = _mm_add_ps(vNorm, _mm_mul_ps(vTemp ,vTemp));
michael@0: 
michael@0:         pVec1 += 16;
michael@0:         pVec2 += 4;
michael@0:     }
michael@0: 
michael@0:     // return value = vSum[0] + vSum[1] + vSum[2] + vSum[3]
michael@0:     float *pvNorm = (float*)&vNorm;
michael@0:     norm = (pvNorm[0] + pvNorm[1] + pvNorm[2] + pvNorm[3]);
michael@0: 
michael@0:     float *pvSum = (float*)&vSum;
michael@0:     return (double)(pvSum[0] + pvSum[1] + pvSum[2] + pvSum[3]) / sqrt(norm < 1e-9 ? 1.0 : norm);
michael@0: 
michael@0:     /* This is approximately corresponding routine in C-language yet without normalization:
michael@0:     double corr, norm;
michael@0:     uint i;
michael@0: 
michael@0:     // Calculates the cross-correlation value between 'pV1' and 'pV2' vectors
michael@0:     corr = norm = 0.0;
michael@0:     for (i = 0; i < channels * overlapLength / 16; i ++) 
michael@0:     {
michael@0:         corr += pV1[0] * pV2[0] +
michael@0:                 pV1[1] * pV2[1] +
michael@0:                 pV1[2] * pV2[2] +
michael@0:                 pV1[3] * pV2[3] +
michael@0:                 pV1[4] * pV2[4] +
michael@0:                 pV1[5] * pV2[5] +
michael@0:                 pV1[6] * pV2[6] +
michael@0:                 pV1[7] * pV2[7] +
michael@0:                 pV1[8] * pV2[8] +
michael@0:                 pV1[9] * pV2[9] +
michael@0:                 pV1[10] * pV2[10] +
michael@0:                 pV1[11] * pV2[11] +
michael@0:                 pV1[12] * pV2[12] +
michael@0:                 pV1[13] * pV2[13] +
michael@0:                 pV1[14] * pV2[14] +
michael@0:                 pV1[15] * pV2[15];
michael@0: 
michael@0:     for (j = 0; j < 15; j ++) norm += pV1[j] * pV1[j];
michael@0: 
michael@0:         pV1 += 16;
michael@0:         pV2 += 16;
michael@0:     }
michael@0:     return corr / sqrt(norm);
michael@0:     */
michael@0: }
michael@0: 
michael@0: 
michael@0: 
michael@0: double TDStretchSSE::calcCrossCorrAccumulate(const float *pV1, const float *pV2, double &norm) const
michael@0: {
michael@0:     // call usual calcCrossCorr function because SSE does not show big benefit of 
michael@0:     // accumulating "norm" value, and also the "norm" rolling algorithm would get 
michael@0:     // complicated due to SSE-specific alignment-vs-nonexact correlation rules.
michael@0:     return calcCrossCorr(pV1, pV2, norm);
michael@0: }
michael@0: 
michael@0: 
michael@0: //////////////////////////////////////////////////////////////////////////////
michael@0: //
michael@0: // implementation of SSE optimized functions of class 'FIRFilter'
michael@0: //
michael@0: //////////////////////////////////////////////////////////////////////////////
michael@0: 
michael@0: #include "FIRFilter.h"
michael@0: 
michael@0: FIRFilterSSE::FIRFilterSSE() : FIRFilter()
michael@0: {
michael@0:     filterCoeffsAlign = NULL;
michael@0:     filterCoeffsUnalign = NULL;
michael@0: }
michael@0: 
michael@0: 
michael@0: FIRFilterSSE::~FIRFilterSSE()
michael@0: {
michael@0:     delete[] filterCoeffsUnalign;
michael@0:     filterCoeffsAlign = NULL;
michael@0:     filterCoeffsUnalign = NULL;
michael@0: }
michael@0: 
michael@0: 
michael@0: // (overloaded) Calculates filter coefficients for SSE routine
michael@0: void FIRFilterSSE::setCoefficients(const float *coeffs, uint newLength, uint uResultDivFactor)
michael@0: {
michael@0:     uint i;
michael@0:     float fDivider;
michael@0: 
michael@0:     FIRFilter::setCoefficients(coeffs, newLength, uResultDivFactor);
michael@0: 
michael@0:     // Scale the filter coefficients so that it won't be necessary to scale the filtering result
