GitHub - rofl0r/memcpy-test: testing speed of memcpy implementations

Branches Tags

Name		Name	Last commit message	Last commit date
Latest commit History 21 Commits
README		README
amd.c		amd.c
amd.s		amd.s
build.sh		build.sh
compare.pl		compare.pl
compare.py		compare.py
cosmopolitan.s		cosmopolitan.s
dummyfuncs.c		dummyfuncs.c
memcpy-64.s		memcpy-64.s
memcpy_1_i386.s		memcpy_1_i386.s
memcpy_2_i386.s		memcpy_2_i386.s
memcpy_noalign.c		memcpy_noalign.c
memcpy_test.c		memcpy_test.c
musl_0911.c		musl_0911.c
musl_093.c		musl_093.c
musl_1124.c		musl_1124.c
musl_1124.s		musl_1124.s
rep_movsb.c		rep_movsb.c
rofl_avx.c		rofl_avx.c
rofl_ult.c		rofl_ult.c

Repository files navigation

pick one of the memcpy implementations or add a file containing a memcpy function called

`mymemcpy`.

it needs to be called this way in order to keep optimizations/system libc from
interfering.

then run
`CC=gcc OPTS="-O3 -funroll-all-loops" ./compare.py impl1.c [impl2.s ... implX.c]`

(both .s and .c suffixes are supported)

if you use a modern CPU using speedstep, you can get much better results by
putting the CPU into "userspace" governor and then lock the CPU freq to 1 GHz.