Test mdpsan for handling memory #91
Conversation
- Support n dimensional mdspan as in- and output. - The iteration order is fixed to iterate first over most right dimension and then iterate from the right to the left dimension. - Only single core CPU is supported at the moment. - Some code for developing with cppinsight is left. - Requires the alpaka branch: https://github.com/bernhardmgruber/alpaka/tree/mdspan
|
@bernhardmgruber ping |
There was a problem hiding this comment.
I think you should use an example where an mdspan actually makes sense over a 1D span. For the transform you have here, a 1D span is easier and potentially faster to handle.
Maybe mdspan is generally not a good fit for vikunja, since vikunja's primitives (e.g. transform, scan, for_each) are typically used on 1D sequences. Or in other words: there is no benefit of the n-dimensionality of the data structure.
One example I could think of would be batched matrix multiplication. You would have e.g. three mdspan<dynamic_extent, 4, 4> a, b, c and compute as many individual 4x4 matrix multiplications as the dynamic extent.
| // Copy the data back to the host for validation. | ||
| alpaka::memcpy(queueAcc, hostMem, deviceMem, extent); | ||
|
|
||
| Data resultSum = std::accumulate(hostNativePtr, hostNativePtr + extent.prod(), 0); |
There was a problem hiding this comment.
| Data resultSum = std::accumulate(hostNativePtr, hostNativePtr + extent.prod(), 0); | |
| Data resultSum = std::reduce(hostNativePtr, hostNativePtr + extent.prod()); |
- only for testing purpose - fix bug, which wrote everything to input
|
I think any tensor example would fit extremely well for mdspan. Tensors are THE thing for machine learning. |
I fully agree that |
The PR tests mdspan as alternative for raw pointers to handle the input and output memory.
Benefits of mdspan:
The PR requires the mdspan support, which is not merged yet: alpaka-group/alpaka#1788
Is solving the same problem like in PR #88 with another approach.