def gpu_version(self, sig, filt, boundary, mode):
with cp.cuda.Stream.null:
out = cusignal.convolve2d(sig,
filt,
boundary=boundary,
mode=mode)
cp.cuda.Stream.null.synchronize()
return out
def test_convolve2d(num_samps, num_taps, boundary, mode):
cpu_sig = np.random.rand(num_samps, num_samps)
cpu_filt = np.random.rand(num_taps, num_taps)
gpu_sig = cp.asarray(cpu_sig)
gpu_filt = cp.asarray(cpu_filt)
cpu_convolve2d = signal.convolve2d(
cpu_sig, cpu_filt, boundary=boundary, mode=mode
)
gpu_convolve2d = cp.asnumpy(
cusignal.convolve2d(gpu_sig, gpu_filt, boundary=boundary, mode=mode)
)
assert array_equal(cpu_convolve2d, gpu_convolve2d)