pothosware / pothosopencl

Support for OpenCL-based processing blocks

Support for OpenCL-based processing blocks

Purpose

This component provides support for using OpenCL in the Pothos framework. The OpenClKernel block allows the execution of array-based OpenCl kernels inside a Pothos Topology. The block uses the Pothos DMA API to integrate OpenCL allocated buffers with the processing topology.

In addition, this component provides a device info plugin so the PothosGui and others can query information about OpenCl on a particular system.

Documentation

• https://github.com/pothosware/PothosOpenCL/wiki

Dependencies

• Pothos library
• OpenCl development libraries and headers

Building

configure, build, and install with CMake

Licensing information

Use, modification and distribution is subject to the Boost Software License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)

pothosopencl's Issues

test with FreeOCL - cpu based opencl

ConfLoader idea

A directory with .cl files and configuration file(s). The config file(s) have the parameters needed to configure the block (input/output types, kernel name, etc) and a block path. PothosOpenCL reads these file(s) and makes blocks with the given name, registry path, and PothosDoc. At this point, it could be accessed and would appear in PothosFlow like any other block, with all the underlying type and OpenCL source taken care of.

deal with blocks using back-to-back GPU memory

• optimization to avoid DMAs?
  • all consumers DMA to GPU

hardware dependent issue - keeping buffers on the gpu

It looks like the buffer optimization to avoid dma/xfers between blocks in the same domain doesnt work for some hardware. This means the feature is either hardware dependent, or the flags for buffer mapping etc may be incorrect. Needs investigation...

• Data dump: https://blackhole.pmtu.de/.dump/pothos/

OpenCL JIT block that demonstrates the DMA API

• use the C api to compile and execute kernels
• integrate kernel parameters with the call registry?

OpenCL system query support

Provide an implementation of the GPU system query API that can query for all GPUs and other CL devices on the system and properties per device. The implementation will be installed into the plugin registry.

Note: the main library will also need to define this "GPU system query API" somewhere.