Hands-On GPU Computing with Python

What is reduction?

Reduction is an optimization strategy that is an expansion of the global kernel. With reduction, you can use multiple thread blocks and decompose a computation into multiple kernel invocations. So, reduction maximizes parallelization and lowers computation times by a huge extent.

To achieve reduction in CUDA-C/C++, you need to expand the __global__ kernel with multiple lines of code.

With PyCUDA, you can perform reduction in a single line by importing the reduction module. It is known as ReductionKernel and can be implemented with GPUArrays. You also have the option to use actual CUDA C/C++ reduction within the __global__ kernel itself, on PyCUDA.

We will revisit this perspective again in the next chapter when we compare ...