def test_three_elements(self):
""" Make sure only 3D spaces can be successfully created. """
self.assertRaises(TypeError, space.initialize_space, (100, ))
self.assertRaises(TypeError, space.initialize_space, (100, 2))
self.assertRaises(TypeError, space.initialize_space, (100, 2, 3, 4))
self.assertRaises(TypeError, space.initialize_space, (100, 2, 3, 4, 5))
space.initialize_space((100, 2, 3))
def test_grid_shape(self):
""" Make sure the grid shapes of the exec configs are correct. """
tot_threads = lambda gs, bs: (gs[0] * bs[0], gs[1] * bs[1])
for case in self.cases:
space.initialize_space(case['shape'])
z = Out(case['dtype'])
fun = Kernel('', ('z', 'out', z.dtype))
for cfg in fun.exec_configs:
for ind in range(2):
self.assertTrue(cfg['grid_shape'][ind] * \
cfg['block_shape'][ind] >= \
case['shape'][ind+1])
self.assertTrue((cfg['grid_shape'][ind]-1) * \
cfg['block_shape'][ind] < \
case['shape'][ind+1])
# One padded case.
fun = Kernel('', ('z', 'out', z.dtype), padding=(1,2,3,4))
pad = [3, 7]
for cfg in fun.exec_configs:
for ind in range(2):
self.assertTrue(cfg['grid_shape'][ind] * \
(cfg['block_shape'][ind]-pad[ind]) >= \
case['shape'][ind+1])
self.assertTrue((cfg['grid_shape'][ind]-1) * \
(cfg['block_shape'][ind]-pad[ind]) < \
case['shape'][ind+1])
def test_three_elements(self):
""" Make sure only 3D spaces can be successfully created. """
self.assertRaises(TypeError, space.initialize_space, (100,))
self.assertRaises(TypeError, space.initialize_space, (100, 2))
self.assertRaises(TypeError, space.initialize_space, (100, 2, 3, 4))
self.assertRaises(TypeError, space.initialize_space, (100, 2, 3, 4, 5))
space.initialize_space((100, 2, 3))
def test_padded_kernel(self):
""" Implement a simple padded kernel. """
for case in self.cases:
# Form data to work on.
space.initialize_space(case['shape'])
x_np = comm.allreduce(np.random.randn(*case['shape']).astype(case['dtype']))
x = Grid(x_np, x_overlap=1)
s_np = comm.allreduce(np.random.randn(1).astype(case['dtype']))
s = Const(s_np)
z = Out(case['dtype'])
# Make a kernel.
code = Template("""
if (_in_local && _in_global) {
x(0,0,0) = s(0) * x(0,0,0);
z += a * x(0,0,0);
}
""").render()
fun = Kernel(code, \
('a', 'number', case['dtype']), \
('x', 'grid', x.dtype), \
('s', 'const', s.dtype, s.data.size), \
('z', 'out', z.dtype), \
padding=(1,1,1,1))
# Execute and check the result.
fun(case['dtype'](2), x, s, z)
gpu_sum = z.get()
cpu_sum = np.sum(2.0 * s_np * x_np)
err = abs(gpu_sum - cpu_sum) / abs(cpu_sum)
# print case, err
if case['dtype'] in (np.float32, np.complex64):
self.assertTrue(err < 1e-2, (case, err))
else:
self.assertTrue(err < 1e-6, (case, err))
def test_batch_sum(self):
""" Make sure batch summing works. """
num_outs = 3
for case in self.cases:
space.initialize_space(case['shape'])
x = [Out(case['dtype'], op='sum') for k in range(num_outs)]
x_cpu_data = [np.random.randn(*case['shape'][1:])\
.astype(case['dtype']) for k in range(num_outs)]
if case['dtype'] in (np.complex64, np.complex128):
for k in range(num_outs):
x_cpu_data[k] = (1 + 1j) * x_cpu_data[k]
res_gold = []
for k in range(num_outs):
x[k].data.set(x_cpu_data[k])
res_gold.append(comm.allreduce(np.sum(x_cpu_data[k].flatten())))
batch_reduce(*x)
res_gpu = [x_indiv.get() for x_indiv in x]
for k in range(num_outs):
err = abs(res_gold[k] - res_gpu[k]) / abs(res_gold[k])
if case['dtype'] in (np.float32, np.complex64):
self.assertTrue(err < 1e-3)
else:
self.assertTrue(err < 1e-10)
def test_init(self):
""" Test initialize function. """
for case in self.cases:
untype_array = np.zeros(case['shape']).astype(np.int)
space.initialize_space(case['shape'])
self.assertRaises(TypeError, Const, np.int)
self.assertRaises(TypeError, Const, untype_array)
self.assertRaises(TypeError, Const, 'string')
Const(np.random.randn(10).astype(case['dtype']))
def test_get_info(self):
""" Test the get_space_info function. """
