__threadfence() across kernels? Fermi

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

The manual says that __threadfence() waits until memory accesses made by the thread are visible to all the threads in the device. Does it guarantee this kind of coherency between threads in different kernels?

Thanks

2

It sounds like you’re trying to pass information between two concurrent kernels at runtime. You really, really shouldn’t do that.

3

Actually, i wonder if I could have the host send notifications to a running kernel. A ‘notifier’ kernel could deliver the notification to the running kernel if this info could be passed between kernels in the device. Is there another way to pass notifications to a running kernel?

Thanks

4

zero-copy memory so you read host-memory in your kernel. If you mark it volatile it should read the value over the PCI-E bus every time you read it.

5

this is another one of those things you really shouldn’t depend on working…

6

Thanks tmurray

7

From a parallel programming perspective I agree wholeheartedly, but technically this is documented to work as far as I know (not that i rely on it ;)).

8

Nope, it’s actually not. Zero-copy to a running kernel is a bad idea because it is really tricky to do right and you may encounter chipset bugs. I I don’t think we’ve ever claimed that you can communicate in that way.