GPGPU API for C# that can use multiple GPUs and CPU at the same time

[tugrul_SIMD]

"Cekirdekler API" is an open-source project which I uploaded to github newly.

This API helps developer to rewrite a bottlenecking hotspot loop or somewhat simple algorithm as C99 code and have it run on all selected OpenCL-capable devices at the same time. At each compute iteration, all devices get fair amount of work depending on their performances and capabilities. They can be totally different vendors and different segments' GPUs.

You can find it in github:

(wiki) https://github.com/tugrul512bit/Cekirdekler/wiki

(download) https://github.com/tugrul512bit/Cekirdekler

also there is a short tutorial about it in here:

https://www.codeproject.com/Articles/1181213/Easy-OpenCL-Multiple-Device-Load-Balancing-and-Pip

Traditional hello-world looks like this:

            ClNumberCruncher cr = new ClNumberCruncher(
                AcceleratorType.GPU, @"
                    __kernel void hello(__global char * arr)
                    {
                        printf(""hello world"");
                    }
                ");

            ClArray<byte> array = new ClArray<byte>(1000);
            array.compute(cr, 1, "hello", 1000, 100);

 

[tugrul_SIMD]

Here you can see the load balancer in action

 

[tugrul_SIMD]

As of version 1.2.0, device to device pipelining feature is working.

If there are more than one OpenCL kernels that are needed to run consecutively and if none of them are distributable to multiple GPUs, then this new feature can run all of them at the same time as a single pipeline's stages with doublebuffering to overlap both computations and data movements between stages.

each stage is built from a list of kernels, input-output arrays and an OpenCL device. Then stages are added together to create a pipeline that works whenever client code pushes data to entrance of it. Each push makes a new result pop from the end point.

https://github.com/tugrul512bit/Cekirdekler/wiki/Pipelining:-Device-to-Device

 

[eidairaman1]

@FordGT90Concept

 

[tugrul_SIMD]

Now it has batch computing option with task-pool and device-pool features.

it uses all pipelines of a GPU and multi GPU scaling is higher than load balancing even with asymmetric GPU setups.

 

[tugrul_SIMD]

Now one can use OpenCL 2.0 dynamic parallelism feature.