• When no placeholder of _2 is passed, degrade to transform_reduce. An existing feature? This is for performance sake, just like partial specializations in c++ std library.
• It's said that make_tuple and get can support only 10 components. Then we can write a more general version using "tuple of tuple".

following original README.

---

CUDA Samples

All samples make use of the runtime API.

vectorAdd

This sample adds two vectors of floats.

zeroCopy

This sample maps device pointers to pinned host memory so that kernels can directly read from and write to pinned host memory.

bandwidthTest

This sample measures host-to-device and device-to-host bandwidth via PCIe for pageable and pinned memory of four transfer sizes of 3KB, 15KB, 15MB and 100MB, and outputs them in CSV format.

checkError

This sample checks the return value of every runtime API.

matrixMul

This sample uses shared memory to accelerate matrix multiplication.

atomicAdd

This sample uses atomic functions, assertions and printf.

asyncEngine

This sample uses asynchronous engines to overlap data transfer and kernel execution.

hyperQ

This sample uses multiple streams to overlap multiple kernel execution, known as the HyperQ technology.

deviceQuery

This sample enumerates the properties of the CUDA devices present in the system. multiDevice

This sample uses cudaSetDevice within a single thread to utilize multiple GPUs.

openmp

This sample uses OpenMP to create multiple CPU threads to utilize multiple GPUs.

mpi

This sample uses MPI to create multiple CPU processes to utilize multiple GPUs.

cublas

This sample uses CUBLAS, a CUDA implementation of BLAS (Basic Linear Algebra Subprograms), for matrix multiplication.

thrust

This sample uses thrust, a CUDA implementation of STL (Standard Template Library), for vector reduction.