Justin Luitjens
Justin Luitjens
the function GetFFTInputView will not compile for non-owning types. https://github.com/NVIDIA/MatX/blob/main/include/matx_fft.h#L706-L738 The issue is that owning/non-owning is part of the type. The output of this function will return the same owning...
**Describe the bug** If you enable this line in the unit tests the unit tests will show memory errors in compute sanitizer. In particular it will try to write to...
we should write a simple kernel that dispatches 1 sort batch per CTA using cub when the sorting dimension is small. This would increase throughput and lower latency for batched...
The goal here is to allow people to create tensors on the fly without syncing. These tensors could not use managed memory at this time. API example: auto t =...
Convolution needs more unit tests for both 1D and 2D. There are things that look like they are probably bugs but are not currently covered by unit tests. For example:...
The deriv operator will return the first derviative with respect to "X" of the underlying operation. The derivative of a constant (or tensor) will return 0. The wrt operator signifies...
Random Tensor APIs were some of our first APIs. We have moved away from that style of API and need to update. We should depericate/delete tensorShape_t references. We should get...
https://www.mathworks.com/help/images/ref/padarray.html We should add a pad operator which functionally changes the size of a tensor and set's how out of bounds accesses get values. We should fully support padding pre,...
Currently our cub sorting code calls a few members which require tensor inputs. They include IsContigious() and Data(). We should implement a path that takes operators.
We should support a user getting just the tensor class from matx by including for example /matx/core/matx_tensor.h without pulling in all the operators/transforms/etc.