Multi2C: An Open-Source GPU compilation framework from OpenCL to Southern Islands ISA

q Popularity of GPU computing has led to extensive research in the design and optimization of workloads for GPUs. q Efficient utilization of hardware resources is the key to improve GPU application performance. q Compiler level optimizations directly influence the architecture and provide superior control over hardware resources. Architectural complexity of the GPU makes it challenging to implement advanced optimizations at the compiler level. q Additionally, the lack of open source tools (compilers, drivers, runtimes, etc.) from hardware vendors have limited the scope of optimizations. q We develop an open source framework that translates OpenCL to LLVM (frontend), and translates LLVM IR to the AMD Southern Islands GPU ISA (backend). q We have incorporated a backend transformation pass to optimize vector/scalar register usage. q Our compiler framework shows a speedup of 1.25x and 21.7 % savings in vector registers over the baseline greedy register allocation