SIMD implementation of dot-product for complex numbers (sum(a_i * conj(b_i])))

complexdotproduct.d

	for(; ap < sbe; ap+=4*M, bp+=4*M)
	{
		auto a0 = load(cast(VP)ap+0);
		auto a1 = load(cast(VP)ap+1);
		auto a2 = load(cast(VP)ap+2);
		auto a3 = load(cast(VP)ap+3);

		auto b0 = load(cast(VP)bp+0);
		auto b1 = load(cast(VP)bp+1);
		auto b2 = load(cast(VP)bp+2);
		auto b3 = load(cast(VP)bp+3);

		s0 += a0 * b0;
		s1 += a1 * b1;
		s2 += a2 * b2;
		s3 += a3 * b3;

		//swap real and imaginary parts
		a0 = swapReIm(a0);
		a1 = swapReIm(a1);
		a2 = swapReIm(a2);
		a3 = swapReIm(a3);

		r0 += a0 * b0;
		r1 += a1 * b1;
		r2 += a2 * b2;
		r3 += a3 * b3;		
	}