# # We should get an error if we haven't initialized a space yet.
# self.assertRaises(TypeError, space.get_space_info)
shape = (100, 2, 3)
space.initialize_space(shape)
info = space.get_space_info()
self.assertEqual(info['shape'], shape)
def test_get_info(self):
""" Test the get_space_info function. """
# # We should get an error if we haven't initialized a space yet.
# self.assertRaises(TypeError, space.get_space_info)
shape = (100,2,3)
space.initialize_space(shape)
info = space.get_space_info()
self.assertEqual(info['shape'], shape)
def test_partition(self):
""" Make sure the x_ranges span the entire space without any gaps. """
shapes = ((200, 30, 10), (33, 10, 10), (130, 5, 5), (111, 2, 2))
for shape in shapes:
space.initialize_space(shape)
x = comm.gather(space.get_space_info()['x_range'])
if comm.Get_rank() == 0:
self.assertEqual(x[0][0], 0)
self.assertEqual(x[-1][-1], space.get_space_info()['shape'][0])
for k in range(len(x) - 1):
self.assertEqual(x[k][1], x[k + 1][0])
def test_init(self):
""" Test initialize function. """
for case in self.cases:
untype_array = np.zeros(case['shape']).astype(np.int)
space.initialize_space(case['shape'])
self.assertRaises(TypeError, Out, np.int)
self.assertRaises(TypeError, Out, untype_array)
self.assertRaises(TypeError, Out, 'string')
self.assertRaises(TypeError, Out, np.complex128, op='bad')
Out(case['dtype'])
Out(case['dtype'], op='sum')
def test_partition(self):
""" Make sure the x_ranges span the entire space without any gaps. """
shapes = ((200,30,10), (33,10,10), (130,5,5), (111,2,2))
for shape in shapes:
space.initialize_space(shape)
x = comm.gather(space.get_space_info()['x_range'])
if comm.Get_rank() == 0:
self.assertEqual(x[0][0], 0)
self.assertEqual(x[-1][-1], space.get_space_info()['shape'][0])
for k in range(len(x)-1):
self.assertEqual(x[k][1], x[k+1][0])
def test_init(self):
""" Just make sure we can initialize the kernel. """
for case in self.cases:
# Form data to work on.
space.initialize_space(case['shape'])
x_np = np.random.randn(*case['shape']).astype(case['dtype'])
x = Grid(x_np)
fun = Kernel('', ('x', 'grid', x.dtype))
fun = Kernel('', ('x', 'grid', x.dtype), shape_filter='all')
fun = Kernel('', ('x', 'grid', x.dtype), shape_filter='skinny')
fun = Kernel('', ('x', 'grid', x.dtype), shape_filter='square')
self.assertRaises(TypeError, Kernel, '', ('x', 'grid', x.dtype), \
shape_filter1='all')
self.assertRaises(TypeError, Kernel, '', ('x', 'grid', x.dtype), \
shape_filter='blah')
def test_to_and_from_gpu(self):
""" Make sure we can load and unload data off the gpu. """
shape = (100,100,100)
d = Data()
space.initialize_space(shape)
for dtype in self.valid_dtypes:
# Create data to load.
d_cpu = np.random.randn(*shape).astype(dtype)
if dtype in (np.complex64, np.complex128):
