Clang build OpenMP requirement about hptt

[LaurentPlagne]

I agree.
In addition, I am not totally confident about the zero overhead of a single threaded (numThreads=1) OMP version compared to a non OMP one.
Several sequential transpositions could run in concurrent threads...

[springer13]

I have just pushed a version which doesn't rely on openmp anylonger. Indeed, HPTT is now building with clang. However, you have to make sure that you are on an AVX-enabled CPU since the explicit vectorization requies AVX (the makefile currently uses -march=native). I might change the build system to CMAKE in the near future to detect these settings automatically.

If you don't want to run on an AVX-enabled CPU you currently only have the option to build the scalar version ("make scalar").

If you want to compile HPTT on system that doesn't support AVX, you have to make sure that you enable avx (e.g., -mavx).

@LaurentPlagne regarding the overhead, please notice that HPTT also exposes an expert interface via execute_expert<...> which indeed entirely avoid the overhead of spawning threads. Moreover, this interface is especially useful if HPTT is called from within a parallel region where threads have already been spawned.

[LaurentPlagne]

Oops, I am discovering HPTT and I missed the execute_expert<...> API. Thank you for mentioning it !

[springer13]

No worries, HPTT is not (well) documented. I will have to find the time to do this at some point. In the meantime, please feel free to open issues if you run into any problems.

[solomonik]

Now builds with clang for me and runs with CTF without error when building with make scalar. My desktop evidently does not have AVX. Will test on some larger machines eventually that ought to have AVX.

Thanks for your help!