level1wendell / memcpy_sse Goto Github PK

I'm sure you'll catch this one, but the copy loop is unrolled 8 times so you need to step the pointers by 1024 bytes and not 128. Should be a nice speedup!

With this change memcpy_sse is almost twice as fast as memcpy (gcc 5.4 on Linux/Nehalem). FWIW, this CPU doesn't exhibit and performance change between x32 and x64.

This is not bug free.

When low 4 bits destination and source addresses are different the function will use _mm_load_si128 to read unaligned memory at the later loop at line 33.
When destination address is unaligned and size is < 16 bytes function will copy 16 bytes and if the pad is > size the size will wrap to 32 ^ 2 - (size - pad) on 32 bit systems and to 64 ^ 2 - (size - pad) on 64.

I tried to compile the code, with the following command, as provided in the readme:

gcc -march=sse3 -O3 -m32 testmem_modified.c -o tm32

but i get this error:

cc1: error: bad value (‘sse3’) for ‘-march=’ switch
cc1: note: valid arguments to ‘-march=’ switch are: i386 i486 i586 pentium lakemont pentium-mmx winchip-c6 winchip2 c3 samuel-2 c3-2 nehemiah c7 esther i686 pentiumpro pentium2 pentium3 pentium3m pentium-m pentium4 pentium4m prescott nocona core2 nehalem corei7 westmere sandybridge corei7-avx ivybridge core-avx-i haswell core-avx2 broadwell skylake skylake-avx512 cannonlake icelake-client icelake-server bonnell atom silvermont slm knl knm geode k6 k6-2 k6-3 athlon athlon-tbird athlon-4 athlon-xp athlon-mp x86-64 eden-x2 nano nano-1000 nano-2000 nano-3000 nano-x2 eden-x4 nano-x4 k8 k8-sse3 opteron opteron-sse3 athlon64 athlon64-sse3 athlon-fx amdfam10 barcelona bdver1 bdver2 bdver3 bdver4 znver1 btver1 btver2 native