Onednn: performance comparison for refcode vs optimised code

[uu-praveeng]

Hi,

Is there a way to build the oneDNN library on Intel platforms disabling the optimization kernels. So that library only includes ref c implementations of all the layers.

We want to do performance analysis, how much we have gained by enabling optimizations wrt ref C implementation of Layers code.

Thanks,
Praveen.

[nSircombe]

Hi @uu-praveeng,

As it happens, there is! @diaena was trying to to the same thing just recently - https://github.com/oneapi-src/oneDNN/issues/957

I think you need to set DNNL_TARGET_ARCH=ARCH_GENERIC.

[sergiomatiz1986]

Hi @nSircombe,

I was wondering if DNNL_TARGET_ARCH=ARCH_GENERIC would also disable BLAS. In my application, I would like to benchmark the performance of MKL-DNN without the use of SIMD (but without deactivating BLAS/GEMM for instance)

[nSircombe]

As far as I'm aware, if you've chosen to build and link with an external BLAS lib, this will still be the case if you choose ARCH_GENERIC. But you might want to double-check with an ldd of the final .so?