d_cpu = (1 + 1j) * d_cpu
# Load and retrieve.
d.to_gpu(d_cpu)
self.assertTrue((d_cpu == d.get()).all())
def test_init(self):
""" Test initialize function. """
for case in self.cases:
unfit_array = np.zeros(10)
untype_array = np.zeros(case['shape']).astype(np.int)
space.initialize_space(case['shape'])
self.assertRaises(TypeError, Grid, np.int)
self.assertRaises(TypeError, Grid, unfit_array)
self.assertRaises(TypeError, Grid, untype_array)
self.assertRaises(TypeError, Grid, 'string')
self.assertRaises(TypeError, Grid, np.float32, x_overlap='a')
self.assertRaises(TypeError, Grid, np.float32, x_overlap=2.2)
self.assertRaises(TypeError, Grid, np.float32, x_overlap=-2)
Grid(np.random.randn(*case['shape']).astype(case['dtype']))
Grid(case['dtype'])
Grid(np.random.randn(*case['shape']).astype(case['dtype']), x_overlap=1)
Grid(case['dtype'], x_overlap=2)
def test_kernel_self_opt(self):
""" Make sure the kernel settles on the fastest configuration. """
for case in (self.cases[0],):
space.initialize_space(case['shape'])
z = Out(case['dtype'])
fun = Kernel('', ('z', 'out', z.dtype), shape_filter='square')
# Run through all configurations.
hist = []
while fun.exec_configs:
hist.append(fun(z))
# Find fastest config, early-bird wins ties.
best_time, best_cfg = min(hist, key=lambda x: x[0])
time, next_cfg = fun(z) # Run once more, should use fastest configuration.
# Make sure we have chosen the fastest configuration.
self.assertEqual(best_cfg, next_cfg)
def test_sum(self):
""" Make sure summing works. """
for case in self.cases:
space.initialize_space(case['shape'])
x = Out(case['dtype'], op='sum')
x_cpu_data = np.random.randn(*case['shape'][1:]).astype(case['dtype'])
if case['dtype'] in (np.complex64, np.complex128):
x_cpu_data = (1 + 1j) * x_cpu_data
x.data.set(x_cpu_data)
res_gold = comm.allreduce(np.sum(x_cpu_data.flatten()))
x.reduce()
err = abs(res_gold - x.get()) / abs(res_gold)
if case['dtype'] in (np.float32, np.complex64):
self.assertTrue(err < 1e-3)
else:
self.assertTrue(err < 1e-10)
def test_simple_example(self):
""" Implement a simple kernel. """
# Form data to work on.
shape = (100, 100, 100)
space.initialize_space(shape)
x = Grid((1 + 1j) * np.ones(shape).astype(np.complex128))
z = Out(np.float64)
# Make a kernel.
code = """
if (_in_global) { // Need to be in the space.
z += real(x(0,0,0)) + imag(x(0,0,0));
} """
fun = Kernel(code, \
('x', 'grid', x.dtype), \
('z', 'out', z.dtype))
# Execute and check the result.
fun(x, z)
gpu_sum = z.get()
def test_simple_example(self):
""" Implement a simple kernel. """
# Form data to work on.
shape = (100,100,100)
space.initialize_space(shape)
x = Grid((1 + 1j) * np.ones(shape).astype(np.complex128))
z = Out(np.float64)
# Make a kernel.
code = """
if (_in_global) { // Need to be in the space.
z += real(x(0,0,0)) + imag(x(0,0,0));
} """
fun = Kernel(code, \
('x', 'grid', x.dtype), \
('z', 'out', z.dtype))
# Execute and check the result.
fun(x, z)
gpu_sum = z.get()
def test_recover(self):
""" Make sure we can store and retrieve information from the GPU. """
for case in self.cases:
space.initialize_space(case['shape'])
data = np.random.randn(*case['shape']).astype(case['dtype'])
cpu_data = np.empty_like(data)
comm.Allreduce(data, cpu_data)
g = Grid(cpu_data)
gpu_data = g.get()
if comm.Get_rank() == 0:
self.assertTrue((cpu_data == gpu_data).all())
