def test_from_api_id(mock_backend_factory):
for api_id in all_api_ids():
with pytest.raises(ImportError):
API.from_api_id(api_id)
mock_backend_factory.mock(api_id)
api = API.from_api_id(api_id)
assert api.id == api_id
def test_eq(mock_backend_factory):
api_id0 = all_api_ids()[0]
api_id1 = all_api_ids()[1]
mock_backend_factory.mock(api_id0)
mock_backend_factory.mock(api_id1)
api0_v1 = API.from_api_id(api_id0)
api0_v2 = API.from_api_id(api_id0)
api1 = API.from_api_id(api_id1)
assert api0_v1 is not api0_v2 and api0_v1 == api0_v2
assert api0_v1 != api1
def test_hash(mock_backend_factory):
api_id0 = all_api_ids()[0]
api_id1 = all_api_ids()[1]
mock_backend_factory.mock(api_id0)
mock_backend_factory.mock(api_id1)
api0 = API.from_api_id(api_id0)
api1 = API.from_api_id(api_id1)
d = {api0: 0, api1: 1}
assert d[api0] == 0
assert d[api1] == 1
def test_builtin_globals(mock_backend_pycuda):
mock_backend_pycuda.add_devices([
PyCUDADeviceInfo(max_threads_per_block=1024),
PyCUDADeviceInfo(max_threads_per_block=512)
])
source_template = DefTemplate.from_string(
'mock_source', [], """
KERNEL void test()
{
int max_total_local_size = ${device_params.max_total_local_size};
}
""")
api = API.from_api_id(mock_backend_pycuda.api_id)
context = Context.from_devices(
[api.platforms[0].devices[0], api.platforms[0].devices[1]])
src = MockDefTemplate(kernels=[MockKernel('test', [None])],
source_template=source_template)
program = Program(context.devices, src)
assert 'max_total_local_size = 1024' in program.sources[
context.devices[0]].source
assert 'max_total_local_size = 512' in program.sources[
context.devices[1]].source
def test_all(mock_backend):
mock_backend.add_devices(['Device2', 'Device3'])
api = API.from_api_id(mock_backend.api_id)
platform = Platform.from_index(api, 0)
devices = Device.all(platform)
device_names = {device.name for device in devices}
assert device_names == {'Device2', 'Device3'}
def test_bound_multi_device_creation(mock_backend_pyopencl):
mock_backend_pyopencl.add_platform_with_devices(
'Platform1', ['Device1', 'Device2', 'Device3'])
api = API.from_api_id(mock_backend_pyopencl.api_id)
platform = api.platforms[0]
devices = platform.devices[:]
context = Context.from_devices(devices)
context2 = Context.from_devices(devices)
with pytest.raises(
ValueError,
match=
"All devices in a multi-device must belong to the same context"):
BoundMultiDevice.from_bound_devices(
[context.devices[0], context2.devices[1]])
with pytest.raises(ValueError,
match="All devices in a multi-device must be distinct"):
context.devices[[1, 1]]
sub_device = context.devices[[2, 1]]
assert len(sub_device) == 2
assert sub_device[0].name == 'Device3'
assert sub_device[1].name == 'Device2'
def test_device_parameters_cuda(mock_backend_pycuda):
device_info = PyCUDADeviceInfo(max_threads_per_block=512,
max_block_dim_x=512,
max_block_dim_y=512,
max_block_dim_z=32,
max_grid_dim_x=2**16,
max_grid_dim_y=2**16,
max_grid_dim_z=256,
warp_size=7,
max_shared_memory_per_block=58 * 1024,
multiprocessor_count=3,
compute_capability=10)
mock_backend_pycuda.add_devices([device_info])
api = API.from_api_id(mock_backend_pycuda.api_id)
d = api.platforms[0].devices[0]
assert d.params.type == DeviceType.GPU
assert d.params.max_total_local_size == device_info.max_threads_per_block
assert d.params.max_local_sizes == (device_info.max_block_dim_x,
device_info.max_block_dim_y,
device_info.max_block_dim_z)
assert d.params.warp_size == device_info.warp_size
assert d.params.max_num_groups == (device_info.max_grid_dim_x,
