Haralick’s Texture Features Computation Accelerated by GPUs for Biological Applications

In biological applications, features are extracted from microscopy images of cells and are used for automated classification. Usually, a huge number of images has to be analyzed so that computing the features takes several weeks or months. Hence, there is a demand to speed up the computation by orders of magnitude. This paper extends previous results of the computation of co-occurrence matrices and Haralick texture features, as used for analyzing images of cells, by general-purpose graphics processing units (GPUs). New GPUs include more cores (480 stream processors) and their architecture enables several new capabilities (namely, computing capabilities). With the new capabilities (by atomic functions) we further parallelize the computation of the co-occurrence matrices. The visually profiling tool was used to find the most critical bottlenecks which we investigated and improved. Changes in the implementation like using more threads, avoiding costly barrier synchronizations, a better handling with divergent branches, and a reorganization of the thread tasks yielded the desired performance boost. The computing time of the features for one image with around 200 cells is compared to the original software version as a reference, to our first CUDA version with computing capability v1.0 and to our improved CUDA version with computing capability v1.3. With the latest CUDA version we obtained an improvement of 1.4 to the previous CUDA version, computed on the same GPU (gForce GTX 280). In total, we achieved a speedup of 930 with the most recent GPU (gForce GTX 480, Fermi) compared to the original CPU version and a speedup of 1.8 compared to the older GPU with the optimized CUDA version.