Bench.dpr

program Benchmark;

{$APPTYPE CONSOLE}

uses
  Windows,
  SysUtils,
  FastMove;

const
  ITERATIONS = 1000000;
var
  Buffer: array[0..128 shl 20] of Byte; // should be enough to not overwrite in memory... maybe? :P

  function BenchFastCode(AForward: Boolean; N: Integer): Extended;
  var
    qs, qe, qf: Int64;
    i: Integer;
  begin
    QueryPerformanceCounter(qs);
    if AForward then
    begin
      for i := 1 to ITERATIONS do
        FastMove.Move(Buffer[i], Buffer[i + N], N);
    end
    else
    begin
      for i := 1 to ITERATIONS do
        FastMove.Move(Buffer[i + N], Buffer[i], N);
    end;
    QueryPerformanceCounter(qe);
    QueryPerformanceFrequency(qf);
    Result := (qe - qs) * 1e9 / (qf * ITERATIONS);
  end;

const
  at: array[0..25] of Integer = (1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 13, 16, 17, 29, 31, 32, 45, 54, 62, 79, 93, 121, 128, 4096, 4096*4, 4096*16);
var
  i: Integer;
begin
  for i := Low(at) to High(at) do
    Writeln(at[i]: 10, BenchFastCode(True, at[i]): 10: 3, BenchFastCode(False, at[i]): 10: 3);
  readln;
end.

main.cpp

#include <windows.h>
#include <stdio.h>
#include "stdint.h"
#include "apex_memmove.h"
#include "apex_memmove.c"

#define ITERATIONS 1000000

uint8_t Buffer[128<<20];

double runbench(bool forward, size_t n){
	LARGE_INTEGER qs, qe, qf;
	QueryPerformanceCounter(&qs);
	if(forward){
		for(size_t i = 0; i < ITERATIONS; i++){
			apex::memmove(&Buffer[i+n], &Buffer[i], n);
		}
	} else {
		for(size_t i = 0; i < ITERATIONS; i++){
			apex::memmove(&Buffer[i], &Buffer[i+n], n);
		}
	}
	QueryPerformanceCounter(&qe);
	QueryPerformanceFrequency(&qf);

	return (qe.QuadPart - qs.QuadPart) * 1e9 / (qf.QuadPart * ITERATIONS);
}

double runbenchstd(bool forward, size_t n){
	LARGE_INTEGER qs, qe, qf;
	QueryPerformanceCounter(&qs);
	if(forward){
		for(size_t i = 0; i < ITERATIONS; i++){
			memmove(&Buffer[i+n], &Buffer[i], n);
		}
	} else {
		for(size_t i = 0; i < ITERATIONS; i++){
			memmove(&Buffer[i], &Buffer[i+n], n);
		}
	}
	QueryPerformanceCounter(&qe);
	QueryPerformanceFrequency(&qf);

	return (qe.QuadPart - qs.QuadPart) * 1e9 / (qf.QuadPart * ITERATIONS);
}

int main() {
	int at[26] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 13, 16, 17, 29, 31, 32, 45, 54, 62, 79, 93, 121, 128, 4096, 4096*4, 4096*16};
	for(size_t i = 0; i < 26; i++){
		printf("%10d %10.3f %10.3f\n", at[i], runbench(true, at[i]), runbench(false, at[i]));
	}
	for(size_t i = 0; i < 26; i++){
		printf("%10d %10.3f %10.3f\n", at[i], runbenchstd(true, at[i]), runbenchstd(false, at[i]));
	}
	return 0;
}