const cache instead of texture cache about rapidcfd-dev HOT 4 CLOSED

Hello,

The commit you linked seems to be private and not accessible for me.

In theory nvcc should itself generate the __ldg instruction for pointers marked const and __restrict __. However in our case the pointers are not regular function arguments but are passed around inside functors, which makes the compiler hesitate about generating this optimization. In my experience with CUDA 7.0 and 6.5, the compiler treated __restrict __ pointers inside functors as regular pointers, therefore I decided to include this instruction manually.

from rapidcfd-dev.

Hello,

For reasons I cannot explain, clinking on the embedded link does lead to a page not found error. But if I copy and paste the URL into a browser I can see the commit on my fork of RapidCFD. You can also try copying and pasting this tinyURL into your browser to get to my commit: https://tinyurl.com/y6v7xgrs Again, simply clicking on the link fails, but copy and paste works.

Your feedback on the __restrict __ pointers is very helpful. If you can access the commit, I am interested in your feedback on where the intrinsic __ldg function should be applied. If you do not have time, I will continue to think about this and work it out in time.

from rapidcfd-dev.

Hello,

I may have figured out where to apply the intrinsic __ldg function. In addition to the URL above, please look at https://github.com/TonkomoLLC/RapidCFD-dev/commit/27dfbb054849e2e7b622aa69921f9fd8bd9357ca
You may need to copy and paste the above URL into your browser.

The code compiles and executes with the proposed changes. However, I am interested in feedback on this implementation.

from rapidcfd-dev.

Since commit 6b40d01 constant cache is for devices with CC >= 3.5.

from rapidcfd-dev.