NVIDIA GPU Offloading Support For GCC Is Still Up & Coming
It's been nearly two years that there's been work going on for OpenACC 2.0 with GPU offloading for GCC, primarily geared for NVIDIA GPUs. That work continues taking shape and hopefully for GCC 6 the support will be in better standing.
I was asked by a Phoronix reader to run some GCC benchmarks when using the NVIDIA/NVPTX offloading to the graphics processor, but as far as I know, it's really not yet ready at that stage for doing performance benchmarks. The code has been mainline for a while and initial support is present in GCC 5, but there's still much left to be desired -- including OpenMP with GPU offloading support.
Progress though is being made with the GCC patches mailing list routinely seeing threads like new work around NVPTX and related improvements in libgomp, etc. The whole GCC accelerator scene is in fact still evolving for Intel's MIC, AMD HSA support, etc. With the NVIDIA PTX target, the issue remains though is that you'll need to be using the proprietary NVIDIA graphics driver for compute support.
If you wish to learn more about the current state of GCC offloading support, see their Wiki page.
I was asked by a Phoronix reader to run some GCC benchmarks when using the NVIDIA/NVPTX offloading to the graphics processor, but as far as I know, it's really not yet ready at that stage for doing performance benchmarks. The code has been mainline for a while and initial support is present in GCC 5, but there's still much left to be desired -- including OpenMP with GPU offloading support.
Progress though is being made with the GCC patches mailing list routinely seeing threads like new work around NVPTX and related improvements in libgomp, etc. The whole GCC accelerator scene is in fact still evolving for Intel's MIC, AMD HSA support, etc. With the NVIDIA PTX target, the issue remains though is that you'll need to be using the proprietary NVIDIA graphics driver for compute support.
If you wish to learn more about the current state of GCC offloading support, see their Wiki page.
1 Comment