How to add scalar to vector?

[Matt Phillips]

I've asked this on SO,http://stackoverflow.com/questions/14051064/add-scalar-to-vector-in-blas-cublas-cuda but the answer I got--use axpy with a unit vector scaled by the number I want to add--seems inefficient, requiring unnecessary memory allocation and multiplication. So I wanted to check here and see if there is a better way. I'm also curious about the rationale for leaving out such an obvious and essential operation, if anybody has some insight there.

[Julien Langou]

Code: Select all

CALL DAXPY( N, 1.0D+00 , X, INCX, ALPHA, 0 )

where ALPHA is a DOUBLE PRECISION number.

So the "trick" is to use a ZERO increment for your SCALAR.

A good implementation of DAXPY should have an IF statement to avoid the multiplication by 1.0D+00.

For CPU, a compiler should be able to optimize this kind of loop very efficiently. (This is a bandwidth bound problem so you will not go faster than the bandwidth from your main memory to your CPU.) So writing your own kernel makes sense.
For GPU, yes, if you do not want to mess up with programming on GPU and simply use a GPU BLAS, DAXPY should do the work.

Cheers,
Julien.

[Matt Phillips]

Thank you, that is exactly the kind of answer I am looking for. However it is slower than doing the addition 'by hand' in a for-loop, to my surprise. However this is using the CBLAS implementation that comes with the GSL, which I did not manually 'tune' in any way. My implementation could very well be sub-optimal, I haven't gotten to that stage yet.

[sven]

If you were using Fortran, then you could just use the Fortran array statement

Code: Select all

x = alpha + x

!!

A minor comment on Julien's code, strictly speaking alpha should be declared as at least a one element array for the call to DAXPY to be legal Fortran (and I presume C as well).

[Matt Phillips]

Hi, thanks for the tip about Fortran, I did not know that it had vector operations built in like that, that's very good to know. I'm in fact coding in C; I don't have the test code in front of me but I think passing the address of a

Code: Select all

const double

element as ALPHA worked without compiler warning/error/unexpected behavior.