device_info.max_grid_dim_y,
device_info.max_grid_dim_z)
assert d.params.local_mem_size == device_info.max_shared_memory_per_block
assert d.params.local_mem_banks == 32
assert d.params.compute_units == device_info.multiprocessor_count
def test_flags(mock_backend_pyopencl):
backend = mock_backend_pyopencl
normal_flags = backend.pyopencl.mem_flags.READ_WRITE
special_flags = normal_flags | backend.pyopencl.mem_flags.ALLOC_HOST_PTR
backend.add_platform_with_devices('Apple', ['GeForce', 'Foo', 'Bar'])
backend.add_platform_with_devices('Baz', ['GeForce', 'Foo', 'Bar'])
# Multi-device on Apple platform with one of the devices being GeForce: need special Buffer flags
api = API.from_api_id(backend.api_id)
context = Context.from_devices([api.platforms[0].devices[0], api.platforms[0].devices[1]])
buf = Buffer.allocate(context.devices[0], 100)
assert buf._buffer_adapter.pyopencl_buffer.flags == special_flags
# None of the devices is GeForce
context = Context.from_devices([api.platforms[0].devices[1], api.platforms[0].devices[2]])
buf = Buffer.allocate(context.devices[0], 100)
assert buf._buffer_adapter.pyopencl_buffer.flags == normal_flags
# Only one device
context = Context.from_devices([api.platforms[0].devices[0]])
buf = Buffer.allocate(context.device, 100)
assert buf._buffer_adapter.pyopencl_buffer.flags == normal_flags
# Not an Apple platform
context = Context.from_devices([api.platforms[1].devices[0], api.platforms[1].devices[1]])
buf = Buffer.allocate(context.devices[0], 100)
assert buf._buffer_adapter.pyopencl_buffer.flags == normal_flags
def check_select_devices(mock_stdin,
mock_backend_factory,
capsys,
platforms_devices,
inputs=None,
**kwds):
# CUDA API has a single fixed platform, so using the OpenCL one
backend = mock_backend_factory.mock_pyopencl()
for platform_name, device_infos in platforms_devices:
platform = backend.add_platform_with_devices(platform_name,
device_infos)
if inputs is not None:
for line in inputs:
mock_stdin.line(line)
api = API.from_api_id(backend.api_id)
try:
devices = select_devices(api, **kwds)
assert mock_stdin.empty()
finally:
# Otherwise the output will be shown in the console
captured = capsys.readouterr()
return devices
def test_all_by_shortcut(mock_backend):
api_id = mock_backend.api_id
mock_backend.add_devices(['Device1', 'Device2'])
apis = API.all_by_shortcut(api_id.shortcut)
assert len(apis) == 1
assert apis[0].id == api_id
def test_all(mock_backend_factory):
api_id = all_api_ids()[0]
mock_backend_factory.mock(api_id)
apis = API.all_available()
assert len(apis) == 1
assert apis[0].id == api_id
def test_virtual_sizes_error_propagated(mock_backend_pycuda):
# Testing for PyCUDA backend only since mocked PyOpenCL backend does not have a way
# to set maximum global sizes (PyOpenCL devices don't have a corresponding parameter),
# and PyCUDA is enough to test the required code path.
device_info = PyCUDADeviceInfo(
max_threads_per_block=2**4,
max_block_dim_x=2**4,
max_block_dim_y=2**4,
max_block_dim_z=2**4,
max_grid_dim_x=2**10,
max_grid_dim_y=2**10,
max_grid_dim_z=2**8)
mock_backend_pycuda.add_devices([device_info])
api = API.from_api_id(mock_backend_pycuda.api_id)
device = api.platforms[0].devices[0]
context = Context.from_devices([device])
kernel = MockKernel('test', [None], max_total_local_sizes={0: 16})
src = MockDefTemplate(kernels=[kernel])
# Just enough to fit in the grid limits
multiply = StaticKernel(context.device, src, 'test', (2**14, 2**10, 2**8), (2**4, 1, 1))
