I'm a compiler engineer on the ARM LLVM team at Qualcomm. My job is getting machines to run code faster, then checking that the speedup is real.
- LLVM backends: AArch64. I read IR and assembly until one of us blinks. NEON, SVE, SME.
- MLIR and ClangIR. I lower things one dialect at a time, upstream and in public, with reviewers.
- OpenMP, MPI and GPU offload. Teaching N cores to argue productively. In Google Summer of Code '25 I built OpenMP 6.0 infrastructure and GPU offloading for LFortran.
- HPC research. Parallel graph algorithms, four publications. Graphs are matrices in denial, but BIG data indeed.
- Fortran. Older than my dependencies, faster than my opinions.
Tools: C++ · LLVM · MLIR · AArch64 asm · OpenMP · MPI · CUDA · Fortran
$ man aditya | grep -A10 "KNOWN ISSUES"
- asks "what was the baseline?" at parties
- says "it's memory-bound" a lot. is right about as often as anyone
- has opinions on phase ordering and nobody to share them with
- will tell you your loop didn't vectorize because of the if-condition.
his own didn't either. we're fine.
- found out sanitizers can quietly stop the vectorizer. filed the
upstream issue. is not proud of how long it took.
- considers printf() a debugger for people who trust themselves
- believes Fortran was fast in 1957 and still is. cope.
Forge is a practice gym for compiler, systems and low-latency engineers.
- 120 problems and 64 guided labs covering compilers, systems and performance work.
- Hints, patterns and solutions are timed. They unlock only after you've had a real go, and Forge remembers what you peeked at.
- A market-making game, because why not.
Want to talk about vectorizers, AArch64 codegen, OpenMP or a benchmark you don't quite trust? Message me on LinkedIn.
measure twice, optimize once.