def test_reduction(self):
if self.dtype == numpy.float16:
self.skipTest('Not supported.')
c = testing.shaped_random((30, ), cupy, self.dtype, seed=2)
c_orig = c.copy()
desc_a = cutensor.create_tensor_descriptor(self.a)
desc_c = cutensor.create_tensor_descriptor(c)
d = cutensor.reduction(self.alpha, self.a, desc_a, self.mode_a,
self.beta, c, desc_c, ('x', ))
assert c is d
testing.assert_allclose(
self.alpha * self.a_transposed.sum(axis=(1, 2)) +
self.beta * c_orig,
d,
rtol=self.tol,
atol=self.tol)
def test_reduction(self):
c = testing.shaped_random((30,), cupy, numpy.float32, seed=2)
c_orig = c.copy()
desc_a = cutensor.create_tensor_descriptor(self.a, ct.OP_COS)
desc_c = cutensor.create_tensor_descriptor(c, ct.OP_TANH)
d = cutensor.reduction(
self.alpha, self.a, desc_a, self.mode_a,
self.beta, c, desc_c, ('x',),
reduce_op=ct.OP_MAX
)
assert c is d
testing.assert_allclose(
self.alpha * cupy.cos(self.a_transposed).max(axis=(1, 2)) +
self.beta * cupy.tanh(c_orig),
d,
rtol=1e-6, atol=1e-6
)
def test_reduction(self):
c = testing.shaped_random((30,), cupy, numpy.float32, seed=2)
c_orig = c.copy()
desc_a = cutensor.create_tensor_descriptor(self.a)
desc_c = cutensor.create_tensor_descriptor(c)
mode_c = cutensor.create_mode('x')
d = cutensor.reduction(
self.alpha, self.a, desc_a, self.mode_a,
self.beta, c, desc_c, mode_c
)
assert c is d
testing.assert_allclose(
self.alpha.item() * self.a_transposed.sum(axis=(1, 2)) +
self.beta.item() * c_orig,
d,
rtol=1e-6, atol=1e-6
)
extent = {'m': 196, 'h': 256, 'k': 64, 'v': 64}
a = cupy.random.random([extent[i] for i in mode_a])
c = cupy.random.random([extent[i] for i in mode_c])
a = a.astype(dtype)
c = c.astype(dtype)
desc_a = cutensor.create_tensor_descriptor(a)
desc_c = cutensor.create_tensor_descriptor(c)
alpha = 1.0
beta = 0.1
# rehearsal
c = cutensor.reduction(alpha, a, desc_a, mode_a, beta, c, desc_c, mode_c)
ev_start = stream.Event()
ev_end = stream.Event()
st = stream.Stream()
with st:
# measurement
ev_start.record()
c = cutensor.reduction(alpha, a, desc_a, mode_a, beta, c, desc_c, mode_c)
ev_end.record()
st.synchronize()
elapsed_ms = stream.get_elapsed_time(ev_start, ev_end)
transfer_byte = a.size * a.itemsize + c.size * c.itemsize
if beta != 0.0:
transfer_byte += c.size * c.itemsize
def time_reduction(self):
if self.mode == 'cute':
cutensor.reduction(1, self.array, self.desc_x, self.mode_x, 0,
self.out, self.desc_out, self.mode_out)
else:
cupy.sum(self.array, self.axis, None, self.out)
