Learning CUDA

CUDA exercises from Programming Massively Parallel Processors (PMPP), with correctness tests (GoogleTest) and benchmarks against optimized libraries (Google Benchmark + Thrust/cuBLAS).

Build

cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc)

Run tests

ctest --test-dir build --output-on-failure

Run benchmarks

./build/exercises/bench_vecadd

Adding a new exercise

1. Create exercises/name/ with:
   • name.cuh — your kernel implementation
   • test_name.cu — correctness tests (GoogleTest)
   • bench_name.cu — benchmarks vs optimized library
2. Add add_exercise(name) in exercises/CMakeLists.txt

Project structure

common/              Shared utilities (error checking, timing, device helpers)
exercises/
  vecadd/            Vector addition — kernel, test, benchmark
  ...