Playing around with GPU kernel optimization.

Current Functions:

copy -- Simple copy as benchmark.

saxpy -- Single-precision ax+y; benchmark.

transposeNaive -- Transposes a matrix out-of-place naively. One element per thread, all memory accesses global, some not coalesced.

transposeFast -- Transposes a matrix using shared memory tiles as an intermediate step to coalesce gmem access.

transposeFastNoBankConf -- Same as transposeFast, but eliminates bank conflicts in smem tiles.

matxmatNaive -- Naive matrix multiply. Each thread handles one element of output array; all memory accesses are global, some not coalesced.

matxmatTiles -- Matrix multiply using shared memory tiles to reduce total number of gmem accesses and ensure that all gmem accesses are coalesced. Smem bank conflicts are also eliminated.

reduceY -- Reduces (sums) columns of matrix into top row.

scan -- Parallel prefix sum capable of handling arbitrarily large data arrays, based on GPU Gems 3 ch. 39. Scans input array within blocks, stores their sums, recursively scans array of sums, then adds elements of the scanned sums array to appropriate regions of the output array.