michael@0:     // also rearrange coefficients suitably for SSE
michael@0:     // Ensure that filter coeffs array is aligned to 16-byte boundary
michael@0:     delete[] filterCoeffsUnalign;
michael@0:     filterCoeffsUnalign = new float[2 * newLength + 4];
michael@0:     filterCoeffsAlign = (float *)SOUNDTOUCH_ALIGN_POINTER_16(filterCoeffsUnalign);
michael@0: 
michael@0:     fDivider = (float)resultDivider;
michael@0: 
michael@0:     // rearrange the filter coefficients for mmx routines 
michael@0:     for (i = 0; i < newLength; i ++)
michael@0:     {
michael@0:         filterCoeffsAlign[2 * i + 0] =
michael@0:         filterCoeffsAlign[2 * i + 1] = coeffs[i + 0] / fDivider;
michael@0:     }
michael@0: }
michael@0: 
michael@0: 
michael@0: 
michael@0: // SSE-optimized version of the filter routine for stereo sound
michael@0: uint FIRFilterSSE::evaluateFilterStereo(float *dest, const float *source, uint numSamples) const
michael@0: {
michael@0:     int count = (int)((numSamples - length) & (uint)-2);
michael@0:     int j;
michael@0: 
michael@0:     assert(count % 2 == 0);
michael@0: 
michael@0:     if (count < 2) return 0;
michael@0: 
michael@0:     assert(source != NULL);
michael@0:     assert(dest != NULL);
michael@0:     assert((length % 8) == 0);
michael@0:     assert(filterCoeffsAlign != NULL);
michael@0:     assert(((ulongptr)filterCoeffsAlign) % 16 == 0);
michael@0: 
michael@0:     // filter is evaluated for two stereo samples with each iteration, thus use of 'j += 2'
michael@0:     for (j = 0; j < count; j += 2)
michael@0:     {
michael@0:         const float *pSrc;
michael@0:         const __m128 *pFil;
michael@0:         __m128 sum1, sum2;
michael@0:         uint i;
michael@0: 
michael@0:         pSrc = (const float*)source;              // source audio data
michael@0:         pFil = (const __m128*)filterCoeffsAlign;  // filter coefficients. NOTE: Assumes coefficients 
michael@0:                                                   // are aligned to 16-byte boundary
michael@0:         sum1 = sum2 = _mm_setzero_ps();
michael@0: 
michael@0:         for (i = 0; i < length / 8; i ++) 
michael@0:         {
michael@0:             // Unroll loop for efficiency & calculate filter for 2*2 stereo samples 
michael@0:             // at each pass
michael@0: 
michael@0:             // sum1 is accu for 2*2 filtered stereo sound data at the primary sound data offset
michael@0:             // sum2 is accu for 2*2 filtered stereo sound data for the next sound sample offset.
michael@0: 
michael@0:             sum1 = _mm_add_ps(sum1, _mm_mul_ps(_mm_loadu_ps(pSrc)    , pFil[0]));
michael@0:             sum2 = _mm_add_ps(sum2, _mm_mul_ps(_mm_loadu_ps(pSrc + 2), pFil[0]));
michael@0: 
michael@0:             sum1 = _mm_add_ps(sum1, _mm_mul_ps(_mm_loadu_ps(pSrc + 4), pFil[1]));
michael@0:             sum2 = _mm_add_ps(sum2, _mm_mul_ps(_mm_loadu_ps(pSrc + 6), pFil[1]));
michael@0: 
michael@0:             sum1 = _mm_add_ps(sum1, _mm_mul_ps(_mm_loadu_ps(pSrc + 8) ,  pFil[2]));
michael@0:             sum2 = _mm_add_ps(sum2, _mm_mul_ps(_mm_loadu_ps(pSrc + 10), pFil[2]));
michael@0: 
michael@0:             sum1 = _mm_add_ps(sum1, _mm_mul_ps(_mm_loadu_ps(pSrc + 12), pFil[3]));
michael@0:             sum2 = _mm_add_ps(sum2, _mm_mul_ps(_mm_loadu_ps(pSrc + 14), pFil[3]));
michael@0: 
michael@0:             pSrc += 16;
michael@0:             pFil += 4;
michael@0:         }
michael@0: 
michael@0:         // Now sum1 and sum2 both have a filtered 2-channel sample each, but we still need
michael@0:         // to sum the two hi- and lo-floats of these registers together.
