Lambda Functions for Segment Scan Operations

Table of Contents

• Fetch Lambda
• Reduction Function Object
• Write Lambda
• Condition Lambda

This page describes the lambda function signatures used in segment scan (prefix-sum) operations.

Fetch Lambda

The fetch lambda retrieves the value of an element for the scan operation. It has one of the following forms:

Full form:

auto fetch = [=] __cuda_callable__ ( IndexType segmentIdx, IndexType localIdx, IndexType globalIdx )
{
   // Return the value at the given position
   return ...;
};

Brief form:

auto fetch = [=] __cuda_callable__ ( IndexType globalIdx )
{
   // Return the value at the given position
   return ...;
};

In both variants:

• segmentIdx is the index of the segment.
• localIdx is the rank of the element within the segment.
• globalIdx is the index of the element in the corresponding container.

Reduction Function Object

The scan operation uses a reduction function object to combine values. See Function objects for reduction operations for available reduction operations like:

• TNL::Plus - Addition
• TNL::Multiplies - Multiplication
• TNL::Min - Minimum
• TNL::Max - Maximum

Write Lambda

The write lambda stores the result of the scan operation. It has the following signature:

auto write = [=] __cuda_callable__ ( IndexType globalIdx, ValueType value )
{
   // Write the value at the given position
   output[globalIdx] = value;
};

• globalIdx is the index where the result should be written.
• value is the scan result to write.

Note: The implementation allows in-place scan (the write function may modify the input array).

Condition Lambda

The condition lambda determines which elements should be included in the scan operation. It has the following signature:

auto condition = [=] __cuda_callable__ ( IndexType segmentIdx, IndexType localIdx, IndexType globalIdx ) -> bool
{
   // Return true if element should be included in scan
   return ...;
};

• segmentIdx is the index of the segment.
• localIdx is the rank of the element within the segment.
• globalIdx is the index of the element in the corresponding container.

The lambda should return true if the element should be included in the scan.