Make reference ukernels MR/NR sizes configurable

[hominhquan]

• Those reference micro-sizes are hardcoded for some performance purposes
  (#pragma omp simd). They may work most of the time, except when user begins
  tuning the MR/NR sizes.
• This commit exposes them as configurable by user in bli_family_.h,
  makes generation of reference ukernels more robust/consistent. Some explanation
  is added to put future developers on guard (should be added in wiki as well).

[hominhquan]

Update with added doc in docs/ConfigurationHowTo.md (113a1a9)

[devinamatthews]

@hominhquan I'm planning to address this issue within a broader update of the microkernel layer over the next few months. Is having a fix for this issue being merged into master a priority for you?

[hominhquan]

@devinamatthews Yes this fix is useful to me. It may also prevent other people from falling to the issue, at least in the next few months. So yes I would like to see it merged into master.

[devinamatthews]

@fgvanzee what do you think? The other fix of course is to disable the vectorized reference kernel.

[fgvanzee]

@fgvanzee what do you think? The other fix of course is to disable the vectorized reference kernel.

Is disabling the vectorized reference gemm ukernel, as you propose, really just a matter of changing #if 1 to #if 0 on line 37 of ref_kernels/3/bli_gemm_ref.c? If so, and if that fix works for @hominhquan, then I'm somewhat partial to that solution. But if @hominhquan needs the individual MR/NR macro constants for other purposes beyond disabling the vectorization, then I suppose that fix would not work.

[devinamatthews]

@fgvanzee I believe it is that easy.

[hominhquan]

@devinamatthews or @fgvanzee Can you tell if this has been addressed in any ongoing development and should I drop this PR ?

[fgvanzee]

@devinamatthews might be able to comment further, but yes, this is part of ongoing development. We've spoken about Devin's idea as it pertains to this topic and I think it will work for your purposes, too, @hominhquan.

[devinamatthews]

@hominhquan after some pending PRs I'm planning to work on overhauling the way kernels are handled. As part of that, I plan to reintroduce the MR/NR macros so that reference kernels can use them to do compile-time optimizations safely.

[hominhquan]

Ok, so I drop this PR and wait for your coming update.