Issue #451 part 1 - use OpenMP for SoundTouch

media/libsoundtouch/src/FIRFilter.cpp

@@ -86,8 +86,11 @@ uint FIRFilter::evaluateFilterStereo(SAMPLETYPE *dest, const SAMPLETYPE *src, ui
     assert((length != 0) && (length == ilength) && (src != NULL) && (dest != NULL) && (filterCoeffs != NULL));
 
     end = 2 * (numSamples - ilength);
+    const SAMPLETYPE *coeffsStereo = filterCoeffsStereo;
 
-    #pragma omp parallel for
+#if defined(_OPENMP)
+    #pragma omp parallel for default(none) shared(src, dest, end, ilength, coeffsStereo) private(j) schedule(static)
+#endif
     for (j = 0; j < end; j += 2) 
     {
         const SAMPLETYPE *ptr;
@@ -98,8 +101,8 @@ uint FIRFilter::evaluateFilterStereo(SAMPLETYPE *dest, const SAMPLETYPE *src, ui
 
         for (int i = 0; i < ilength; i ++)
         {
-            suml += ptr[2 * i] * filterCoeffsStereo[2 * i];
-            sumr += ptr[2 * i + 1] * filterCoeffsStereo[2 * i + 1];
+            suml += ptr[2 * i] * coeffsStereo[2 * i];
+            sumr += ptr[2 * i + 1] * coeffsStereo[2 * i + 1];
         }
 
 #ifdef SOUNDTOUCH_INTEGER_SAMPLES
@@ -133,7 +136,10 @@ uint FIRFilter::evaluateFilterMono(SAMPLETYPE *dest, const SAMPLETYPE *src, uint
     assert(ilength != 0);
 
     end = numSamples - ilength;
-    #pragma omp parallel for
+    const SAMPLETYPE *coeffs = filterCoeffs;
+#if defined(_OPENMP)
+    #pragma omp parallel for default(none) shared(src, dest, end, ilength, coeffs) private(j) schedule(static)
+#endif
     for (j = 0; j < end; j ++)
     {
         const SAMPLETYPE *pSrc = src + j;
@@ -143,7 +149,7 @@ uint FIRFilter::evaluateFilterMono(SAMPLETYPE *dest, const SAMPLETYPE *src, uint
         sum = 0;
         for (i = 0; i < ilength; i ++)
         {
-            sum += pSrc[i] * filterCoeffs[i];
+            sum += pSrc[i] * coeffs[i];
         }
 #ifdef SOUNDTOUCH_INTEGER_SAMPLES
         sum >>= resultDivFactor;
@@ -176,8 +182,11 @@ uint FIRFilter::evaluateFilterMulti(SAMPLETYPE *dest, const SAMPLETYPE *src, uin
     int ilength = length & -8;
 
     end = numChannels * (numSamples - ilength);
+    const SAMPLETYPE *coeffs = filterCoeffs;
 
-    #pragma omp parallel for
+#if defined(_OPENMP)
+    #pragma omp parallel for default(none) shared(src, dest, end, ilength, numChannels, coeffs) private(j) schedule(static)
+#endif
     for (j = 0; j < end; j += numChannels)
     {
         const SAMPLETYPE *ptr;
@@ -194,7 +203,7 @@ uint FIRFilter::evaluateFilterMulti(SAMPLETYPE *dest, const SAMPLETYPE *src, uin
 
         for (i = 0; i < ilength; i ++)
         {
-            SAMPLETYPE coef=filterCoeffs[i];
+            SAMPLETYPE coef=coeffs[i];
             for (c = 0; c < numChannels; c ++)
             {
                 sums[c] += ptr[0] * coef;

media/libsoundtouch/src/TDStretch.cpp

@@ -306,25 +306,31 @@ int TDStretch::seekBestOverlapPositionFull(const SAMPLETYPE *refPos)
     // over the permitted range.
     bestCorr = calcCrossCorr(refPos, pMidBuffer, norm);
     bestCorr = (bestCorr + 0.1) * 0.75;
+    const int seekLen = seekLength;
+    const int nChannels = channels;
+    const SAMPLETYPE *midBuffer = pMidBuffer;
 
