#include <stdlib.h> 

int multiply_buf(void *ptr, void  *with, size_t count)
{
	int i;
	for(i=0;i<count;i++)
	{
	asm volatile ("movupd (%0), %%xmm0\n\t"
		 "mulps (%1), %%xmm0\n\t"
		 "movupd %%xmm0, (%0)"
			:
			: "r" (ptr+4*i), "r" (with));
	}
}

int main()
{
	float with[4] = {0.5, 0.5, 0.5, 0.5};
	float buffer[1048576];
	float f=0.0;
	int i;
	for(i=0;i<1048576;i++)
	{
		buffer[i] = f;
		f += 0.1;
	}
	multiply_buf(buffer, with, 1048576);
	printf("%f %f %f %f", buffer[0], buffer[1], buffer[2], buffer[3]);
}