How to compile a dll for more than one arhcitecure?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

I have developed a library and I want to be able to use it on sm_arch 1.3, 1.1 and 1.1.

How can I do that?

Actually, I compile one lib with the -arch sm10 option and one lib with -arch sm13 option, and I change the .dll in my bin folder, it’s not very user-friendly.

I use CUDA 2.3 and I can’t use 3.0 for some reasons.

Is there a more simple method than use in the reduce example in the SDK?

Thanks for any help!

Letinono

2

Maybe just compile for 1.0 with ptx code, it may scales fine with other gpu. also you may use driver api or make 3 files with 3 kernel copies and compile each file with different compilation options.
Also you may make dll wrapper for your 3 different dlls.

3

Ok, thanks for your help, I will use the first option :)