kAlignmentA Class — pytorch Architecture

Source Code

aten/src/ATen/native/cuda/cutlass_extensions/gemm/threadblock/default_dq_mma.h lines 61–102

template<
    /// Element type for A matrix operand
    typename ElementA_,
    /// Layout type for A matrix operand
    typename LayoutA_,
    /// Access granularity of A matrix in units of elements
    int kAlignmentA,
    /// Element type for B matrix operand
    typename ElementB_,
    /// Layout type for B matrix operand
    typename LayoutB_,
    /// Access granularity of B matrix in units of elements
    int kAlignmentB,
    /// Element type for the input scale
    typename ElementScale_,
    /// Layout for the scale operand
    typename LayoutScale_,
    /// Access granularity of Scales in unit of elements
    int kAlignmentScale,
    /// Element type for internal accumulation
    typename ElementAccumulator_,
    /// Layout type for C and D matrix operands
    typename LayoutC_,
    /// Operator class tag
    typename OperatorClass_,
    /// Tag indicating architecture to tune for
    typename ArchTag_,
    /// Threadblock-level tile size (concept: GemmShape)
    typename ThreadblockShape_,
    /// Warp-level tile size (concept: GemmShape)
    typename WarpShape_,
    /// Instruction-level tile size (concept: GemmShape)
    typename InstructionShape_,
    /// Number of stages used in the pipelined mainloop
    int Stages,
    /// Operation performed by GEMM
    typename Operator_,
    /// Use zfill or predicate for out-of-bound cp.async
    SharedMemoryClearOption SharedMemoryClear = SharedMemoryClearOption::kNone,
    ///
    typename Enable = void>
struct DqMma;