michael@0: 
michael@0:         // post-shuffle & add the filtered values and store to dest.
michael@0:         _mm_storeu_ps(dest, _mm_add_ps(
michael@0:                     _mm_shuffle_ps(sum1, sum2, _MM_SHUFFLE(1,0,3,2)),   // s2_1 s2_0 s1_3 s1_2
michael@0:                     _mm_shuffle_ps(sum1, sum2, _MM_SHUFFLE(3,2,1,0))    // s2_3 s2_2 s1_1 s1_0
michael@0:                     ));
michael@0:         source += 4;
michael@0:         dest += 4;
michael@0:     }
michael@0: 
michael@0:     // Ideas for further improvement:
michael@0:     // 1. If it could be guaranteed that 'source' were always aligned to 16-byte 
michael@0:     //    boundary, a faster aligned '_mm_load_ps' instruction could be used.
michael@0:     // 2. If it could be guaranteed that 'dest' were always aligned to 16-byte 
michael@0:     //    boundary, a faster '_mm_store_ps' instruction could be used.
michael@0: 
michael@0:     return (uint)count;
michael@0: 
michael@0:     /* original routine in C-language. please notice the C-version has differently 
michael@0:        organized coefficients though.
michael@0:     double suml1, suml2;
michael@0:     double sumr1, sumr2;
michael@0:     uint i, j;
michael@0: 
michael@0:     for (j = 0; j < count; j += 2)
michael@0:     {
michael@0:         const float *ptr;
michael@0:         const float *pFil;
michael@0: 
michael@0:         suml1 = sumr1 = 0.0;
michael@0:         suml2 = sumr2 = 0.0;
michael@0:         ptr = src;
michael@0:         pFil = filterCoeffs;
michael@0:         for (i = 0; i < lengthLocal; i ++) 
michael@0:         {
michael@0:             // unroll loop for efficiency.
michael@0: 
michael@0:             suml1 += ptr[0] * pFil[0] + 
michael@0:                      ptr[2] * pFil[2] +
michael@0:                      ptr[4] * pFil[4] +
michael@0:                      ptr[6] * pFil[6];
michael@0: 
michael@0:             sumr1 += ptr[1] * pFil[1] + 
michael@0:                      ptr[3] * pFil[3] +
michael@0:                      ptr[5] * pFil[5] +
michael@0:                      ptr[7] * pFil[7];
michael@0: 
michael@0:             suml2 += ptr[8] * pFil[0] + 
michael@0:                      ptr[10] * pFil[2] +
michael@0:                      ptr[12] * pFil[4] +
michael@0:                      ptr[14] * pFil[6];
michael@0: 
michael@0:             sumr2 += ptr[9] * pFil[1] + 
michael@0:                      ptr[11] * pFil[3] +
michael@0:                      ptr[13] * pFil[5] +
michael@0:                      ptr[15] * pFil[7];
michael@0: 
michael@0:             ptr += 16;
michael@0:             pFil += 8;
michael@0:         }
michael@0:         dest[0] = (float)suml1;
michael@0:         dest[1] = (float)sumr1;
michael@0:         dest[2] = (float)suml2;
michael@0:         dest[3] = (float)sumr2;
michael@0: 
michael@0:         src += 4;
michael@0:         dest += 4;
michael@0:     }
michael@0:     */
michael@0: }
michael@0: 
michael@0: #endif  // SOUNDTOUCH_ALLOW_SSE