Abstract: Multi-processor DSPs offer a high-performance cost-effective method that is critical for many embedded application areas. However, it is currently complex and time-consuming to port existing uniprocessor applications to such parallel architectures. There are no commercially available compilers which will automatically map existing sequential DSP programs to a multi-processor machine (Rijpkema et al., 1999). Users are typically required to rewrite their code as a process network or as a set of sequential processes of communication (Lee, 1995). It is well known that such an approach is highly non-trivial and error-prone, possibly leading to a deadlock. This is due to the fact that Digital Signal Processor programs are written using C language rather than FORTRAN language (Hiranandani et al., 1992) and the wide use of pointer arithmetic is shuffled and the compilation of distributed memory space of Digital Signal Processors is difficult. It is a highly specialized skill to rewrite an application in a parallel way. What is needed is a tool that effectively takes existing programs into the new multi-processor architecture and maps them automatically. Although research on auto-parallelising compilation in scientific computing has been going on for over 20 years (Gupta et al., 2000), this has not occurred in the embedded domain. These problems are solved through the use of the pointer conversion technique and a new address resolution technique based on a new data transformation scheme that allows multiple address spaces to be paralleled without the introduction of complex (and potentially deadlocking) passing code message. An auto-parallel C compiler integrating these two techniques into an overall parallel strategy 

Keywords: Multi-processors DSPs, Digital Signal Processor programs, Data transformation, Prallelization algorithm, data address resolution


PDF | DOI: 10.17148/IJARCCE.2019.8702

Open chat
Chat with IJARCCE