CuPy for HIP vs ClPy for OpenCL for Radeon VII
See original GitHub issueHello,
I’ve been recently trying to run my existing CuPy code (that currently runs on an NVIDIA Titan X (Maxwell) GPU on another new system with an AMD Radeon VII GPU for the same project. On the NVIDIA based system, I was successfully able to install CuPy within a Conda env with Pip.
For the new AMD based system, so far I have followed the instructions as specified at PR #1094 and have been trying to install it. I was able to set it up but was unable to execute import cupy
successfully. It kept giving an error for hipsparse: undefined symbol: _ZN8hip_impl8hip_initEv
. At that time I was using ROCm 2.2. After upgrading to ROCm 2.3, I now get a new error for cupy/core/core.so: undefined symbol: hipModuleGetGlobal
when trying to test import cupy
. Is there a workaround for this?
Yesterday, I also came to know about ClPy. I also found that it received ROCm OpenCL support very recently. Should I be going for ClPy instead? As of now, is ClPy a better option performance wise? When can we expect CuPy for HIP to be updated like ClPy?
Thank you so much for developing CuPy for both platforms! Looking forward to your valuable feedback.
Best, Avimanyu
Issue Analytics
- State:
- Created 4 years ago
- Reactions:1
- Comments:7 (6 by maintainers)
Top GitHub Comments
CuPy v7.0.0rc1 was released (include #1094). It supports HIP environment!
Thank you for testing CuPy HIP branch #1094. It branch has many issues. In now, using ClPy is a smart option.
I will continue to support ROCm. However, work is taking time because ROCm is not stable for CuPy. CuPy uses internal interface of ROCm to support just in time compile.