cutlass2.x源码解读之PredicatedTileAccessIteratorParams

原创

于 2025-09-06 17:53:34 发布 · 186 阅读

1 ·

CC 4.0 BY-SA版权

文章标签：

#cuda #cutlass

结构体定义也参数初始化：

struct PredicatedTileAccessIteratorParams {

  using Index = int32_t;
  using LongIndex = int64_t;

  //
  // Data members
  //
  /// stride of pitch-linear layout (units of Element)
  LongIndex stride_;
  /// amount (in byte) to increment pointer to move to next access along
  /// strided dimension
  LongIndex inc_strided_;
  /// amount (in byte) to increment pointer from last access to first access
  /// of next tile
  LongIndex inc_next_;
  /// amount (in byte) to increment pointer from first access of current tile
  /// to first access of next tile
  LongIndex inc_advance_;

  //
  // Methods
  //

  CUTLASS_HOST_DEVICE
  Status initialize(LongIndex stride, PredicatedTileAccessIteratorDesc desc) {

    stride_ = stride;

    inc_strided_ = (LongIndex(stride_) * desc.threadmap_delta.strided()) *
                     desc.element_size_bits / 8;

    if (desc.advance_rank) {
      // advance alon