# Test with-overlap cases as well.
for k in range(1, 3):
g = Grid(cpu_data, x_overlap=k)
gpu_data = g.get()
if comm.Get_rank() == 0:
self.assertTrue((cpu_data == gpu_data).all())
cpu_raw = get_cpu_raw(cpu_data, k)
self.assertTrue((cpu_raw == g._get_raw()).all())
def test_simple_kernel(self):
""" Implement a simple kernel. """
for case in self.cases:
# Form data to work on.
space.initialize_space(case['shape'])
x_np = comm.allreduce(np.random.randn(*case['shape']).astype(case['dtype']))
x = Grid(x_np, x_overlap=2)
s_np = comm.allreduce(np.random.randn(case['shape'][0],1,1).astype(case['dtype']))
s = Const(s_np)
z = Out(case['dtype'])
# Make a kernel.
code = Template("""
if (_in_local && _in_global) {
z += a * s(_X) * x(0,0,0);
// z += a * x(0,0,0);
}
""").render()
fun = Kernel(code, \
('a', 'number', case['dtype']), \
('x', 'grid', x.dtype), \
('s', 'const', s.dtype), \
('z', 'out', z.dtype), \
shape_filter='all')
# Execute and check the result.
# fun()
while fun.exec_configs:
# for k in range(40):
fun(case['dtype'](2.0), x, s, z)
# fun(case['dtype'](2.0), x, z)
gpu_sum = z.get()
cpu_sum = np.sum(2 * s_np * x_np)
# cpu_sum = np.sum(2 * x_np)
err = abs(gpu_sum - cpu_sum) / abs(cpu_sum)
if case['dtype'] in (np.float32, np.complex64):
self.assertTrue(err < 1e-2, (case, err))
else:
self.assertTrue(err < 1e-6, (case, err))
def test_synchronize(self):
""" Make sure that we can make the overlap spaces accurate. """
for case in self.cases:
space.initialize_space(case['shape'])
data = np.random.randn(*case['shape']).astype(case['dtype'])
cpu_data = np.empty_like(data)
comm.Allreduce(data, cpu_data)
g = Grid(case['dtype'])
self.assertRaises(TypeError, g.synchronize) # No overlap.
# Test with-overlap cases as well.
for k in range(1, 4):
g = Grid(case['dtype'], x_overlap=k)
# Overwrite entire grid
data = np.random.randn(*case['shape']).astype(case['dtype'])
cpu_data = np.empty_like(data)
comm.Allreduce(data, cpu_data)
cpu_raw_bad = get_cpu_raw(cpu_data, k)
cpu_raw_bad[:k,:,:] += 1 # Mess up padding areas.
cpu_raw_bad[-k:,:,:] += 1
drv.memcpy_htod(g.data.ptr, cpu_raw_bad)
# Prove that the data is not synchronized at this time.
cpu_raw = get_cpu_raw(cpu_data, k)
xx = case['shape'][0]
gd = g._get_raw()
self.assertTrue((gd[:k,:,:] != cpu_raw[:k,:,:]).all())
self.assertTrue((gd[-k:,:,:] != cpu_raw[-k:,:,:]).all())
g.synchronize() # Synchronize the overlapping data.
# Make sure that the overlap data is accurate.
gd = g._get_raw()
self.assertTrue((gd[:k,:,:] == cpu_raw[:k,:,:]).all())
self.assertTrue((gd[-k:,:,:] == cpu_raw[-k:,:,:]).all())
comm.Barrier() # Wait for other mpi nodes to finish.
def test_ecc_disabled(self):
""" Make sure ECC is disabled. """
space.initialize_space((100, 2, 3))
self.assertTrue(space.get_space_info()['ecc_enabled'] == False, \
'ECC enabled! Should be disabled for best performance.')
def test_ecc_disabled(self):
""" Make sure ECC is disabled. """
space.initialize_space((100, 2, 3))
self.assertTrue(space.get_space_info()['ecc_enabled'] == False, \
'ECC enabled! Should be disabled for best performance.')