alpha123 / tclopencl

VERY WIP OpenCL extension for Tcl

A VecTcl Extension For Parallel Numeric Computation on Heterogeneous Platforms

TclOpenCL provides both low-level and higher-level access to OpenCL 1.2+ APIs via SWIG and a TclOO wrapper.

THIS IS PROBABLY NOT USEFUL TO YOU YET

Installation

Binaries

Building TclOpenCL on Windows is a bit of a pain. You can download binary releases on GitHub. These have been verified to work with ActiveTcl 8.6 on 64-bit Windows 10 and the NVIDIA platform.

TclOpenCL requires VecTcl and an OpenCL implementation for your platform. Hacking on TclOpenCL requires SWIG.

• NVIDIA: Install the NVIDIA CUDA SDK
• AMD: Install the AMD Accelerated Parallel Processing SDK
• Intel CPU, HD Graphics, and Iris Graphics: Find the Intel OpenCL Drivers for your CPU and iGPU

Once that's installed, git clone https://github.com/alpha123/TclOpenCL and do the standard ./configure; make install. You may need to play with the --with-opencl=path/to/opencl/lib/dir configure flag on Windows. Type ./configure --help to see other flags.

Progress

What works:

• Create and query OpenCL devices and platforms
• Create command queues, contexts, and memory buffers
• Build OpenCL programs and query kernels
• Pass arguments to kernels and run them
• Do all of the above from a TclOO interface

What kind of works:

• Create memory buffers from VecTcl arrays
• Call kernels with VecTcl arrays as arguments

What doesn't work:

• Copy buffers back to VecTcl arrays....
• Fully asynchronous APIs based on OpenCL events
• A more pleasant way of managing workgroups
• A higher-level DSL for defining and running kernels