Julia pitfall

1. BLAS has pre-check system: if the input is zero matrix/vector, then BLAS will return zero matrix/vector directly without calculating. Thus, the speed is faster.

   Example:

   BLAS.axpy!(oldAlpha-α[j],x,yCorr)
   

   If oldAlpha-α[j] is a zero vector, BLAS will return zero vector directly. If oldAlpha-α[j] and x are not zero vectors, BLAS will continue to calculate.

   In conclusion: when testing speed, please check whether there exists zero vector/matrix.

2. reshape is reference, not copy.

a=[1,1,1]
b=reshape(a,1,3)
b[1]=999
b
a

Extra pitfall from Tianjing:

1. I found this pitfall when I changed BLAS function to normal function. In helper function, the BLAS changes ycorr in global scope, but += changes ycorr only in function scope. So, we have to return ycorr in helper function.

   • BLAS version:

   function helper(ycorr)
      BLAS.axpy!(a,x,ycorr)
   end
   
   function f()
       helper(ycorr)
   end
   

   • Normal version #WRONG!

   function helper(ycorr)
      ycorr += a*x
   end
   
   function f()
       helper(ycorr)
   end
   

   • Right version

   function helper(ycorr)
      ycorr += a*x
      return ycorr 
   end
   
   function f()
       ycorr = helper(ycorr)
   end