Can you tell me the name of CUBLAS function for multiplying a MxN matrix with a N element vector?
I will call *gemv with alpha=1 and beta=0.
But the fuction would be performing a lot of useless computions because of multiplications with the 1 and 0.
Is there a faster version available?
or any other way out of it?
Unless this is a bottleneck in your code, why worry? It’s even possible that the CUBLAS library spots a few special cases like that itself, and has custom kernels for them (don’t quote me on that… I’ve not checked the source code myself). Cogitating a little more, I suspect that gemv is bandwidth, not compute bound, which means that special-casing things won’t make any measureable difference. But seriously, unless you know that this is the slowest section in your code, I really wouldn’t worry about it. Programmer time will be better spent elsewhere.
Is the source code of gemv function available?
If yes, where?
If it is then my problem is solved…
Thanx in advance.