// slow
	vbroadcastss xmm16, dword ptr [rcx]
	vbroadcastss xmm17, dword ptr [rcx + 4]
	vbroadcastss xmm18, dword ptr [rcx + 8]
	vbroadcastss xmm19, dword ptr [rcx + 12]

	// fast
	vmovups	   		xmm31, [rcx]
	vbroadcastss 	zmm16, xmm31
	valignd 		xmm17, xmm31, xmm31, 1
	vbroadcastss 	zmm17, xmm17
	valignd 		xmm18, xmm31, xmm31, 2
	vbroadcastss 	zmm18, xmm18
	valignd 		xmm19, xmm31, xmm31, 3
	vbroadcastss 	zmm19, xmm19

	// commmon
	vfmadd231ps		zmm0, zmm16, [rax + 0]
	vfmadd231ps		zmm1, zmm17, [rax + 64]
	vfmadd231ps		zmm2, zmm18, [rax + 128]
	vfmadd231ps		zmm3, zmm19, [rax + 192]

	add rcx, 16
	add rax, 256