high performance computing on graphics processing units: hgpu.org

In this report, we have taken the first steps in investigating the feasibility of using the GPU as a cryptographic accelerator for the AES algorithm on mobile devices. In particular, our focus was on exploring the use of OpenCL as a framework for implementing the algorithm. Using modifications of an existing implementation [11], we first showed that OpenCL schedules work groups to execute exclusively on the compute units of the GPU, and that optimal throughput must be achieved through more than just a simple data parallel implementation. Next, we investigated how performance varied with the implementation as the number of active work items in a single compute unit was varied, discovering a surprising level off in encryption throughput when using only 1/4 of the available work items. Next we investigated two potential reasons for lower than expected performance: strided access to input data within work groups, and redundant random accesses to the T-box lookup table that could be avoided using local memory. We discovered the T-box lookup table optimization provided the best increase in throughput, whereas the coalesced access implementation only added to the runtime. Finally, we determined that using optimal parameters for the number of work groups and work group size, we were able to achieve a 1.79 fold speedup in maximum throughput over a single CPU using the OpenSSL implementation.