// Tile size: 2x6
	// Accumulators: 0-11
	// Col regs: ymm14-15
	// Row regs: ymm12-13

	vbroadcastss	ymm14,	dword ptr [rcx]
	vmovaps			ymm12,	[rax]
	vmovaps			ymm13,	[rax + 32]
	vbroadcastss	ymm15,	dword ptr [rcx + 4]

	vfmadd231ps		ymm0,	ymm12, ymm14
	vfmadd231ps		ymm1,	ymm13, ymm14

	vbroadcastss	ymm14,	dword ptr [rcx + 8]

	vfmadd231ps		ymm2,	ymm12, ymm15
	vfmadd231ps		ymm3,	ymm13, ymm15

	vbroadcastss	ymm15,	dword ptr [rcx + 12]

	vfmadd231ps		ymm4,	ymm12, ymm14
	vfmadd231ps		ymm5,	ymm13, ymm14

	vbroadcastss	ymm14,	dword ptr [rcx + 16]

	vfmadd231ps		ymm6,	ymm12, ymm15
	vfmadd231ps		ymm7,	ymm13, ymm15

	vbroadcastss	ymm15,	dword ptr [rcx + 20]

	vfmadd231ps		ymm8,	ymm12, ymm14
	vfmadd231ps		ymm9,	ymm13, ymm14

	vbroadcastss	ymm14,	dword ptr [rcx+24]

	vfmadd231ps		ymm10,	 ymm12, ymm15
	vfmadd231ps		ymm11,	 ymm13, ymm15

	// Iteration two
	vmovaps			ymm12,	[rax + 64]
	vmovaps			ymm13,	[rax + 96]
	vbroadcastss	ymm15,	dword ptr [rcx + 24 + 4]

	vfmadd231ps		ymm0,	ymm12, ymm14
	vfmadd231ps		ymm1,	ymm13, ymm14

	vbroadcastss	ymm14,	dword ptr [rcx + 24 + 8]

	vfmadd231ps		ymm2,	ymm12, ymm15
	vfmadd231ps		ymm3,	ymm13, ymm15

	vbroadcastss	ymm15,	dword ptr [rcx + 24 + 12]

	vfmadd231ps		ymm4,	ymm12, ymm14
	vfmadd231ps		ymm5,	ymm13, ymm14

	vbroadcastss	ymm14,	dword ptr [rcx + 24 + 16]

	vfmadd231ps		ymm6,	ymm12, ymm15
	vfmadd231ps		ymm7,	ymm13, ymm15

	vbroadcastss	ymm15,	dword ptr [rcx + 24 + 20]

	vfmadd231ps		ymm8,	ymm12, ymm14
	vfmadd231ps		ymm9,	ymm13, ymm14

	vfmadd231ps		ymm10,	 ymm12, ymm15
	vfmadd231ps		ymm11,	 ymm13, ymm15

	add rax, 128
	add rcx, 48