Archives Discussions

[d_a_a_]

I want to compare the serial performance with the parallel one

Hi,

I am working in a one-dimensional domain and I have a fixed number of work-items (global work size), suppose 1000 work-items. I execute a single call to enqueueNDRangeKernel as follows:

   queue.enqueueNDRangeKernel( kernel, cl::NDRange(), cl::NDRange(1000), cl::NDRange(1) );

Since I have a six-core AMD CPU (so CL_DEVICE_MAX_COMPUTE_UNITS = 6), OpenCL executes simultaneously six kernel instances (work-items). However, I would like to tell OpenCL to use only a single CPU core, is this possible?

I know that I could achieve the serial execution by setting global work size to 1, changing the code properly, and calling NDRangeKernel multiple times--unfortunately, given how the code is structured, "changing the code properly" would be a daunting task.

[nou]

set enviroment variable

export CPU_MAX_COMPUTE_UNITS=1

[d_a_a_]

Thank you guys.

Originally posted by: nou set enviroment variable

export CPU_MAX_COMPUTE_UNITS=1

Is this specific to the AMD's OpenCL implementation? Or, is the above environment variable part of the OpenCL specification?

[Illusio]

Alternatively, if you need more control, you can look at the device fission extension.

http://www.khronos.org/registry/cl/extensions/ext/cl_ext_device_fission.txt

[MicahVillmow]

d.a.a.,
If you want an OpenCL way to do it, use the device fission extension. CPU_MAX_COMPUTE_UNITS is not guaranteed to work.

[d_a_a_]

Thank you.