# Global size is too large to fit on the device,
# so virtual size finding fails and the error is propagated to the user.
with pytest.raises(
VirtualSizeError,
match="Bounding global size \\(16384, 2048, 256\\) is too large"):
multiply = StaticKernel(context.device, src, 'test', (2**14, 2**11, 2**8), (2**4, 1, 1))
def test_all(mock_backend_pyopencl):
mock_backend_pyopencl.add_platform_with_devices('Platform1', ['Device1'])
mock_backend_pyopencl.add_platform_with_devices('Platform2', ['Device2'])
api = API.from_api_id(mock_backend_pyopencl.api_id)
platforms = Platform.all(api)
platform_names = {platform.name for platform in platforms}
assert platform_names == {'Platform1', 'Platform2'}
def test_from_index(mock_backend_pyopencl):
mock_backend_pyopencl.add_platform_with_devices('Platform1', ['Device1'])
mock_backend_pyopencl.add_platform_with_devices('Platform2', ['Device2'])
api = API.from_api_id(mock_backend_pyopencl.api_id)
platform = Platform.from_index(api, 1)
assert platform.name == 'Platform2'
def test_max_total_local_sizes(mock_backend):
mock_backend.add_devices(
["Device1", "Device2 - tag", "Device3 - tag", "Device4"])
api = API.from_api_id(mock_backend.api_id)
context = Context.from_criteria(api,
devices_num=2,
device_include_masks=["tag"])
# Providing max_total_local_sizes for all possible devices to make sure
# only the ones corresponding to the context will get picked up
kernel = MockKernel('test',
max_total_local_sizes={
0: 64,
1: 1024,
2: 512,
3: 128
})
src = MockDefTemplate(kernels=[kernel])
program = Program(context.devices, src)
# The indices here correspond to the devices in the context, not in the platform
assert program.kernel.test.max_total_local_sizes == {
context.devices[0]: 1024,
context.devices[1]: 512
}
def test_bound_multi_device_issubset(mock_backend_pyopencl):
mock_backend_pyopencl.add_platform_with_devices(
'Platform1', ['Device1', 'Device2', 'Device3'])
api = API.from_api_id(mock_backend_pyopencl.api_id)
context = Context.from_devices(api.platforms[0].devices[:])
assert context.devices[[2, 1]].issubset(context.devices)
def test_getitem(mock_backend_pyopencl):
api_id = mock_backend_pyopencl.api_id
mock_backend_pyopencl.add_platform_with_devices('Platform0', ['Device0'])
mock_backend_pyopencl.add_platform_with_devices('Platform1', ['Device1'])
api = API.from_api_id(api_id)
assert api.platforms[0].name == 'Platform0'
assert api.platforms[1].name == 'Platform1'
def test_device_parameters_opencl_cpu(mock_backend_pyopencl):
device_info = PyOpenCLDeviceInfo(type=DeviceType.CPU)
mock_backend_pyopencl.add_devices([device_info])
api = API.from_api_id(mock_backend_pyopencl.api_id)
d = api.platforms[0].devices[0]
assert d.params.local_mem_banks == 1
assert d.params.warp_size == 1
def test_params(mock_backend):
mock_backend.add_devices(['Device1'])
api = API.from_api_id(mock_backend.api_id)
platform = Platform.from_index(api, 0)
device = Device.from_index(platform, 0)
params1 = device.params
params2 = device.params
assert params1 is params2 # check caching
def test_all_by_shortcut_none(mock_backend_factory):
api_ids = all_api_ids()
for api_id in api_ids:
backend = mock_backend_factory.mock(api_id)
backend.add_devices(['Device1', 'Device2'])
apis = API.all_by_shortcut()
assert len(apis) == len(api_ids)
assert set(api.id for api in apis) == set(api_ids)
def test_all_by_masks(mock_backend_pyopencl):
mock_backend_pyopencl.add_platform_with_devices('foo-bar', ['Device1'])
mock_backend_pyopencl.add_platform_with_devices('bar-baz', ['Device2'])
mock_backend_pyopencl.add_platform_with_devices('foo-baz', ['Device3'])
api = API.from_api_id(mock_backend_pyopencl.api_id)
