I am working to compile an OpenCL program which needs 64bit atomics (atomic_xchg and atomic_add, with long datatype). I have added " #pragma OPENCL EXTENSION cl_khr_int64_base_atomics : enable" and the code works with no issue for NVIDIA GeForce GTX 750 Ti but does not compile for "gfx902" under "Advanced Micro Devices, Inc. OpenCL 2.1 AMD-APP (2686.5)".
I am compiling in Windows 10 64 bit, under cygwin, using SDK 3.0 AMD and driver 19.1.1 Radeon Settings Version
2019.0109.1906.34385 with driver OpenCL™ Version 25.20.15011.1004.
What could the problem be ? Are 64 bit atomics not implemented on Windows for this GPU ?
If that is the case, is it any better in Linux ? Last time I checked (March 2018) there wasn't even a driver for Linux. Windows looks very stable, so I would prefer to continue working in this environment - especially I made both NVIDIA and AMD GPU work together.