Towards “Write Once, Run Anywhere” HPC via Automated Translation
Session ChairWu Feng
Event Type
Emerging Technologies
TimeWednesday, November 16th10am - 6pm
Location155-B
DescriptionLeveraging the vast array of already-written CUDA codes, we propose a “write once in CUDA, run anywhere in OpenCL” approach to enable functional portability across a spectrum of parallel computing devices, including CPUs from AMD, ARM, IBM, and Intel; GPUs from AMD, NVIDIA, and PowerVR; MIC co-processors from Intel; and FPGAs from Altera and Xilinx. This approach, encompassed by our CU2CL: Automated CUDA-to-OpenCL Source-to-Source Translator, ingests an application’s CUDA source files and re-writes them into equivalent OpenCL host and kernel files. The emerging CU2CL tool has already supplanted the tedious translation of hundreds of thousands of lines of CUDA code (and 100+ CUDA-accelerated applications, including three major applications from neuroinformatics, molecular modeling, and molecular dynamics). Our Emerging Technology submission would showcase the following aspects of CU2CL: open-source release, ease of use, correctness, and instantiation across 100+ applications and many hardware devices (CPU/GPU/FPGA/MIC).
