Programming Massively Parallel Processors 4th – CUDA Examples & Exercises

What this project is

CUDA examples following "Programming Massively Parallel Processors". Currently includes basic kernels, image filters, matrix ops with cuBLAS validation, and device query utility.

Project layout

• CMakeLists.txt – builds all examples (CUDA 12+, CMake 3.20+, C++17)
• include/ – shared headers (cdiv.h for ceiling division)
• src/chapter02/ – vectorAdd.cu
• src/chapter03/ – rgb_to_gray.cu, blur.cu, matrixMul.cu, exercise_3_1.cu, exercise_3_2.cu
• src/chapter04/ – query_device.cu
• assets/ – copied next to binaries after build (if present)

Prerequisites

• CUDA Toolkit 11.x (nvcc available on PATH)
• CMake ≥ 3.20
• C++17 toolchain (MSVC/Clang/GCC); on Windows, MSVC 2022 works
• GPU with CUDA support

Configure & build

# from repo root
cmake -B build -S .
cmake --build build --config Release

Build & run on WSL (Ubuntu)

• Ensure NVIDIA drivers and CUDA are installed in WSL (CUDA Toolkit 12.x recommended) and /usr/local/cuda/bin is on PATH.
• Use GCC/Clang toolchain; CMake will detect nvcc automatically.

# inside WSL, from repo root
cmake -B build -S .
cmake --build build --config Release

# run examples
./build/bin/Release/chapter04/query_device    # list devices
./build/bin/Release/chapter03/exercise_3_2 1024

Targets & outputs

• Chapter 02
  • vectorAdd → build/bin/Release/chapter02/vectorAdd.exe
• Chapter 03
  • rgbToGray → build/bin/Release/chapter03/rgbToGray.exe
  • blurFilter → build/bin/Release/chapter03/blurFilter.exe
  • matrixMul (cuBLAS SGEMM validation) → build/bin/Release/chapter03/matrixMul.exe
  • exercise_3_1 (row vs column sweep multiply, cuBLAS compare) → build/bin/Release/chapter03/exercise_3_1.exe
  • exercise_3_2 (matrix–vector multiply, cuBLAS SGEMV compare) → build/bin/Release/chapter03/exercise_3_2.exe
• Chapter 04
  • query_device (device properties dump) → build/bin/Release/chapter04/query_device.exe

Assets (if present) are auto-copied beside the binaries.

Running examples (Windows PowerShell)

# run with default sizes
./build/bin/Release/chapter02/vectorAdd.exe
./build/bin/Release/chapter03/rgbToGray.exe
./build/bin/Release/chapter03/blurFilter.exe
./build/bin/Release/chapter03/matrixMul.exe
./build/bin/Release/chapter03/exercise_3_1.exe
./build/bin/Release/chapter03/exercise_3_2.exe 1024   # optional size argument
./build/bin/Release/chapter04/query_device.exe

Implementation notes

• Timing uses CUDA events; cuBLAS paths include a warm-up call to remove first-call overhead.
• Matrix kernels are row-major; cuBLAS expects column-major, so validation transposes host matrices before GEMM/GEMV.
• exercise_3_1 compares row-sweep and column-sweep kernels against cuBLAS SGEMM.
• exercise_3_2 computes C = A * b (matrix–vector); validated with cuBLAS SGEMV.

Next steps

• Add more chapter04 samples (occupancy, bandwidth tests).
• Extend performance reporting (GFLOP/s, bandwidth) for existing exercises.