-    #pragma omp parallel for
-    for (i = 1; i < seekLength; i ++)
+    #if defined(_OPENMP)
+        #pragma omp parallel for default(none) shared(refPos, bestCorr, bestOffs, seekLen, nChannels, midBuffer) private(i) schedule(static)
+    #endif
+    for (i = 1; i < seekLen; i ++)
     {
         double corr;
+        double localNorm;
         // Calculates correlation value for the mixing position corresponding to 'i'
 #if defined(_OPENMP) || defined(ST_SIMD_AVOID_UNALIGNED)
         // in parallel OpenMP mode, can't use norm accumulator version as parallel executor won't
         // iterate the loop in sequential order
         // in SIMD mode, avoid accumulator version to allow avoiding unaligned positions
-        corr = calcCrossCorr(refPos + channels * i, pMidBuffer, norm);
+        corr = calcCrossCorr(refPos + nChannels * i, midBuffer, localNorm);
 #else
         // In non-parallel version call "calcCrossCorrAccumulate" that is otherwise same
         // as "calcCrossCorr", but saves time by reusing & updating previously stored 
         // "norm" value
-        corr = calcCrossCorrAccumulate(refPos + channels * i, pMidBuffer, norm);
+        corr = calcCrossCorrAccumulate(refPos + nChannels * i, midBuffer, norm);
 #endif
         // heuristic rule to slightly favour values close to mid of the range
-        double tmp = (double)(2 * i - seekLength) / (double)seekLength;
+        double tmp = (double)(2 * i - seekLen) / (double)seekLen;
         corr = ((corr + 0.1) * (1.0 - 0.25 * tmp * tmp));
 
         // Checks for the highest correlation value
@@ -333,7 +339,9 @@ int TDStretch::seekBestOverlapPositionFull(const SAMPLETYPE *refPos)
             // For optimal performance, enter critical section only in case that best value found.
             // in such case repeat 'if' condition as it's possible that parallel execution may have
             // updated the bestCorr value in the mean time
-            #pragma omp critical
+#if defined(_OPENMP)
+            #pragma omp critical(soundtouch_bestcorr)
+#endif
             if (corr > bestCorr)
             {
                 bestCorr = corr;
@@ -910,7 +918,9 @@ double TDStretch::calcCrossCorr(const short *mixingPos, const short *compare, do
     if (lnorm > maxnorm)
     {
         // modify 'maxnorm' inside critical section to avoid multi-access conflict if in OpenMP mode
-        #pragma omp critical
+    #if defined(_OPENMP)
+        #pragma omp critical(soundtouch_maxnorm)
+    #endif
         if (lnorm > maxnorm)
         {
             maxnorm = lnorm;

media/libsoundtouch/src/mmx_optimized.cpp

@@ -119,7 +119,9 @@ double TDStretchMMX::calcCrossCorr(const short *pV1, const short *pV2, double &d
     if (norm > (long)maxnorm)
     {
         // modify 'maxnorm' inside critical section to avoid multi-access conflict if in OpenMP mode
-        #pragma omp critical
+    #if defined(_OPENMP)
+        #pragma omp critical(soundtouch_maxnorm)
+    #endif
         if (norm > (long)maxnorm)
         {
             maxnorm = norm;

media/libsoundtouch/src/moz.build

@@ -41,6 +41,11 @@ else:
     # Windows need alloca renamed to _alloca
     DEFINES['alloca'] = '_alloca'
 
+# Use standardized OpenMP parallelization for SoundTouch hot loops.
+if CONFIG['CC_TYPE'] == 'gcc':
+    CXXFLAGS += ['-fopenmp']
+    OS_LIBS += ['gomp']
+
 DEFINES['BUILDING_SOUNDTOUCH'] = 1
 
 # We allow warnings for third-party code that can be updated from upstream.

media/libsoundtouch/src/sse_optimized.cpp

@@ -250,9 +250,13 @@ uint FIRFilterSSE::evaluateFilterStereo(float *dest, const float *source, uint n
     assert((length % 8) == 0);
     assert(filterCoeffsAlign != NULL);
     assert(((ulongptr)filterCoeffsAlign) % 16 == 0);
+    const uint filterLength = length;
+    const float *coeffsAlign = filterCoeffsAlign;
 
     // filter is evaluated for two stereo samples with each iteration, thus use of 'j += 2'
-    #pragma omp parallel for
+#if defined(_OPENMP)
+    #pragma omp parallel for default(none) shared(source, dest, count, filterLength, coeffsAlign) private(j) schedule(static)
+#endif
     for (j = 0; j < count; j += 2)
     {
         const float *pSrc;
@@ -263,11 +267,11 @@ uint FIRFilterSSE::evaluateFilterStereo(float *dest, const float *source, uint n
 
         pSrc = (const float*)source + j * 2;      // source audio data
         pDest = dest + j * 2;                     // destination audio data
-        pFil = (const __m128*)filterCoeffsAlign;  // filter coefficients. NOTE: Assumes coefficients 
+        pFil = (const __m128*)coeffsAlign;  // 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 < filterLength / 8; i ++) 
         {
             // Unroll loop for efficiency & calculate filter for 2*2 stereo samples 
             // at each pass