tree 822a761ad5cefd55078e3b148f79fba8fcbc4492
parent 1031e0da388ad42b8c33dedb8e6243b9fdab9b15
author ahs <amrita.h.s@intel.com> 1654825274 +0530
committer Elliott Hughes <enh@google.com> 1658526530 +0000

avx2 implementation for memset.

This patch includes handwritten avx2
assembly for memset 64-bit. Uses
non-temporal stores for very large sizes.
Also includes dynamic dispatch for APIs
having multiple implementations.

Convincing benchmark improvements for sizes above 512 bytes, and
although the slight regression for small sizes is unfortunate, it's
probably small enough to be okay?

Before:

  BM_string_memset/8/0            3.06 ns         3.04 ns    222703428 bytes_per_second=2.45261G/s
  BM_string_memset/16/0           3.50 ns         3.47 ns    202569932 bytes_per_second=4.29686G/s
  BM_string_memset/32/0           3.50 ns         3.48 ns    200064955 bytes_per_second=8.57386G/s
  BM_string_memset/64/0           3.49 ns         3.46 ns    201928186 bytes_per_second=17.2184G/s
  BM_string_memset/512/0          14.8 ns         14.7 ns     47776178 bytes_per_second=32.3887G/s
  BM_string_memset/1024/0         27.3 ns         27.1 ns     25884933 bytes_per_second=35.2515G/s
  BM_string_memset/8192/0          203 ns          201 ns      3476903 bytes_per_second=37.9311G/s
  BM_string_memset/16384/0         402 ns          399 ns      1750471 bytes_per_second=38.2725G/s
  BM_string_memset/32768/0         932 ns          925 ns       755750 bytes_per_second=33.0071G/s
  BM_string_memset/65536/0        2038 ns         2014 ns       347060 bytes_per_second=30.3057G/s
  BM_string_memset/131072/0       4012 ns         3980 ns       175186 bytes_per_second=30.6682G/s

After:

  BM_string_memset/8/0            3.32 ns         3.23 ns    208939089 bytes_per_second=2.3051G/s
  BM_string_memset/16/0           4.07 ns         3.98 ns    173479615 bytes_per_second=3.74822G/s
  BM_string_memset/32/0           4.07 ns         3.95 ns    177208119 bytes_per_second=7.54344G/s
  BM_string_memset/64/0           4.09 ns         4.00 ns    174729144 bytes_per_second=14.8878G/s
  BM_string_memset/512/0          10.7 ns         10.4 ns     65922763 bytes_per_second=45.6611G/s
  BM_string_memset/1024/0         18.0 ns         17.6 ns     40489136 bytes_per_second=54.3166G/s
  BM_string_memset/8192/0          109 ns          106 ns      6577711 bytes_per_second=71.7667G/s
  BM_string_memset/16384/0         221 ns          210 ns      3343800 bytes_per_second=72.684G/s
  BM_string_memset/32768/0         655 ns          623 ns      1153501 bytes_per_second=48.9781G/s
  BM_string_memset/65536/0        1547 ns         1495 ns       461702 bytes_per_second=40.8154G/s
  BM_string_memset/131072/0       2991 ns         2924 ns       240189 bytes_per_second=41.7438G/s

This patch drops the wmemset() code because we don't even have a
microbenchmark for it, we have as many implementations checked in as we
have non-test call sites (!), so at this point it seems like we've spent
more time maintaining wmemset() than running it!

Test: bionic/tests/run-on-host.sh 64
Signed-off-by: ahs <amrita.h.s@intel.com>
Change-Id: Ie5047df5300638c1e4c69f8285d33d034f79c83b