Is cuDNN/BLAS, MKL, Neon all BLAS compatible?

[wangkuiyi]

If we represent tensors in the same memory footprint, is it correct that we can call cuDNN, MKL, and Neon to operate tensors on various devices?

[wangkuiyi]

OK, one cent is that BLAS just defined some very primitive and frequently used operations. But all these toolkits are much more powerful than BLAS.

[gangliao]

memory layout should be specified if using third-party libs, like MKL, cuDNN, OpenBlas. All of them supports column major, thus that's not a problem.

In Majel, the internal implementation of gemm(Array, ...) also invoked cublas_gemm(Array.ptr()) and MKL's cblas_gemm(Array.ptr()).

@wangkuiyi

Thus, it's definitely compatible.

[wangkuiyi]

Thanks to @gangliao