I made an OpenACC code for project(to publish in journal), and used NVIDIA Geforce to run that program.
However, since I need to use a lot of double type variables, to reduce calculation time, I bought FirePro W9100 with research fund.
But the problem is that the program does not calculate correctly with W9100(Even though the PGI compiler says that there is no problem when I compile that OpenACC code for Hawaii archtecture via -ta=radeon:hawaii.), though it works fine with NVIDIA GPU or Intel CPU(by using -ta=multicore compiler option). To say it more specifically, simple OpenACC code with matrix addition gives wrong results(nan) for W9100, but gives correct results for NVIDIA GPU or Intel CPU.
(I'm using PGI Community Edition 16.10 to compile OpenACC code)
I posted a question about this in PGI forum, but it seems that the newer version does not support AMD GPU when compiling OpenACC code...
So, may I ask which compiler would compile OpenACC code for W9100 without problem? I heard that AMD joined OpenACC(http://developer.amd.com/community/blog/2015/01/07/amd-joins-openacc/), so I think there would be at least one compiler which compiles OpenACC code for W9100.
Or should I learn OpenCL to use W9100 to perform calculations? I fear that it would take too much time to learn OpenCL and convert OpenACC code to OpenCL code.(My intention is to reduce calculation time for existing code. So, if possible, I want to reduce efforts to make the existing code run for W9100)
Thank you for your attention.
Please let me know if you have found the method to use W9100 with OpenACC?
GCC seems to have some sort of OpenACC support for GCN:
(but for GCN3 while W9100 has GCN2)