VTL Backend Benchmark

This benchmark compares matrix multiplication (matmul) across:

• CPU (vtl.la.matmul)
• Vulkan backend (-d vulkan)
• OpenCL/VCL backend (-d vcl)

What it runs

benchmark/main.v executes matmul over square matrices for sizes:

• 64 x 64
• 128 x 128
• 256 x 256

Each size is repeated 3 times and averaged.

Run commands

##v run benchmark/main.v

##v -d vcl run benchmark/main.v

##v -d vulkan run benchmark/main.v

##v -d vulkan -d vcl run benchmark/main.v

Output format

Each line prints timing and speedup against CPU:

matmul [128x128]: CPU=8.12ms | Vulkan=1.20ms (6.77x) | VCL=1.55ms (5.24x)

If a backend is not enabled, it is omitted.

Notes

• Benchmarks are sensitive to GPU model, thermal state, and driver/runtime.
• First run can include setup overhead (shader/kernel setup, memory init).
• Use repeated runs for stable comparisons.

fn timeit(f fn () !) i64

timeit runs fn and returns elapsed nanoseconds.

struct BenchResult {
pub:
	name      string
	size      string
	cpu_ns    i64
	vulkan_ns i64 // -1 if not available
	vcl_ns    i64 // -1 if not available
}

BenchResult holds timing results for a single operation across backends.

fn (r &BenchResult) str() string

str returns a human-readable summary line for a BenchResult.