// Tile size: 1x8
	// Accumulators: 0-7
	// Col regs: 8-14
	// Row regs: 15


    vmovaps         ymm15,  [rax]

    vbroadcastss    ymm8, dword ptr [rcx + 0 * 4]
    vfmadd231ps     ymm0, ymm15, ymm8

    vbroadcastss    ymm9, dword ptr [rcx + 1 * 4]
    vfmadd231ps     ymm1, ymm15, ymm9

    vbroadcastss    ymm10, dword ptr [rcx + 2 * 4]
    vfmadd231ps     ymm2, ymm15, ymm10

    vbroadcastss    ymm11, dword ptr [rcx + 3 * 4]
    vfmadd231ps     ymm3, ymm15, ymm11

    vbroadcastss    ymm12, dword ptr [rcx + 4 * 4]
    vfmadd231ps     ymm4, ymm15, ymm12

    vbroadcastss    ymm13, dword ptr [rcx + 5 * 4]
    vfmadd231ps     ymm5, ymm15, ymm13

    vbroadcastss    ymm10, dword ptr [rcx + 6 * 4]
    vfmadd231ps     ymm6, ymm15, ymm10

    vbroadcastss    ymm11, dword ptr [rcx + 7 * 4]
    vfmadd231ps     ymm7, ymm15, ymm11


    vmovaps         ymm15,  [rax]

    vbroadcastss    ymm8, dword ptr [rcx + 0 * 4]
    vfmadd231ps     ymm0, ymm15, ymm8

    vbroadcastss    ymm9, dword ptr [rcx + 1 * 4]
    vfmadd231ps     ymm1, ymm15, ymm9

    vbroadcastss    ymm10, dword ptr [rcx + 2 * 4]
    vfmadd231ps     ymm2, ymm15, ymm10

    vbroadcastss    ymm11, dword ptr [rcx + 3 * 4]
    vfmadd231ps     ymm3, ymm15, ymm11

    vbroadcastss    ymm12, dword ptr [rcx + 4 * 4]
    vfmadd231ps     ymm4, ymm15, ymm12

    vbroadcastss    ymm13, dword ptr [rcx + 5 * 4]
    vfmadd231ps     ymm5, ymm15, ymm13

    vbroadcastss    ymm10, dword ptr [rcx + 6 * 4]
    vfmadd231ps     ymm6, ymm15, ymm10

    vbroadcastss    ymm11, dword ptr [rcx + 7 * 4]
    vfmadd231ps     ymm7, ymm15, ymm11