[FEA] Add support for mdspan

[MashrafeeAryan]

What this does

Adds two features:

1. Create a MatX tensor from an mdspan

make_tensor can now create a MatX tensor from a cuda::std::mdspan

2. Ability to use tensor[i, j] with C++23

When the compiler supports C++23 multidimensional indexing, tensors can use:

tensor[2, 3] = value;
auto value = tensor[2, 3];

Testing

Added tests for mdspan and operator[]

Built and ran the tensor tests with:

cmake --build build \
  --target test_00_tensor_MakeTensorTests \
  --parallel 2

./build/test/test_00_tensor_MakeTensorTests \
  --gtest_filter='MakeTensorTests.*Mdspan*'

./build/test/test_00_tensor_MakeTensorTests

Testing was done using:

• Google Colab
• Tesla T4 GPU
• CUDA 12.8.93
• NVIDIA driver 580.82.07
• GCC 11.4.0
• CMake 3.31.10
• C++20

The mdspan changes and the full MakeTensorTests suite passed with C++20

The new operator[] feature requires C++23 multidimensional subscript support,
which is not available with CUDA 12.8. I could not update the CUDA environment
due to administrative permission restrictions

I would really appreciate it if the C++23 tests could be run in a compatible
environment during review.

Documentation

Updated the documentation with:

• How to create a MatX tensor from an mdspan
• The new C++23 tensor[i, j] syntax
• Continued support for tensor(i, j) in C++20 and C++23

Related issue

Closes #582

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[greptile-apps]

Greptile Summary

This PR adds zero-copy construction of MatX tensors from supported cuda::std::mdspan layouts and C++23 multidimensional subscript syntax.

• Preserves mdspan extents, strides, data pointer, and non-owning lifetime semantics.
• Restricts conversion to standard layouts and default accessors.
• Adds host/device indexing support, documentation, and focused tensor tests.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
include/matx/core/make_tensor.h	Adds a constrained mdspan conversion that creates a non-owning tensor from the original pointer, extents, and strides while checking representable ranges.
include/matx/core/tensor_impl.h	Adds const and mutable C++23 multidimensional operator[] overloads that delegate to existing tensor indexing.
test/00_tensor/MakeTensorTests.cu	Covers supported mdspan layouts, dynamic and static extents, custom strides, non-owning storage, range rejection, and host/device multidimensional subscripting.
docs_input/api/creation/tensors/make.rst	Documents zero-copy mdspan tensor creation, supported layouts, and borrowed-memory lifetime requirements.
docs_input/quickstart.rst	Documents C++23 multidimensional subscript syntax while retaining C++20-compatible operator() guidance.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  M["cuda::std::mdspan"] --> V["Validate layout, accessor, extents, and strides"]
  V --> D["Create MatX shape/stride descriptor"]
  D --> T["Non-owning MatX tensor"]
  M -. "shared data_handle()" .-> T

Loading

_{Reviews (2): Last reviewed commit: "Only allow supported mdspan layouts" | Re-trigger Greptile}

[MashrafeeAryan]

Hi @cliffburdick, I addressed the issue Greptile pointed out. Please let me know if I need to change anything else. Thank you for your time!

[cliffburdick]

I think we should support both cuda::std::mdspan for those on older c++ versions, then an include guard for the same ones for std::mdspan. Maybe there's a type trait you can use to simplify the code.

[cliffburdick]

I think if we want to support std::mdspan also we should add a type trait that's true for cuda::std::mdspan and std::mdspan. Have that be the only requires clause, then do separate checks inside? It could be a bit ugly but at least it keeps it in one function

[MashrafeeAryan]

Sounds like a good idea. I am working on it!