Connor Briggs
Connor Briggs
I can add some. Especially for the dumb bugs. Could you explain what you mean by the second bit?
Einsums tensors work very well with NumPy functions, so this is probably not necessary. We may need to work on this for GPU views, though.
Fixes #119
For my piece, I think staying with HDF5 wouldn't be bad, but I would like to get rid of the h5cpp layer and use the official HDF5 C++ wrapper.
I'll look into what needs to be done. I think it may make tiled tensors and block tensors easy to handle, though.
This is in 1.10a1.dev42.
We have thought about it. Just haven’t gotten around to it. I have tried to implement my own version, since it seems that Devin Matthew’s hasn’t updated tblis recently, but...
I added those changes and suggestions
It's the CreatePlan_test. In my testing, it seems that there may be a dependence on the writelock. It seems to fail when the writelock is zero, but succeed when it...
That's what I was thinking. Although, if we're streaming data to the GPU, we could always permute the data as it's being transferred. We wouldn't even need hipTensor. We could...