IPMACC – An Open Source OpenACC to CUDA/OpenCL Translator

IPMACC is  a research-grade open-source framework for translating OpenACC source code to CUDA or OpenCL. Binary executables can then be created with OpenCL or CUDA compilers. The authors (Ahmad Lashgar – University of Victoria, Alireza Majidi – Texas A&M University, Amirali Baniasadi – University of Victoria)  verified correctness and performance using benchmarks from the Rodinia Benchmark Suit and CUDA SDK. IMPACC is of interest due to the recent demise of CAPS-enterprise who provided a commercial OpenACC to OpenCL source translator. IPMACC can be found in it’s github repository. Also note that gcc will start supporting OpenACC and OpenMP 4.0 pragmas in 2015.

IMPACC translation pipeline and performance results (images courtesy Arxiv.org)

Limitations

• Currently, parallel directive is not supported. Notice that with a little effort by the programmer, any parallel region construct can be translated into a kernels region.
• Only 1D array can be transfered in-out the region.
• User-defined data types are not supported for data copy clauses.
• C/C++ main function should be prototyped as normal function with the output. e.g. int main(). Avoid declaring main as main() with no return type.
• Clause support: seq clause for the top-level 1-nested loop is not supported. This is weird case where there is only one loop in the region which is targeted for serial execution.
• There are some issues between NVCC and C’s restrict keyword in CUDA 4.0.
• In case the compiler crashed from pycparser.plyparser.ParseError class, check the last line and look for meaningful prompt.
• Limitations on the Reduction/Private clause of loop
• IPMACC assumes the reduction/private variable is not declared inside the loop.
• If the variable is defined as both private() and reduction(), IPMACC assumes reduction which covers private too.
• Reduction/Private on array/subarray is not supported
• Default reduction type is two-level tree reduction [1]. Alternatively for CUDA, atomic reduction is implemented and it is supported only on recent hardwares (compute capability >= 1.3). Proper flag should be passed to underlying NVCC; add -arch=sm_13 compile flag.
• To gurantee the safety, it is necessary to use acc_init() early in the code to avoid potentially runtime errors. This is essential for the OpenCL target devices.
• IPMACC can parallel the iterations of loops with the following increment steps: +, -, ++, –, *, /