Two benchmarked memory copy algorithms

[Steve Hutchesson]

There are two different memory copy algorithms here, one is a simple to use byte aligned copy that is more than fast enough in most instances, the second is an aligned SSE2 copy that is faster in most instances but requires aligned memory both in and out. The rep movsb code is general purpose copy, the SSE2 version is where you are after the fastest you can run for tasks like streaming and other demanding tasks.

' ¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤

    #include "\basic\include\win32api.inc"

    MACRO memsize = 1024*1024*768

' ¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤

 FUNCTION PBmain as LONG

    LOCAL psrc as DWORD
    LOCAL pdst as DWORD
    LOCAL tcnt as DWORD
    LOCAL lcnt as DWORD

    psrc = GlobalAlloc(%GMEM_FIXED or %GMEM_ZEROINIT,memsize)
    pdst = GlobalAlloc(%GMEM_FIXED or %GMEM_ZEROINIT,memsize)

    StdOut $CRLF+"Running timing for XMM copy and REP MOVSB copy"+$CRLF

    SetPriorityClass GetCurrentProcess,%HIGH_PRIORITY_CLASS

    xmmcopya psrc,pdst,memsize      ' dummy run to defeat cache effects
    bcopy psrc,pdst,memsize         ' dummy run to defeat cache effects

  ' -----------------------------

    ! cpuid
    SleepEx 100,0

    tcnt = GetTickCount

    lcnt = 20

  lbl1:
    xmmcopya psrc,pdst,memsize
    ! sub lcnt, 1
    ! jnz lbl1

    tcnt = GetTickCount - tcnt

    StdOut "XMM Copy "+format$(tcnt)+" ms"

  ' -----------------------------

    ! cpuid
    SleepEx 100,0

    tcnt = GetTickCount

    lcnt = 20

  lbl2:
    bcopy psrc,pdst,memsize
    ! sub lcnt, 1
    ! jnz lbl2


    tcnt = GetTickCount - tcnt

    StdOut "rep movsb copy "+format$(tcnt)+" ms"

  ' -----------------------------

    SetPriorityClass GetCurrentProcess,%NORMAL_PRIORITY_CLASS

    GlobalFree psrc
    GlobalFree pdst

    waitkey$

 End FUNCTION

' ¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤

FUNCTION xmmcopya(ByVal src as DWORD, ByVal dst as DWORD, ByVal cnt as DWORD) as DWORD
  ' ------------------------
  ' SSE2 aligned memory copy
  ' ------------------------
    #REGISTER NONE

    PREFIX "!"

    mov esi, src
    mov edi, dst
    mov ecx, cnt
    cmp ecx, 16
    jbe bypass
    shr ecx, 4
    xor ebx, ebx

  lpst:
    movdqa xmm0, [esi+ebx]          ' aligned read
    movntdq [edi+ebx], xmm0         ' aligned write
    add ebx, 16
    sub ecx, 1
    jnz lpst

    mov eax, cnt
    and eax, 15
    test eax, eax
    jnz lbl
    jmp bye

  lbl:
    mov dl, [esi+ebx]
    mov [edi+ebx], dl
    add ebx, 1
    sub eax, 1
    jnz lbl
    jmp bye

  bypass:
    rep movsb

  bye:
    END PREFIX

END FUNCTION

' ¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤

FUNCTION bcopy(ByVal src as DWORD, ByVal dst as DWORD, ByVal cnt as DWORD) as DWORD
  ' -------------------
  ' unaligned byte copy
  ' -------------------
    #REGISTER NONE

    PREFIX "!"

    mov esi, src
    mov edi, dst
    mov ecx, cnt
    rep movsb

    END PREFIX

END FUNCTION

' ¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤

[Tim Lakinir]

Thank you Sir Steve, the results of 3 runs on my laptop are

XMM copy            -->    2234  ms   , 2250  ms , 2125 ms

Rep movsb copy  -->    2578 ms ,    2547  ms , 2547 ms

XMM copy is much faster