Multiple arch and -maxrregcount

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

The Fermi compatibility guide mentions that you can create more than one version of your kernel, per arcitecture, and CUDART will decide which kernel

to run at runtime.

-gencode=arch=compute_10,code=sm_10 

-gencode=arch=compute_10,code=compute_10 

-gencode=arch=compute_20,code=sm_20 

-gencode=arch=compute_20,code=compute_20

If i have a mixed Tesla and Fermi cluster this is a great feature. However I have a kernel that I had to limit the reg count on Tesla with the -maxrregcount

flag but I dont want this to effect the sm_20 code for Fermi (and the -maxrregcount is in the makefile for both Tesla and Fermi)

How do I do this?

thanks

eyal

2

Did you try using the launch_bounds qualifier? I found it awkward and difficult to do what I wanted to do but maybe it was just me.

3

Did you try using the launch_bounds qualifier? I found it awkward and difficult to do what I wanted to do but maybe it was just me.

4

I’ll check it tomorrow :) I guess maybe the maxrregcount would work for sm < 2.0 and maybe the launch_bounds will override it

for sm == 2.0.

Can you please post how to use it? I tried it but it didnt compile and I couldnt find any reference to it in the SDK or the documents.

thanks

eyal

5

I’ll check it tomorrow :) I guess maybe the maxrregcount would work for sm < 2.0 and maybe the launch_bounds will override it

for sm == 2.0.

Can you please post how to use it? I tried it but it didnt compile and I couldnt find any reference to it in the SDK or the documents.

thanks

eyal

6

It’s section B.14 of the CUDA 3.0 Programming Guide.

7

It’s section B.14 of the CUDA 3.0 Programming Guide.