// Tile size: 6x1
	// Accumulators: 0-5
	// Col regs: 6-11
	// Row regs: 15

    vbroadcastss    ymm15,  dword ptr [rcx]

	vmovups     ymm10, [rax]
	vmulps     ymm10, ymm10, ymm15
	vaddps     ymm0, ymm0, ymm10
    vmovups     ymm11, [rax + 32]
	vmulps     ymm11, ymm11, ymm15
	vaddps     ymm1, ymm1, ymm11
    vmovups     ymm12, [rax + 64]
	vmulps     ymm12, ymm12, ymm15
	vaddps     ymm2, ymm2, ymm12
    vmovups     ymm13, [rax + 96]
	vmulps     ymm13, ymm13, ymm15
	vaddps     ymm3, ymm3, ymm13
    vmovups     ymm14, [rax + 128]
	vmulps     ymm14, ymm14, ymm15
	vaddps     ymm4, ymm4, ymm14
    vmovups     ymm15, [rax + 160]
	vmulps     ymm15, ymm15, ymm15
	vaddps     ymm5, ymm5, ymm15


    add rcx, 4
	add rax, 192