GPU trace options (Linux, Windows only)

You can trace operations that take place with the graphics processing unit (GPU) by using the standard SDK trace options, including sending the trace output to a file.

Tracing can be set up in the standard way by invoking the -Xtrace option on the command line. For example, you can use method trace to investigate problems with the CUDA4J or com.ibm.gpu classes.

For tracing CUDA operations at a native level you can use the ibm_gpu trace component.

For more information about tracing, trace components, and tracepoints, see the following topics:

-Xtrace in the OpenJ9 user documentation
Determining the tracepoint ID of a trace point in the J9 VM reference

CUDA4J example

To trace all methods of all CUDA4J classes, you can specify the following option on the command line:

-Xtrace:print="mt,methods={com/ibm/cuda/*.*}"

com.ibm.gpu example

To trace all methods of all com.ibm.gpu classes, you can specify the following option on the command line:

-Xtrace:print="mt,methods={com/ibm/gpu/*.*}"

To trace native CUDA operations, you can use the following command line option to output all trace data to the file trace.out:

java -Xtrace:maximal=ibm_gpu,output=trace.out

Here is a sample of the trace output that is generated and formatted:

10:22:41.289702969  0x0000000000010500 ibm_gpu.0           Entry      
    >Called IBM_GPU_sortIntArray with params: numElements=5, device number: 0
10:22:41.289708961  0x0000000000010500 ibm_gpu.1           Event       
    IBM_GPU_sortIntArray starting CUDA malloc, numBytes=20
10:22:41.359505039  0x0000000000010500 ibm_gpu.2           Event       
    IBM_GPU_sortIntArray completed CUDA malloc, starting CUDA memcpy (host to 
    device), deviceData=0x8900400000, hostData=0x3fffa4698428, numBytes=20
10:22:41.359529008  0x0000000000010500 ibm_gpu.3           Event       
    IBM_GPU_sortIntArray completed CUDA memcpy (host to device), starting sort
10:22:41.363024404  0x0000000000010500 ibm_gpu.4           Event       
    IBM_GPU_sortIntArray completed sort, transferring from device to host: 
    deviceData=0x8900400000, hostData=0x3fffa4698428, numBytes=20
10:22:41.363119279  0x0000000000010500 ibm_gpu.5           Event       
    IBM_GPU_sortIntArray completed device to host memcpy
10:22:41.363637597  0x0000000000010500 ibm_gpu.6           Exit       
    <IBM_GPU_sortIntArray - return code=0
10:22:41.363686532  0x0000000000010500 ibm_gpu.7           Entry      
    >Called IBM_GPU_sortFloatArray with params: numElements=5, device number: 0
10:22:41.363687531  0x0000000000010500 ibm_gpu.8           Event       
    IBM_GPU_sortFloatArray starting CUDA malloc, numBytes=20
10:22:41.363793392  0x0000000000010500 ibm_gpu.9           Event       
    IBM_GPU_sortFloatArray completed CUDA malloc, starting CUDA memcpy (host to 
    device), deviceData=0x8900400000, hostData=0x3fffa4b98b08, numBytes=20
10:22:41.363805376  0x0000000000010500 ibm_gpu.10          Event       
    IBM_GPU_sortFloatArray completed CUDA memcpy (host to device), starting sort
10:22:41.365438225  0x0000000000010500 ibm_gpu.11          Event       
    IBM_GPU_sortFloatArray completed sort, transferring from device to host: 
    deviceData=0x8900400000, hostData=0x3fffa4b98b08, numBytes=20
10:22:41.365531103  0x0000000000010500 ibm_gpu.12          Event       
    IBM_GPU_sortFloatArray completed device to host memcpy
10:22:41.366046424  0x0000000000010500 ibm_gpu.13          Exit       
    <IBM_GPU_sortFloatArray - return code=0
10:22:41.366062403  0x0000000000010500 ibm_gpu.21          Entry      
    >Called IBM_GPU_sortDoubleArray with params: numElements=5, device number: 0
10:22:41.366063402  0x0000000000010500 ibm_gpu.22          Event       
    IBM_GPU_sortDoubleArray starting CUDA malloc, numBytes=40
10:22:41.366170261  0x0000000000010500 ibm_gpu.23          Event       
    IBM_GPU_sortDoubleArray completed CUDA malloc, starting CUDA memcpy (host to device), deviceData=0x8900400000, hostData=0x3fffa4b79a18, numBytes=40
10:22:41.366182245  0x0000000000010500 ibm_gpu.24          Event       
    IBM_GPU_sortDoubleArray completed CUDA memcpy (host to device), starting sort
10:22:41.369051467  0x0000000000010500 ibm_gpu.25          Event       
    IBM_GPU_sortDoubleArray completed sort, transferring from device to host: deviceData=0x8900400000, hostData=0x3fffa4b79a18, numBytes=40
10:22:41.369144344  0x0000000000010500 ibm_gpu.26          Event       
    IBM_GPU_sortDoubleArray completed device to host memcpy
10:22:41.369662662  0x0000000000010500 ibm_gpu.27          Exit       
    <IBM_GPU_sortDoubleArray - return code=0
10:22:41.369677642  0x0000000000010500 ibm_gpu.14          Entry      
    >Called IBM_GPU_sortLongArray with params: numElements=5, device number: 0
10:22:41.369678641  0x0000000000010500 ibm_gpu.15          Event       
    IBM_GPU_sortLongArray starting CUDA malloc, numBytes=40
10:22:41.369784501  0x0000000000010500 ibm_gpu.16          Event       
    IBM_GPU_sortLongArray completed CUDA malloc, starting CUDA memcpy (host to device), deviceData=0x8900400000, hostData=0x3fffa4b98b98, numBytes=40
10:22:41.369796486  0x0000000000010500 ibm_gpu.17          Event       
    IBM_GPU_sortLongArray completed CUDA memcpy (host to device), starting sort
10:22:41.371647048  0x0000000000010500 ibm_gpu.18          Event       
    IBM_GPU_sortLongArray completed sort, transferring from device to host: deviceData=0x8900400000, hostData=0x3fffa4b98b98, numBytes=40
10:22:41.371739926  0x0000000000010500 ibm_gpu.19          Event       
    IBM_GPU_sortLongArray completed device to host memcpy
10:22:41.372258243  0x0000000000010500 ibm_gpu.20          Exit       
    <IBM_GPU_sortLongArray - return code=0