i’m a computational science college student. i need the source file of cublas to study on especially funtions on the blas level seems that it was once available on the forum during version 1.1 but unfortunately now its download link is removed
and i also know that registering for the nvidia developers group will help gain access to the latest and complete source files of cublas and cufft, but i am afraid that i am not qualified to be accepted into the group.but i really need it. so is anyone here could help me with it? :unsure: please email me .THANKS IN ADVANCE!

You can find the source for one of the current CUBLAS SGEMM implementations here. That is about all you can expect, I am afraid.

THANKS! is it mean the program from the user is accepted into the official cublas?! it’s really cool!

here’s some information maybe useful for people here-magma-a project to realize lapack on cuda

the latest version of it is the program the developers have optimize some of the function in blas ,which is most on level 2,especially those they need to improve the function in lapack which is already available.

but also unfortunately, it’s source is not open yet.but some of them is released to meet the need of users!(the link i provide)

and on the whitepaper provided it is said the blas they realized is far more efficient than cublas, which i have not yet tried

and the version 0.3 will coming soon, which is said to conclude many eigenvalue programm which involved a lot of computations on blas level 3,so perhaps they will provide better *gemm soon.

hope my information is useful for you !

No it just means that somebody wrote a faster gemm kernel that the kernel that was in earlier versions of CUBLAS, and because it was made available under a permissive license (BSD in this case, I think), NVIDIA took the code and integrated it in into CUBLAS.

ok…it seems that nowadays there are less people devoted to the optimization of blas on gpus, not as many as those focus on that on cpus at that time.

maybe it’s just the architecture of gpus is not stable enough,so devotion can be worthless as soon as the new type of hardware is released. am i right? so if i were right, if that means focus on it is some kind of time waste…but i am really interested in it

sorry for my awkward english, i’m not from an english speaking country.