An Introduction To Thrust.ppt

zip_iterator “zips” two or more sequences together. When the zip_iterator is dereferenced, it returns a tuple() containing (two or more) elements at the same position in their respective sequences. Here the seuqences [A,B,C] and [X,Y,Z] are zipped into a virtual sequence [tuple(A,X), tuple(B,Y), tuple(C,Z)]. This is important because storing the data in separate sequences (so-called “structure of arrays”) is generally better for performance than the alternative (so-called “array of structs”). If we were to store these tuples in an actual array, then when we loaded them from memory we’d suffer a performance penalty due to uncoalesed memory accesses. zip_iterator let’s us imagine that we’re processing a AoS when we’ve really stored things in a SoA. * This example zips two sequences together and then computes a maximum reduction over the virtual tuple sequence. Note that we never need to store the tuples in memory since they are generated on-the-fly. The maximum reduction works because the tuple() data structure defines a comparison operator (specifically operator). * SNRM is the standard (Euclidean) vector norm (AKA the vector 2-norm) for an array of floats The fused implementation reads the array once while the un-fused implementation performs 2 reads and 1 write per element (3 operations vs. 1 for the fused version) * * This is an example of kernel fusion, where the square() transformation is fused with the reduction operation. Although we could have implemented this with transform_iterator, we use the transform_reduce function for convenience. Note that transform_reduce is implemented with transform_iterator and reduce(). Note that this implementation provides excellent performance without much effort. * Fancy Iterators zip_iterator Looks like an array of structs (AoS) Stored in structure of arrays (SoA) * A B C X Y Z A B C X Y Z Fancy Iterators zip_iterator * // initialize vectors device_vectorint A(3); device_vectorchar B(3); A[0] = 10; A[1] = 20; A[2] = 30;