Windows Support for hip #13
Conversation
|
wait, is the goal to have the changes in the generate scripts? |
|
I meant to call it |
|
Err, and you can't really modify |
|
yeah redoing to inject it from the script |
|
The script injection stuff really needs to be cleaned up if it's going to be more than a one line change |
test/test_hip.py
Outdated
|
|
||
| def test_compile(self): | ||
| prg = cuda_compile("int test() { return 42; }", ["--offload-arch=gfx1100"], HIPCompile, check) | ||
| prg = compile("int test() { return 42; }", ["--offload-arch=gfx1100"], HIPCompile, check, filename=None) |
There was a problem hiding this comment.
Windows isn't okay with <null>?
There was a problem hiding this comment.
it would throw an error about the profiler not being initialized, not sure how the filename relates to it tho but can debug when I have some spare time
|
#14 should make CI fail if the autogeneration doesn't match |
|
The |
dc-dc-dc
force-pushed
the
hip_windows
branch
from
5a8e0a5
to
3b13ecc
Compare
|
fixed CI error |
| if 'linux' in sys.platform: | ||
| return ctypes.CDLL(os.path.join('/opt/rocm/lib/libhiprtc.so')) | ||
| elif 'win' in sys.platform: | ||
| hip_path = os.getenv('HIP_PATH', None) |
There was a problem hiding this comment.
Is there a way to not need this? How come it can find amdhip64?
There was a problem hiding this comment.
amdhip64 gets installed in System32 where cdll searches for dlls, the rtc does not so have to use HIP_PATH to find the hip installation. HIP_PATH is added to the env on install of the hip sdk
Added windows support for hip
cuda_compile->compilecompileLIB_HIP,LIB_HIPRTCconstants togpuctypes/hip.py