Fast and guaranteed C compilation onto the PACT-XPP/spl trade/ reconfigurable computing platform

Abstract

We introduce the XPP-VC high-level compiler, which maps C programs to the coarse-grained XPP architecture. XPP-VC's main feature is the integration of pipeline vectorization and temporal partitioning techniques. The former provides high-throughput inner loop computations and the later allows the compilation of large programs or the use of fewer XPP processing elements. Although the preliminary results are very encouraging, improvements on the generation of configurations are still required. The evaluation we have conducted shows that only a few seconds are required to generate, from algorithms in C, the binaries to program the XPP. To our knowledge this compilation performance is unmatched by any other compiler targeting reconfigurable architectures. Moreover the compiler still achieves high-performance implementations. Since the XPP is delivered as an IP core or device to be coupled to a host processor, a future version of XPP-VC will consider co-compilation, i.e., compilation to hybrid microprocessor/XPP architectures.