// Accumulators: 0-9
	// Columns: 14-15
	// Rows: 10-13
    vbroadcastss    ymm10,  dword ptr [rcx]
    vbroadcastss    ymm11,  dword ptr [rcx + 4]
    vbroadcastss    ymm12,  dword ptr [rcx + 8]
    vbroadcastss    ymm13,  dword ptr [rcx + 12]

    vmovaps         ymm14,  [rax]
    vmovaps         ymm15,  [rax + 32]

    vfmadd231ps     ymm0,   ymm14, ymm10
    vfmadd231ps     ymm1,   ymm15, ymm10

    vfmadd231ps     ymm2,   ymm14, ymm11
    vfmadd231ps     ymm3,   ymm15, ymm11

    vbroadcastss    ymm11,  dword ptr [rcx + 16]

    vfmadd231ps     ymm4,   ymm14, ymm12
    vfmadd231ps     ymm5,   ymm15, ymm12

    vfmadd231ps     ymm6,   ymm14, ymm13
    vfmadd231ps     ymm7,   ymm15, ymm13

    vfmadd231ps     ymm8,   ymm14, ymm11
    vfmadd231ps     ymm9,   ymm15, ymm11

    vbroadcastss    ymm10,  dword ptr [rcx + 20]
    vbroadcastss    ymm11,  dword ptr [rcx + 24]
    vbroadcastss    ymm12,  dword ptr [rcx + 28]
    vbroadcastss    ymm13,  dword ptr [rcx + 32]

    vmovaps         ymm14,  [rax + 64]
    vmovaps         ymm15,  [rax + 96]

    vfmadd231ps     ymm0,   ymm14, ymm10
    vfmadd231ps     ymm1,   ymm15, ymm10

    vfmadd231ps     ymm2,   ymm14, ymm11
    vfmadd231ps     ymm3,   ymm15, ymm11

    vbroadcastss    ymm11,  dword ptr [rcx + 36]

    vfmadd231ps     ymm4,   ymm14, ymm12
    vfmadd231ps     ymm5,   ymm15, ymm12

    vfmadd231ps     ymm6,   ymm14, ymm13
    vfmadd231ps     ymm7,   ymm15, ymm13

    vfmadd231ps     ymm8,   ymm14, ymm11
    vfmadd231ps     ymm9,   ymm15, ymm11