Olajide Badejo

Selected work

• CUDA-Kernel-Library

  A GEMM ladder from a 1,908 GFLOP/s naive kernel to 54,985 GFLOP/s using mma.sync, ldmatrix, and swizzled shared memory, which is 90% of cuBLAS in FP16 on the same GPU, with the tensor pipe at 83% of speed of light. Each rung is a separate kernel with its profiler run committed.

• MLIR-Backend-for-a-Simulated-Edge-NPU

  A deep learning compiler backend lowering ONNX through custom MLIR dialects to a binary instruction stream for a simulated NPU. On LeNet, -O2 takes 91 instructions to 70, 23,421 simulated cycles to 12,710, and DRAM traffic from 339 KB to 177 KB, with output matching onnxruntime to 3e-8.

• Parallel-Numerical-Library

  A C++23 library whose iterative solvers run unchanged over serial, OpenMP, threads, MPI, and CUDA backends and return bit identical results. Measured against each device's own STREAM bandwidth, 62.3 GiB/s on the CPU and 549.7 GiB/s on the GPU, red black Gauss Seidel reaches 45.7% and 47.1% of them. 440 configurations, reproducible end to end in 23 minutes.

• Particle-Physics-Simulation

  A CUDA engine for N-body gravity and SPH fluids sustaining 17.47 TFLOP/s in FP32, with shared memory tiling 1.93x over the naive kernel at 200,000 particles. Correctness is pinned by ten gates, including energy drift of 1.33e-05 over 10,000 steps.

• Ring-AllReduce

  Ring allreduce built from MPI point to point primitives, then fitted to the alpha beta cost model at 0.250 us per message and 305.4 ps per byte, R squared 0.981 across 375 points. The write up is explicit about where it loses to the vendor collective and why.

• Low-Power-Async-CDC-FIFO-IP

  A parameterized asynchronous clock domain crossing FIFO in SystemVerilog with gray code pointers and low power intent, proven bounded in SymbiYosys and swept across 60 configurations for area, power, and timing.

More repositories, including an out of tree LLVM pass framework, CPU microbenchmarks, GPU roofline profiling, MPI collective benchmarking, and ML experiment tooling, are on my GitHub profile.