tcgu-amd
commented
3 weeks ago Hi @AngryLoki Thanks for reporting the issue! It seems like I am unable to reproduce your issue on ubuntu. Here are the steps I took:
- Use a rocm docker image with ubuntu (chose rocm/pytorch since I already had it on my system)
docker run -it rocm/pytorch /bin/bash - Install libc ++ from atp.llvm.org
sudo apt install lsb-release wget software-properties-commonwget https://apt.llvm.org/llvm.sh chmod +x llvm.sh sudo ./llvm.sh 20 sudo apt install libc++-20-dev libc++abi-20-dev - Create a test script
vi test.cpp - Save the following to the test script
#include <half/half.cpp> #include <cmath>
int main(){ half_float::half v(0.2); std::isfinite(v); return 0; }
- Compile with hipcc and libc++
```bash
hipcc -O2 --stdlib=libc++ test.cppAll I got was some warning regarding unused return value:
test.cpp:6:9: warning: ignoring return value of function declared with 'nodiscard' attribute [-Wunused-result]
6 | std::isfinite(v);
| ^~~~~~~~~~~~~ ~
1 warning generated when compiling for gfx906.
test.cpp:6:9: warning: ignoring return value of function declared with 'nodiscard' attribute [-Wunused-result]
6 | std::isfinite(v);
| ^~~~~~~~~~~~~ ~Would you be able to double check on your end and see if the issue persists?
Also, please keep in mind that ROCm does not officially support Gentoo. I will label this issue with "enhancement" to reflect this properly.
Thanks :D
AngryLoki
Problem Description
Hi,
I noticed a few problems affecting ROCm/half and upstream (v2.2 incl. trunk), so I'll report in both places.
Both are related to
std::isfinite. First problem is that it does not work with libc++. Here is a breakdown why: 1) Since C++11 std::isfinite is defined for integral types (having specific traits) 2) Used trait is std::is_arithmetic, it uses std::is_floating_point. Both traits have no specialization forhalf_float::half3) What happens in gcc/libstdc++/msvc is thatstd::isfinite(half_value)does implicit conversion to float and then checks for "isfinite". But it does not work in libc++.My current understanding is that it is a bug of libc++ (caused by misunderstanding of https://wg21.cmeerw.net/lwg/issue2086), so I reported it to https://github.com/llvm/llvm-project/issues/98816
Anyways, the second problem is that this conversion to float should not be called at all, because check could be much simpler:
Here is link to conformance view: https://godbolt.org/z/hhWebb9os
It looks safe, it fixes both problems mentioned above. Could you add it to half.hpp?
This affects ROCm/miopen, which currently fails with libc++ with:
Operating System
Gentoo
CPU
-
GPU
AMD Instinct MI300X
ROCm Version
ROCm 6.1.0
ROCm Component
No response
Steps to Reproduce
No response
(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support
No response
Additional Information
No response