platforms = Platform.all_by_masks(api,
include_masks=['foo'],
exclude_masks=['bar'])
assert len(platforms) == 1
assert platforms[0].name == 'foo-baz'
def test_device_parameters_opencl_apple_cpu(mock_backend_pyopencl):
device_info = PyOpenCLDeviceInfo(type=DeviceType.CPU,
max_work_group_size=512)
mock_backend_pyopencl.add_platform_with_devices('Apple', [device_info])
api = API.from_api_id(mock_backend_pyopencl.api_id)
d = api.platforms[0].devices[0]
assert d.params.max_total_local_size == 1
assert d.params.max_local_sizes == (1, 1, 1)
def test_getitem(mock_backend_pyopencl):
mock_backend_pyopencl.add_platform_with_devices('Platform0', ['Device0'])
mock_backend_pyopencl.add_platform_with_devices('Platform1',
['Device1', 'Device2'])
api = API.from_api_id(mock_backend_pyopencl.api_id)
p1 = Platform.from_index(api, 1)
assert p1.devices[0].name == 'Device1'
assert p1.devices[1].name == 'Device2'
def test_from_devices(mock_backend):
mock_backend.add_devices(['Device2', 'Device3'])
api = API.from_api_id(mock_backend.api_id)
platform = api.platforms[0]
devices = platform.devices[:]
context = Context.from_devices(devices)
assert context.platform == platform
assert [device.as_unbound() for device in context.devices] == devices
def test_attributes(mock_backend):
mock_backend.add_devices(['Device1'])
api = API.from_api_id(mock_backend.api_id)
p = Platform.from_index(api, 0)
d = Device.from_index(p, 0)
assert d.platform == p
assert d.name == 'Device1'
assert d.shortcut == p.shortcut + ',0'
assert str(d) == 'device(' + d.shortcut + ')'
def test_device_parameters_cuda_1(mock_backend_pycuda):
# CUDA 1.x
device_info = PyCUDADeviceInfo(compute_capability=1)
mock_backend_pycuda.add_devices([device_info])
api = API.from_api_id(mock_backend_pycuda.api_id)
d = api.platforms[0].devices[0]
assert d.params.local_mem_banks == 16
def test_eq(mock_backend):
mock_backend.add_devices(['Device0', 'Device1'])
api = API.from_api_id(mock_backend.api_id)
platform = Platform.from_index(api, 0)
d0_v1 = Device.from_index(platform, 0)
d0_v2 = Device.from_index(platform, 0)
d1 = Device.from_index(platform, 1)
assert d0_v1 is not d0_v2 and d0_v1 == d0_v2
assert d0_v1 != d1
def test_hash(mock_backend):
mock_backend.add_devices(['Device0', 'Device1'])
api = API.from_api_id(mock_backend.api_id)
platform = Platform.from_index(api, 0)
d0 = Device.from_index(platform, 0)
d1 = Device.from_index(platform, 1)
d = {d0: 0, d1: 1}
assert d[d0] == 0
assert d[d1] == 1
def test_deactivate(mock_backend_pyopencl, mock_backend_pycuda):
mock_backend_pyopencl.add_platform_with_devices('Platform1', ['Device1'])
mock_backend_pycuda.add_devices(['Device1'])
api = API.from_api_id(mock_backend_pyopencl.api_id)
context = Context.from_devices(api.platforms[0].devices[0])
with pytest.raises(RuntimeError,
match="`deactivate\\(\\)` only works for CUDA API"):
context.deactivate()
backend_context = mock_backend_pycuda.pycuda_driver.Device(
0).make_context()
backend_context.pop()
api = API.from_api_id(mock_backend_pycuda.api_id)
context = Context.from_backend_contexts(backend_context,
take_ownership=True)
assert backend_context.is_stacked()
context.deactivate()
assert not backend_context.is_stacked()
def test_from_devices_different_platforms(mock_backend_pyopencl):
mock_backend_pyopencl.add_platform_with_devices('Platform1',
['Device1', 'Device2'])
mock_backend_pyopencl.add_platform_with_devices('Platform2',
['Device3', 'Device4'])
api = API.from_api_id(mock_backend_pyopencl.api_id)
with pytest.raises(ValueError,
match="All devices must belong to the same platform"):
Context.from_devices(
[api.platforms[0].devices[0], api.platforms[1].devices[0]])