Module gpu 

GPU utilities for kernel microbenchmarks.

Initializes AtlasRegistry, allocates device buffers, and provides CUDA event-based timing for Criterion iter_custom benchmarks.

Functions

bench_kernel_ms

Measure kernel execution time in milliseconds using CUDA events. Runs warmup warmup iterations, then iters timed iterations, returning the minimum time across 3 rounds.

ensure_registry

Ensure the registry is loaded (idempotent).

get_kernel

Resolve module::func for a bench.

gpu_alloc_zeroed

Allocate bytes of GPU memory, zero-initialized.

gpu_free

Free GPU memory.

gpu_sync

Synchronize the stream.

launch^⚠

Launch a kernel with raw parameters.