VSL vs NumPy baselines

Run from ~/.vmodules:

v run vsl/benchmarks/vs_numpy/matmul_bench.v
v run vsl/benchmarks/vs_numpy/gemv_bench.v
v run vsl/benchmarks/vs_numpy/conv2d_bench.v

With OpenBLAS (recommended):

v -d vsl_blas_cblas run vsl/benchmarks/vs_numpy/matmul_bench.v

NumPy reference (numpy_baseline.py)

python3 vsl/benchmarks/vs_numpy/numpy_baseline.py matmul
python3 vsl/benchmarks/vs_numpy/numpy_baseline.py gemv
python3 vsl/benchmarks/vs_numpy/numpy_baseline.py conv2d

Compare GFLOPS / ms from the V scripts with the Python output for the same sizes.

Output and reporting

Use the same host, flags, and matrix sizes for both VSL and NumPy. A useful release note includes:

Field	Example
CPU/GPU	Ryzen 9, RTX 4060 Laptop
V flags	-d vsl_blas_cblas, -d cuda, -d vulkan
Operation	matmul, gemv, conv2d
Size	512, 1024, or Conv2D shape
Result	ms, GFLOPS, ratio vs baseline

CUDA/Vulkan benchmark variants should be treated as opt-in until dedicated CI coverage is added. Prefer scoped smoke tests for PR validation.

Priority sizes

Op	Shape
GEMM	128, 256, 512, 1024
GEMV	square m=n same sizes
Conv2D	1×1×32×32, kernel 3×3, stride 1

Tracked in issue #282.