cl-gpu-example: CPU vs GPU Vector Addition Benchmark — Common Lisp

CPU vs GPU vector addition (100 million single-floats) in Common Lisp, using CFFI to call CUDA via a shared library.

Prepared with the help of Claude Code.

Results

Version	Time (s)	Speedup
CPU	0.6207	1×
GPU	0.0168	~37×

Architecture

Common Lisp (SBCL)
    │
    ├── CPU benchmark: pure CL with optimized (simple-array single-float)
    │
    └── GPU benchmark: CL orchestrates full CUDA pipeline via CFFI
            │
            └── libclgpu.so (thin C wrappers around CUDA runtime API)
                    │
                    └── CUDA kernel (vector_add_kernel)

The CL side controls the entire GPU workflow — memory allocation, host/device transfers, kernel launch with timing, verification, and cleanup — making the GPU programming model fully transparent and reusable for other kernels.

Requirements

• SBCL with Quicklisp (for CFFI dependency)
• CUDA Toolkit (nvcc compiler)
• NVIDIA GPU with CUDA support

WSL2 Setup (Windows)

1. Install NVIDIA drivers on Windows (regular GeForce/Studio drivers)
2. Install CUDA Toolkit in WSL2:

   wget https://developer.download.nvidia.com/compute/cuda/repos/wsl-ubuntu/x86_64/cuda-keyring_1.1-1_all.deb
   sudo dpkg -i cuda-keyring_1.1-1_all.deb
   sudo apt-get update
   sudo apt-get install -y cuda-toolkit-12-8

3. Add to ~/.bashrc:

   export PATH=/usr/local/cuda/bin:$PATH
   export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH

Build & Run

make cuda       # Build cuda/libclgpu.so
make cpu        # Run CPU benchmark only (no GPU needed)
make gpu        # Build CUDA + run GPU benchmark
make run        # Build CUDA + run full comparison with speedup
make clean      # Remove build artifacts

Run make from within WSL2, not from Windows cmd/PowerShell.

--dynamic-space-size 4096 is set in the Makefile because the benchmark allocates 3 arrays of 100M single-floats (~1.2 GB), exceeding SBCL's default 1 GB heap.

Bash usage

# Full benchmark (CPU vs GPU with speedup)
sbcl --dynamic-space-size 4096 --non-interactive \
  --eval '(push *default-pathname-defaults* asdf:*central-registry*)' \
  --eval '(asdf:load-system "cl-gpu")' \
  --eval '(cl-gpu:SHOW-benchmark)'

# CPU only (no GPU needed)
sbcl --dynamic-space-size 4096 --non-interactive \
  --eval '(push *default-pathname-defaults* asdf:*central-registry*)' \
  --eval '(asdf:load-system "cl-gpu")' \
  --eval '(cl-gpu:SHOW-cpu-vector-add)'

# GPU only
sbcl --dynamic-space-size 4096 --non-interactive \
  --eval '(push *default-pathname-defaults* asdf:*central-registry*)' \
  --eval '(asdf:load-system "cl-gpu")' \
  --eval '(cl-gpu:SHOW-gpu-vector-add)'

REPL Usage

From an interactive SBCL session (started with sbcl --dynamic-space-size 4096):

(push *default-pathname-defaults* asdf:*central-registry*)
(asdf:load-system "cl-gpu")

(cl-gpu:SHOW-benchmark)       ; Full comparison
(cl-gpu:SHOW-cpu-vector-add)  ; CPU only
(cl-gpu:SHOW-gpu-vector-add)  ; GPU only

CL-Specific Notes

• CPU version uses (declare (type (simple-array single-float (*)))) with (optimize (speed 3) (safety 0)) for SBCL to generate efficient native code
• GPU version uses cffi:foreign-alloc for host arrays (required for CUDA memcpy compatibility)
• with-cuda-memory and with-cuda-timing macros provide clean resource management with unwind-protect
• All CUDA bindings are exported from the :cl-gpu package for reuse with custom kernels

end of file