IntroductionToThrust.pptVIP

Acknowledgments Mark Harris, Michael Garland, et al. Work on sort, scan, etc. Bastiaan Aarts et al. Enhancements to nvcc * Further Reading Thrust Homepage /p/thrust/ Tutorial /p/thrust/wiki/Tutorial Documentation /p/thrust/wiki/Documentation NVIDIA Research /research CUDA /cuda * We should encourage programmers to think in terms of parallel primitives instead of ad hoc kernels. Many operations can be efficiently implemented with combinations of reduce(), scan(), sort(), etc. * Thrust is open source software with liberal licensing terms. Thrust is a large scale CUDA library with extensive testing. * Thrust mimics the STL and implements many of the same algorithms. When we can’t offer the same exact semantics, we choose a different name. For example, std::accumulate() is similar to thrust::reduce(), but the former guarantees sequential operation while the latter does not. * Thrust provides two containers, host_vector device_vector Like the vector container in the STL, Thrust’s containers provide dynamic array functionality. Copying between host and device is trivial with Thrust. We can access and manipulate device_vector elements directly from the host using standard bracket [] notation. * Thrust containers are compatible with STL containers such as vector, list, map, etc. One can first allocate a Thrust vector and then copy() the data, or use the vector’s constructor to initialize the vector in one step. Note that Thrust handles these operations in an optimized manner. For example, copying a host listT to a device_vectorT is implementing by first copying the listT to a temporary host_vectorT and then copying the host_vectorT to a device_vectorT. The only exception to this is where a STL container is initialized with a device_vector since we can’t intercept this usage. This operation does work, but it will result in one cudaMemcpy() for each element of the sequence, as opposed to a single cudaMemcpy() for the whole vector. * We borrow the STL concept of it