gebp: Add new ½ and ¼ packet rows per (peeling) round on the lhs

MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The patch works by altering the gebp lhs packing routines to also
consider ½ and ¼ packet lenght rows when packing, besides the original
whole package and row-by-row attempts. Finally, gebp itself will try
to fit a fraction of a packet at a time if:

i) ½ and/or ¼ packets are available for the current context (e.g. AVX2
   and SSE-sized SIMD register for x86)

ii) The matrix's height is favorable to it (it may be it's too small
    in that dimension to take full advantage of the current/maximum
    packet width or it may be the case that last rows may take
    advantage of smaller packets before gebp goes row-by-row)

This helps mitigate huge slowdowns one had on AVX512 builds when
compared to AVX2 ones, for some dimensions. Gains top at an extra 1x
in throughput. This patch is a complement to changeset 4ad359237aeb519dbd4b55eba43057b37988838c
.

Since packing is changed, Eigen users which would go for very
low-level API usage, like TensorFlow, will have to be adapted to work
fine with the changes.

0 files changed, 0 insertions, 0 deletions