I've had a little look at it, and I think LibRapid aims to solve a different problem. LibRapid aims to be as fast as physically possible (where reasonably possible to do so), and hence is able to achieve performance at or above libraries like XTensor and Eigen.
It also has seamless CUDA integration, allowing you to perform operations on the GPU by simply changing the tempate argument to your array definitions. For example, `auto cpuArray = lrc::Array<float>(...);` and `auto gpuArray = lrc::Array<float, lrc::device::GPU>(...)`.
It appears to me that TensorStore's primary focus is storing the arrays, not providing high-performance routines to manipulate them.
Regarding the features you mentioned, LibRapid provides an ArrayView class, which provides strided read and write access to any Array type. I'm also planning on impementing a method to save and load arrays from files, so it might be possible to implement some form of compression into that process.
I am the lead developer of LibRapid, a high-performance array and mathematics library with some additional utility functions.
LibRapid's array class is incredibly intuitive to use while simultaneously achieving performance at or above leading libraries such as Eigen and XTensor. The array library has full support for CUDA, with JIT-compiled kernels and support for cuBLAS and cuFFT. The array library integrates seamlessly with the rest of LibRapid's mathematics functionality. Almost all functions are overloaded for primitive types, multi-precision types, vectors and arrays, enabling a single function to operate on the whole LibRapid ecosystem.
LibRapid has a powerful CMake script capable of finding and linking many popular BLAS libraries, such as OpenBLAS and Intel's OneAPI Math Kernel Library, as well as detecting CUDA support and MPIR and MPFR libraries. Furthermore, the script provides many configuration options to fine-tune LibRapid to your system, further increasing the library's performance.
As mentioned earlier, LibRapid also fully supports multi-precision arithmetic using MPIR and MPFR. If you choose to enable multi-precision (via a CMake option), the CMake script will first search for the relevant libraries already on your system. If they cannot be found, LibRapid will use custom, modified versions of MPIR and MPFR.
LibRapid's vector library is also incredibly powerful, supporting SIMD operations, many functions and an intuitive interface.
In the future, we plan to develop a machine-learning library on top of LibRapid's mathematics and array functionality.
Hopefully, this gives you a brief idea of what LibRapid is capable of. If you have any questions or want to try LibRapid out for yourself, have a look at the links below.
It also has seamless CUDA integration, allowing you to perform operations on the GPU by simply changing the tempate argument to your array definitions. For example, `auto cpuArray = lrc::Array<float>(...);` and `auto gpuArray = lrc::Array<float, lrc::device::GPU>(...)`.
It appears to me that TensorStore's primary focus is storing the arrays, not providing high-performance routines to manipulate them.
Regarding the features you mentioned, LibRapid provides an ArrayView class, which provides strided read and write access to any Array type. I'm also planning on impementing a method to save and load arrays from files, so it might be possible to implement some form of compression into that process.
Hopefully this helps!