CUTLASS: fragment_iterator_complex_tensor_op.h File Reference

| | CUTLASS

CUDA Templates for Linear Algebra Subroutines and Solvers |

• Main Page

• Modules

• Namespaces

• Classes

• Files

• File List

• File Members

• include

• cutlass

• epilogue

• warp

Classes | Namespaces

fragment_iterator_complex_tensor_op.h File Reference

This defines a "fragment" iterator for visiting the fragments of an accumulator tile that participate in one warp-level store operation. More...

#include "cutlass/array.h"
#include "cutlass/layout/matrix.h"
#include "cutlass/epilogue/warp/tensor_op_policy.h"

Include dependency graph for fragment_iterator_complex_tensor_op.h:

This graph shows which files directly or indirectly include this file:

[Go to the source code of this file.](fragment iterator complex tensor op_8h_source.html)

|

Classes

| | class | cutlass::epilogue::warp::FragmentIteratorComplexTensorOp< WarpShape, OperatorShape, OperatorElementC, OperatorFragmentC, Layout > | | | | class | cutlass::epilogue::warp::FragmentIteratorComplexTensorOp< WarpShape_, OperatorShape_, OperatorElementC_, OperatorFragmentC_, layout::RowMajor > | | | Partial specialization for row-major shared memory. More...
| | |

|

Namespaces

| | | cutlass | | | | | cutlass::epilogue | | | | | cutlass::epilogue::warp | | |

Detailed Description

Typically, the accumulator tile is the largest single block of register-backed storage within the kernel. Storing it to memory is best accomplished by partitioning it into smaller tiles and storing these sequentially.

Round trips through shared memory during the Epilogue phase require partitioning, as shared memory capacity is typically insufficient for a threadblock's total accumulator size.

---

Generated by 1.8.11