GMX with Mesa Opencl on radeon

[alexxy]

I tryed to run opencl enabled mdrun with Mesa opencl however it seems that it cannot run opencl code due to undefined vars
alexxy@fermi ~/Tests/speptide $ cat nbnxn_ocl_kernels.cl.FAILED
Compilation of source file failed!
-- Used build options: -DWARP_SIZE_TEST=1 -D_WARPLESS_SOURCE_ -Iopencl
--------------LOG START---------------
In file included from input.cl:57:
In file included from opencl/nbnxn_ocl_kernels.clh:60:
In file included from opencl/nbnxn_ocl_kernel_utils.clh:36:
opencl/vectype_ops.clh:74:12: warning: implicit declaration of function 'distance' is invalid in C99
opencl/vectype_ops.clh:91:18: warning: implicit declaration of function 'atomic_cmpxchg' is invalid in C99
In file included from input.cl:57:
In file included from opencl/nbnxn_ocl_kernels.clh:60:
opencl/nbnxn_ocl_kernel_utils.clh:59:44: error: use of undeclared identifier 'CLK_NORMALIZED_COORDS_FALSE'
opencl/nbnxn_ocl_kernel_utils.clh:60:47: error: use of undeclared identifier 'CLK_ADDRESS_NONE'
opencl/nbnxn_ocl_kernel_utils.clh:61:47: error: use of undeclared identifier 'CLK_FILTER_NEAREST'
In file included from input.cl:60:
In file included from opencl/nbnxn_ocl_kernels.clh:140:
opencl/nbnxn_ocl_kernel_nowarp.clh:482:76: warning: implicit declaration of function 'erf' is invalid in C99
---------------LOG END----------------
alexxy@fermi ~/Tests/speptide $ grep CLK_NORMALIZED_COORDS_FALSE * -R
nbnxn_ocl_kernels.cl.FAILED:opencl/nbnxn_ocl_kernel_utils.clh:59:44: error: use of undeclared identifier 'CLK_NORMALIZED_COORDS_FALSE'
opencl/nbnxn_ocl_kernel_utils.clh:__constant sampler_t generic_sampler = CLK_NORMALIZED_COORDS_FALSE /* Natural coords /
alexxy@fermi ~/Tests/speptide $ grep CLK_NORMALIZED_COORDS_FALSE ~/Src/gromacs/ -R
/home/alexxy/Src/gromacs/src/gromacs/mdlib/nbnxn_ocl/nbnxn_ocl_kernel_utils.clh:__constant sampler_t generic_sampler = CLK_NORMALIZED_COORDS_FALSE /* Natural coords */

[ancahamuraru]

Are you running the latest version? Did you set correctly the OCL_FILE_PATH environment variable?

If not, try with the latest version from the scdev branch and follow the instructions here: https://github.com/StreamComputing/gromacs/wiki/GROMACS-GPU-ACCELERATION-USING-OPENCL#OpenCL_Setup

Are you running on the same machine as the one mentioned here: #35 ?

What does the output say about detected GPUs and the device selected for the current run? It should be similar to what you have previously send us:

2 GPUs detected:
#0: name: AMD TURKS, vendor: X.Org device version: OpenCL 1.1 MESA 10.4.0-devel, comp. units: 6, stat: compatible
#1: name: AMD TURKS, vendor: X.Org device version: OpenCL 1.1 MESA 10.4.0-devel, comp. units: 6, stat: compatible

1 GPU auto-selected for this run.
Mapping of GPU ID to the 1 PP rank in this node: AMD TURKS

[dkarkoulis]

Alexy,

distance, atomic_cmpxchg and erf are all OpenCL built-in function
so are the CLK_* names (type __constant sampler_t)
Check the OpenCL 1.1 reference card!

Are they implemented in the MESA driver?

[alexxy]

Hi!
They may be missing in MESA opencl driver. I'll ask MESA devs about it.

PS yes machine was the same as in previous report