opal/mca/accelarator: remove unused flag - #13784
Conversation
Remove unused device only accelerator flag Signed-off-by: Alexia Ingerson <alexia.ingerson@intel.com>
|
I'm not sure that removing this flag is the right approach. Why not use it in |
|
That's what I was originally doing in my first implementation but then I looked closer and realized that this path would never be taken with host memory (whether allocated with the accelerator API or not). So then you only have two options for the memory - unified or device only. So then I simplified it to just checking that it is not unified memory. But then I forgot to remove the flag :/ |
|
The flag was just recently introduced by @aingerson in this PR here: #13582, I haven't thought through whether was truly required for ofi mtl/btl operations or not. So I am neutral on removing it again, the ofi experts will have to weigh in on that. |
|
In regards to how it interacts with OFI, there are certain optimizations we can or cannot enable depending on whether the memory was allocated with the accelerator interface using host, shared/unified, or device only memory. The device only flag in the original PR basically allows us to enable IPC transfers (cannot be used with host or unified memory). As long as device_only is always the same as !unified, then I think removing should be sufficient. But if you'd rather be safe or you think host memory can go through this path, I can keep the flag and add the explicit check. |
|
@aingerson if I understand the logic in the original PR correctly, I agree that this flag is not really required and could be removed |
Remove unused device only accelerator flag