-
-
Notifications
You must be signed in to change notification settings - Fork 56.5k
ocl4dnn convolution - division by zero #20833
Copy link
Copy link
Open
Labels
bugcategory: dnnfeatureincompleteneeds reproducerProvide complete minimal reproducer with input dataProvide complete minimal reproducer with input data
Description
I'm getting some weird crashes when running a simple Resnet model on a GTX 3080Ti in OPENCL backend.
The crash is a division by zero in:
| M_ = num_output_ / group_; |
group_ is computed here| config.group = inputs[0].size[1] / umat_blobs[0].size[1]; |
In my case
inputs[0].size[1] equal 1 and umat_blobs[0].size[1] equals 64
It is going to be a little difficult to provide a simple program reproducing this but maybe someone can already spot an obvious bug.
It seems that it will only crash once, the next run will use the cached kernels and work fine (group_=1)
Reactions are currently unavailable
Metadata
Metadata
Assignees
Labels
bugcategory: dnnfeatureincompleteneeds reproducerProvide complete minimal reproducer with input dataProvide complete minimal reproducer with input data