// Accumulators: 0-7
	// Columns: 14-15
	// Rows: 8-13


    vbroadcastss    ymm15,  dword ptr [rcx]
    vbroadcastss    ymm14,  dword ptr [rcx + 4]

    vmovaps     ymm8, [rax]
    vmovaps     ymm9, [rax + 32]
    vmovaps     ymm10, [rax + 64]
    vmovaps     ymm11, [rax + 96]
    vmovaps     ymm12, [rax + 128]
    vmovaps     ymm13, [rax + 160]

    vfmadd231ps     ymm0, ymm15, ymm8
    vfmadd231ps     ymm1, ymm15, ymm9
    vfmadd231ps     ymm2, ymm15, ymm10
    vfmadd231ps     ymm3, ymm15, ymm11
    vfmadd231ps     ymm4, ymm15, ymm12
    vfmadd231ps     ymm5, ymm15, ymm13

	vmovaps     ymm8, [rax + 192]
    vmovaps     ymm9, [rax + 224]
    vmovaps     ymm10, [rax + 256]
    vmovaps     ymm11, [rax + 288]
    vmovaps     ymm12, [rax + 320]
    vmovaps     ymm13, [rax + 352]

    vfmadd231ps     ymm6, ymm15, ymm8
    vfmadd231ps     ymm7, ymm15, ymm9
    vfmadd231ps     ymm0, ymm14, ymm10
    vfmadd231ps     ymm1, ymm14, ymm11
    vfmadd231ps     ymm2, ymm14, ymm12
    vfmadd231ps     ymm3, ymm14, ymm13

    vmovaps     ymm8, [rax + 384]
    vmovaps     ymm9, [rax + 416]
    vmovaps     ymm10, [rax + 448]
    vmovaps     ymm11, [rax + 480]

    vfmadd231ps     ymm4, ymm14, ymm8
    vfmadd231ps     ymm5, ymm14, ymm9
    vfmadd231ps     ymm6, ymm14, ymm10
    vfmadd231ps     ymm7, ymm14, ymm11

    add rcx, 8
	add rax, 512