numa aware allocation of FFT scratch space

bioem.cpp

@@ -319,10 +319,14 @@ int bioem::run()
 		const int count = omp_get_max_threads();
 		localCCT = new mycomplex_t*[count];
 		lCC = new myfloat_t*[count];
-		for (int i = 0;i < count;i++)
+		#pragma omp parallel
 		{
-			localCCT[i] = (mycomplex_t *) myfftw_malloc(sizeof(mycomplex_t) * param.param_device.NumberPixels * param.param_device.NumberFFTPixels1D);
-			lCC[i] = (myfloat_t *) myfftw_malloc(sizeof(myfloat_t) * param.param_device.NumberPixels * param.param_device.NumberPixels);
+			#pragma omp critical
+			{
+				const int i = omp_get_thread_num();
+				localCCT[i] = (mycomplex_t *) myfftw_malloc(sizeof(mycomplex_t) * param.param_device.NumberPixels * param.param_device.NumberFFTPixels1D);
+				lCC[i] = (myfloat_t *) myfftw_malloc(sizeof(myfloat_t) * param.param_device.NumberPixels * param.param_device.NumberPixels);
+			}
 		}
 	}