-
Notifications
You must be signed in to change notification settings - Fork 67
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Porting rocPRIM onto HIP-Clang #22
Comments
If you just want to make sure everything compiles, then you can locally remove |
Thanks, I'll work on the other build failures |
I need some help with changing the implementation here: https://github.com/ROCmSoftwarePlatform/rocPRIM/blob/master/rocprim/include/rocprim/device/detail/device_reduce_by_key.hpp#L343 It seems that in CUDA specs,
I tried alternatives like using typedef, and initializing later. Or removing the |
I guess some type in that structure has nontivial default ctor. Is this the only place with that problem? |
Seems to only occur in rocprim.hip.device_run_length_encode and hipcub.DeviceRunLengthEncode, both showing errors here |
My guess is that You would need to replace it with raw storage: |
You're right, it seems the compiler cannot accept result_type carry_in with attribute |
On Monday I'll pass that issue to my colleague who implemented that function. It's odd that |
I tried your method and it seems to work partially. Of the 26 tests in rocprim.hip.device_run_length_encode, only 3 are failing now. I'll make a PR so your colleague has reference for what I tried. |
Ok, thanks. We'll look into it. If you can, please provide test output in the PR. Question: are you certain that |
I think it should work in the same way as HCC. The implementation follows that in HCC. Do you know any tests in rocPRIM will test these shfl functions extensively? |
That would mean they don't work for HIP, unless it [hip-clang] already includes change you made.
No, we don't have any direct test for shfl functions. However, implementations of warp reduction and warp scan are quite straightforward, if they fail it's a strong indication there's something not right. |
Yes HIP-Clang is built with the master branch of HIP. It isn't completed yet, and still in the works. the master branch does include the change for shfl that I made. Okay, currently with HIP-Clang, my results look like this with -Werror disabled.
|
You can't compile |
No there is still build/link issue:
|
ok, linking problem. |
@aaronenyeshi, we've been thinking that you can take full control of the |
That sounds like a good workflow to me. Thank you @ncrkblacksmith |
Hi rocPRIM team, I'm working on porting rocPRIM to be compiled with HIP-Clang. I'm running into some issues that I'm not sure how to resolve since I don't understand the nature of the test. Here is what it looks like:
I've had this same error for a few other tests as well. I think that HIP-Clang compiler is a little more strict than the HCC option. Let me know what I should do to fix this. Thanks!
The text was updated successfully, but these errors were encountered: