def _do_for_params(self, horizon):
# setup test data
numpy.random.seed(42)
n_step, n_node, n_cvar, n_svar, n_thread = 100, 5, 2, 4, self.n_thread
cvars = numpy.random.randint(0, n_svar, n_cvar).astype(numpy.int32)
out = numpy.zeros((n_step, n_node, n_cvar, n_thread), numpy.float32)
delays = numpy.random.randint(0, horizon - 2,
(n_node, n_node)).astype(numpy.int32)
weights = numpy.random.randn(n_node, n_node).astype(numpy.float32)
weights[numpy.random.rand(*weights.shape) < 0.25] = 0.0
state = numpy.random.randn(n_step, n_node, n_svar,
n_thread).astype(numpy.float32)
buf = numpy.zeros((n_node, horizon, n_cvar, n_thread), numpy.float32)
# debugging
delayed_step = numpy.zeros_like(delays)
# setup cu functions
pre = cu_linear_cfe_pre(0.0, 1.0, 0.0)
post = cu_linear_cfe_post(1.0, 0.0)
dcf = cu_delay_cfun(horizon,
pre,
post,
n_cvar,
self.block_dim[0],
step_stride=1,
aff_node_stride=1)
# run it
@self.jit_and_run(out, delays, weights, state, cvars,
buf) #,delayed_step)
def kernel(out, delays, weights, state, cvars, buf): #, delayed_step):
i_thread = cuda.blockDim.x * cuda.blockIdx.x + cuda.threadIdx.x
for step in range(state.shape[0]):
for i_post in range(state.shape[1]):
dcf(out, delays, weights, state, i_post, i_thread, step,
cvars, buf) #,delayed_step)
# ensure buffer is updating correctly
buf_state = numpy.roll(state[:, :, cvars][-horizon:].transpose(
(1, 0, 2, 3)),
n_step,
axis=1)
numpy.testing.assert_allclose(buf, buf_state)
# ensure buffer time indexing is correct
# numpy.testing.assert_equal(delayed_step, (n_step - 1 - delays + horizon) % horizon)
# replay
nodes = numpy.tile(numpy.r_[:n_node], (n_node, 1))
for step in range(horizon + 3, n_step):
delayed_state = state[:, :, cvars][
step - delays, nodes] # (n_node, n_node, n_cvar, n_thread)
afferent = (weights.reshape(
(n_node, n_node, 1, 1)) * delayed_state).sum(
axis=1) # (n_node, n_cvar, n_thread)
numpy.testing.assert_allclose(afferent, out[step], 1e-5, 1e-6)
def test_linear_pre(self):
ai, aj, intercept = -0.2, 0.3, 0.25
out = numpy.zeros((self.n_thread,), 'f')
xj, xi = numpy.random.rand(2, self.n_thread).astype('f')
pre = cu_linear_cfe_pre(ai, aj, intercept)
@self.jit_and_run(out, xi, xj)
def kernel(out, xi, xj):
t = cuda.blockDim.x * cuda.blockIdx.x + cuda.threadIdx.x
out[t] = pre(xi[t], xj[t])
numpy.testing.assert_allclose(out, ai * xi + aj * xj + intercept, 1e-4, 1e-5)