minor formatting

app/audio/audiovisualwaveform.cpp

@@ -17,15 +17,15 @@
   along with this program.  If not, see <http://www.gnu.org/licenses/>.
 
 ***/
-#include <QtGlobal>
-
-#ifdef Q_PROCESSOR_X86_64
-#include <xmmintrin.h>
-#endif
 
 #include "audiovisualwaveform.h"
 
 #include <QDebug>
+#include <QtGlobal>
+
+#ifdef Q_PROCESSOR_X86
+#include <xmmintrin.h>
+#endif
 
 #include "config/config.h"
 #include "common/functiontimer.h"
@@ -291,8 +291,8 @@ AudioVisualWaveform::Sample AudioVisualWaveform::SumSamples(const float *samples
   return summed_samples;
 }
 
-#ifdef Q_PROCESSOR_X86_64
+#ifdef Q_PROCESSOR_X86
-void min_max_sse(float *a, int start, int end, float &min_val, float &max_val) 
+void ExpandMinMaxSSE(float *a, int start, int end, float &min_val, float &max_val)
 {
   // load the first 4 elements of 'a' into min and max (they are 4 * 32 = 128 bits)
   __m128 max = _mm_loadu_ps(a + start);
@@ -331,9 +331,9 @@ AudioVisualWaveform::Sample AudioVisualWaveform::SumSamples(SampleBufferPtr samp
   int channels = samples->audio_params().channel_count();
   AudioVisualWaveform::Sample summed_samples(channels);
 
-  #ifdef Q_PROCESSOR_X86_64
+  #ifdef Q_PROCESSOR_X86
     for (int channel=0; channel<samples->audio_params().channel_count(); channel++) {
-      min_max_sse(samples->data(channel), start_index, length, summed_samples[channel].min, summed_samples[channel].max);
+      ExpandMinMaxSSE(samples->data(channel), start_index, length, summed_samples[channel].min, summed_samples[channel].max);
     }
   #else
     int end_index = start_index + length;

app/node/math/math/mathbase.cpp

@@ -23,6 +23,10 @@
 #include <QMatrix4x4>
 #include <QVector2D>
 
+#ifdef Q_PROCESSOR_X86
+#include <xmmintrin.h>
+#endif
+
 #include "common/tohex.h"
 #include "node/distort/transform/transformdistortnode.h"
 #include "render/color.h"
@@ -349,9 +353,32 @@ void MathNodeBase::ValueInternal(Operation operation, Pairing pairing, const QSt
       if (IsInputStatic(number_param)) {
         if (!NumberIsNoOp(operation, number)) {
           for (int i=0;i<job.samples()->audio_params().channel_count();i++) {
+#ifdef Q_PROCESSOR_X86
+            // Use SSE instructions for optimization
+
+            float *a = job.samples()->data(i);
+            int end = job.samples()->sample_count();
+
+            int end_divisible_4 = (end / 4) * 4;
+
+            // loop over 'a' and compare current elements with min and max 4 by 4.
+            // we need to make sure we don't read out of boundaries should 'a' lenght be not mod. 4
+            __m128 mult = _mm_load1_ps(&number);
+            for(int j = 0; j < end_divisible_4; j+=4) {
+              __m128 cur = _mm_loadu_ps(a + j);
+              __m128 res = _mm_mul_ps(cur, mult);
+              _mm_storeu_ps(a + j, res);
+            }
+
+            // Do last three numbers, if non-divisible by 4
+            for (int j=end_divisible_4; j<end; j++) {
+              a[j] *= number;
+            }
+#else
             for (int j=0;j<job.samples()->sample_count();j++) {
               job.samples()->data(i)[j] = PerformAll(operation, job.samples()->data(i)[j], number);
             }
+#endif
           }
         }