ocular: Basic example of routing using OpenCL

[daveshah1]

This is currently still very experimental and needs a week or two's more work, but opening a PR in case anyone is interested in giving feedback on the current state or OpenCL integration.

Currently it can legally route some smaller ECP5 designs, with mediocre performance (about 20% slower than router1 with a GTX 960, with overhead being a big problem, only half the runtime is spent in the router kernel). Larger designs are currently hitting big congestion issues (the termination criterion is too close to a BFS and doesn't count the cost of the found solution) and also occasional illegal route tree failures (probably a race condition somewhere).

Things still TODO:

• Improved termination/congestion handling and bounding box expansion
• Remove duplicate queue entries
• Implement the near/far pile separation to improve runtime (some parts of this are present but it's not being used yet)
• Better selection options for OpenCL device
• Finding ways to reduce overhead
• Consider non-rectangular bounding boxes for long but low-fanout nets
• Benchmark performance on a FPGA accelerator card using the Xilinx/Intel OpenCL stacks
• Buy a better GPU ;)

[pepijndevos]

Very cool!

My limited understanding of OpenCL programming is that it's easy to write bad OpenCL and hard to write the good stuff, mostly because of memory locality and the ideal amount of parallelization depending on the hardware.

My only experience with GPGPU programming is using Futhark, which aims to take care of these details, and ideally allow you to write straightforward functional code and get optimized OpenCL out. So I admire your courage to do it the hard way :)

You mentioned FPGA accelerator cards as an OpenCL target, which is super interesting and appropriate for an FPGA tool haha. I asked the author of Futhark if he had considered supporting FPGAs, but he said the trade-offs for good performance are very different between GPUs and FPGAs, so curious to see if you will manage to get good performance out of both.

[rowanG077]

+1 on Futhark. Very easy to get relatively high quality GPGPU code.