Andreas Borgen Longva

You bring up some very good points! The problem of uninitialized input is difficult to resolve within the confines of a safe API. As you say, we can write an...

If I understand you correctly, you want to construct a matrix from uninitialized data and then pass it into the function. I assume that you intend it to be called...

(This is an answer only to the question of uninitialized data) I think it should be possible to encapsulate the possibility of uninitialized data in a slightly more safe manner....

> I meant specifically the Float trait bound was dangerous. Really we only want these functions to work with f32 and f64 (or maybe complex types in the future). But...

> However, I'm not convinced we can allow the user to safely construct an UninitializedBuffer. After all - this constructor is really just an abstraction for unsafe { vec.set_len(n*m) }....

> I think that this could work actually. At the very least we should try this idea out and see if it gives us what we want with the additional...

Thanks for your input, @bluss! Did you have the impression that overhead for small matrices is a fundamental restriction of the architecture of BLIS, or that it is/was an implementation...

> I worry that we lose some functionality this way. For example what if a user has their own custom type which implements Float. We would like our implementations to...

That was a very interesting talk, @bluss! I've never really had the need to work with tensors so it's not my field of expertise, but I quite enjoyed to see...

Thanks for the PR @vks! Yes, I agree that the original error message is indeed misleading. Note that the documentation for the method also needs to be updated as it...