A novel efficient inplace, multithreaded, and cache-friendly parallel 2-D wavelet transform algorithm based on the lifting transform is introduced. In order to maximize the cache utilization and consequently minimize the memory bus bandwidth use, the threads compete to work on a small memory area maximizing the chance of finding it in the cache and their synchronization is done with very low overhead without the use of any locks and relying solely on the basic compare-and-swap (CAS) atomic primitive. An implementation in the C programming language with and without the use of vector (single instruction multiple data - SIMD) instructions is provided for both single (serial) and multi (parallel) threaded single-loop DWT implementations as well as serial and parallel naive implementations using linear (row order) and strided (column order) memory access patterns for comparison. Results show a significant improvement over the single-threaded optimized implementation and a much greater improvement over both the single and multi threaded naive implementations, reaching minimum running time depending on the number of processor cores and the available memory bus bandwidth, i.e., it becomes memory bound using the minimum number of memory accesses. Given the simplicity and high speed of the lifting steps, an analysis based on the number of memory bus operations (read and write) is done for images that are larger than twice the shared cache size which establishes a lower bound for the running time of all linear memory access algorithms and also determines the maximum speed gains to be expected in relation to currently implemented parallel schemes based on the parallel execution of independent lifting steps. It also shows the optimality of the parallel algorithm presented. Finally, a comparison with currently available implementations shows the gains achieved by the proposed algorithm.
展开▼
