Optimize gemv_n_sve kernel

[manaalmj]

Loop-unrolling of sgemv_n kernel is implemented with svmla_lane, along with a main loop and a tail loop.
The graph below shows performance improvement for OMP_NUM_THREADS=1 on Graviton-3:

[fadara01]

Can we also update the plot s.t:

• it's discontinuous - don't connect the points to each other
• showing clearly what the current state of affairs is vs this PR - maybe a red horizontal line at speedup=1
• It might also make sense to use log scale for the Y-axis.
• X-axis label is MxKxN, right?
• add the number of threads/machine to the title

[martin-frbg]

yes, an explanation of the benchmark graph would be nice - I'm not sure I understand why your PR would result in two very narrow regions where there is more than tenfold speedup ?

[manaalmj]

yes, an explanation of the benchmark graph would be nice - I'm not sure I understand why your PR would result in two very narrow regions where there is more than tenfold speedup ?

Please refer to the updated graph. For the regions where we see spikes, The baseline performance of these problem sizes are poor and that's why we see spike in speedup.

[manaalmj]

Can we also update the plot s.t:

• it's discontinuous - don't connect the points to each other
• showing clearly what the current state of affairs is vs this PR - maybe a red horizontal line at speedup=1
• It might also make sense to use log scale for the Y-axis.
• X-axis label is MxKxN, right?
• add the number of threads/machine to the title

Please see the updated graph.

[fadara01]

@martin-frbg could you please re-review?

[martin-frbg]

Looks good to me, thank you very much.