oclkit is a small set of C functions, to avoid writing the same OpenCL boiler plate over and over again, yet keeping full control over executing OpenCL. It also contains a set of binaries to check and test for the installed OpenCL runtime.
The API should be pretty self-explanatory by examining ocl.h.
Run make in the top-level directory and change into build/examples. Most
binaries are compiled with command-line flags to chose the OpenCL platform and
device type.
Measures the time for typical boilerplate operations such as clCreateContext,
clBuildProgram etc.
$ ./check-infrastructure-times
Create context: 0.150983 s
Build program : 0.002060 s
Create kernel : 0.000004 s
Create buffer : 0.000003 s
Cleanup : 0.061502 s
Runs a dummy kernel and measures the OpenCL profiling times and wall clock time for submission and execution.
$ ./check-launch-latencies
GeForce GTX TITAN Black
wait for submission: 2.88220 us
wait for execution : 7.48316 us
wall clock : 22.12242 us
GeForce GTX 580
wait for submission: 2.38133 us
wait for execution : 4.08007 us
wall clock : 16.33144 us
Uses a single blocking, an out-of-order, two or three queues to write data, execute a kernel and read back data. The total
Outputs the queue profiling timer resolution for each device.
$ ./test-profile-timer
GeForce GTX TITAN Black : 1000 ns
GeForce GTX 580 : 1000 ns
Outputs the double extension flags for each device.
$ ./test-double-flags
GeForce GTX TITAN Black
cl_khr_fp64 = 1
cl_amd_fp64 = 0
GeForce GTX 580
cl_khr_fp64 = 1
cl_amd_fp64 = 0
Outputs the compiled binary (which might be PTX assembly for NVIDIA GPUs) for each device by appending a counter, i.e.
$ ./dump-opencl-binary test.cl
Generates output files test.cl.0, .test.cl.1 etc.
The code is licensed under GPL v3.