topic Thanks Jeffrey for in OpenCL* for CPU

device fission on gen9 GPU

Biren_Doshi — Wed, 30 Nov 2016 05:10:13 GMT

Is device fission is supported on Gen 9 GPU?

I want to run multiple OpenCL kernel concurrently on GPU. Is there any way i can run?

Thanks,

Biren Doshi

Device fission is implemented

Jeffrey_M_Intel1 — Wed, 30 Nov 2016 07:28:13 GMT

Device fission is implemented for Intel CPUs only. https://software.intel.com/en-us/articles/opencl-device-fission-for-cpu-performance. On the GPU the kernel will use all EUs. However, you may be able to increase the efficiency of your GPU utilization for small workloads by enabling out of order execution mode when you create your queue.

Quote:Jeffrey M. (Intel)

Huseyin_Tugrul_B_ — Thu, 01 Dec 2016 22:08:47 GMT

Jeffrey M. (Intel) wrote:

Device fission is implemented for Intel CPUs only. https://software.intel.com/en-us/articles/opencl-device-fission-for-cpu-performance. On the GPU the kernel will use all EUs. However, you may be able to increase the efficiency of your GPU utilization for small workloads by enabling out of order execution mode when you create your queue.

Does Intel gpu have context switching on compute-unit level? When I enqueue 100 different kernels using different number of compute units(such as one using only 2 units, some 5, some all,..), does the gpu set unused units idle or do they get their own work from another kernel?

Out of order execution may be

Jeffrey_M_Intel1 — Fri, 02 Dec 2016 16:03:25 GMT

Out of order execution may be a step in the direction you want. It isn't the same thing as context switching on CPUs but it can allow more efficient use of the hardware for small jobs. More info here: http://www.iwocl.org/wp-content/uploads/IWOCL2016-employing-ooqueues.pdf

Thanks Jeffrey for

Biren_Doshi — Thu, 08 Dec 2016 07:42:22 GMT

Thanks Jeffrey for explanation.