Merge branch 'DanielMerget-DanielMerget/fix_atomicAdd' into nikosk/pascal-and-cuda8-fixes

Source/Math/GPUMatrixCUDAKernels.cuh

@@ -42,7 +42,7 @@
 
 #define IDX2C(i, j, ld) (((j) * (ld)) + (i)) // 0 based indexing
 
-// CUDA atomicAdd() only exists for 'float'. This is the 'double' version.
+// On older GPUs, CUDA atomicAdd() only exists for 'float'. This is the 'double' version.
 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 600
 static __inline__ __device__ double atomicAdd(double* address, double val)
 {