2018 has been a very busy year for me personally, and I thought for the sake of my own memory of the events I’d do a review of the things I’ve managed to achieve this year. Vulkan 1.1 / SPIR-V 1.3 & Subgroup Operations The single biggest item that I shipped this year was adding subgroup operations to Vulkan 1.1 and SPIR-V 1.3. It took just under two years of wrangling with all the hardware vendors to get a set of functionality that we could expose on all Vulkan 1.1 implementations.

So I’ve been playing around with Rust in my spare time for a few months now, and have been doing a lot of stuff that bridges C/C++ and Rust content. And one thing that has always bugged me is that I can use a super awesome allocator like jemalloc in Rust, but that the allocations being done via C++ were not being routed into this allocator. So as a tiny baby step into the Rust community, I’ve released my first crate cpp_new.

After 11 years since my first internship, and 9 years full time, I’m leaving Codeplay to join AMD. A Brief History of Codeplay and I I was a nineteen year old ned when I started my first internship at Codeplay, and was given the code-monkey job of hand optimizing shaders for an OpenGL ES 2.0 shader compiler, so that Codeplay’s compiler engineers could prove the capability of the compiler. Slated to take 10 weeks, I finished the task in 2 weeks and to stave off boredom in the remaining 8 weeks I ended up proving 100% code coverage of the compiler and writing a test generator to break it.

I’ve been meaning to try out Intel’s VTune now that it has a permanent trial mode (you can renew the 30-day trial license every 30-days without paying!), as I’ve long heard people rave about it as a profiling aid. I decided to pick one of my open source C libraries and see what performance problems I could find in the code, and I choose my JSON parser json.h. I’ve long known that the recursive approach I used in the parser will limit the performance I can get out of the code as compared to some of my competitors, but I support a metric-ton of extensional functionality (simplified JSON and JSON5 to name two major extensions that my users love).

After my previous post on how to read & write LLVM bitcode, I thought I’d follow it up with a post on actually modifying LLVM bitcode files after you’ve read them. LLVM comes with extensive built-in optimization passes, but also plenty of scope to do your own optimizations too. First off, remember that that when we parse an LLVM bitcode file, we get an LLVM module. So what exactly is an LLVM module?