You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: README.md
+2-7Lines changed: 2 additions & 7 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -50,23 +50,18 @@ Read README-build.md
50
50
51
51
## How to Use
52
52
53
-
Keep it mind... it is very very initial version that misses a lot of functionality and it isn't fully tested yet.
53
+
Keep it mind... it is earky version that misses a lot of functionality and it isn't fully tested yet.
54
54
So if something fails. It is either not implemented or it is implemented incorrectly
55
55
56
56
Note: pytorch backend is based on dlprimitives library that actually implements all the operators and
57
57
it is relatively well tested.
58
58
59
-
If you still want to try:
59
+
If you still want to try: import package `pytorch_ocl`
60
60
61
-
Import package `pytorch_ocl`
62
-
63
-
torch.ops.load_library("/path/to/libpt_ocl.so")
64
-
65
61
Keep in mind you may have several OpenCL devices. Refer to `clinfo --list` to list
66
62
of the devices and their order. Now instead of calling `something.to('cuda')` you call `something.to('ocl:0')` or
67
63
`something.to('privateuseone:0' for pytorch 1.13)` or another `ocl:1` etc.
68
64
69
-
70
65
## Known Issues
71
66
72
67
1. Many operators not implemented and there may be fallbacks to CPU. Sometimes it is minor but sometimes it may hamper the performance, some may just fail
0 commit comments