Runtimes

dspk runs every operation on a runtime: cuda (NVIDIA), wgpu (Vulkan / Metal / DirectX, any GPU) or cpu. The same call gives the same result on each; only the time differs. Every function takes runtime=; without it, the one dspk.runtime.set chose (default: the first available).

Which runtimes this build has, and which one is in use

import time

import numpy as np

import dspk
from plating import signals

print("available:", dspk.runtime.available())
print("in use:   ", dspk.runtime.current())

The same operation on every runtime: identical results, different times

s = signals.synthetic(channels=64, seconds=2.0)
results, times = {}, {}
for rt in dspk.runtime.available():
    try:
        dspk.ops.spatial.car(s.data, runtime=rt)  # the first call compiles the kernels (cached on disk)
    except RuntimeError as e:  # compiled in, but no such device on this machine
        print(f"{rt}: not usable here ({e})")
        continue
    start = time.perf_counter()
    results[rt] = dspk.ops.spatial.car(s.data, runtime=rt)
    times[rt] = (time.perf_counter() - start) * 1e3
first = next(iter(results.values()))
for rt, out in results.items():
    print(f"{rt:5s} {times[rt]:7.1f} ms   max difference from {next(iter(results))}: {np.max(np.abs(out - first)):.1e}")

The whole script is playground/01_basics/01_runtimes.py. Run it with uv run python playground/01_basics/01_runtimes.py.