def AddStencilRuntime(self, device_name, device_count,
north, south, east, west, data_width,
data_height, type_in, type_out, source,
max_wg_size, wg_c, wg_r, runtime):
"""
Add a new stencil runtime.
Args:
device_name (str): The name of the execution device.
device_count (int): The number of execution devices.
north (int): The stencil shape north direction.
south (int): The stencil shape south direction.
east (int): The stencil shape east direction.
west (int): The stencil shape west direction.
data_width (int): The number of columns of data.
data_height (int): The number of rows of data.
type_in (str): The input data type.
type_out (str): The output data type.
source (str): The stencil kernel source code.
max_wg_size (int): The maximum kernel workgroup size.
wg_c (int): The workgroup size used (columns).
wg_r (int): The workgroup size used (rows).
runtime (double): The measured runtime in milliseconds.
"""
# Parse arguments.
device_name = util.parse_str(device_name)
device_count = int(device_count)
north = int(north)
south = int(south)
east = int(east)
west = int(west)
data_width = int(data_width)
data_height = int(data_height)
type_in = util.parse_str(type_in)
type_out = util.parse_str(type_out)
source = util.parse_str(source)
max_wg_size = int(max_wg_size)
wg_c = int(wg_c)
wg_r = int(wg_r)
runtime = float(runtime)
# Lookup IDs
device = self.db.device_id(device_name, device_count)
kernel = self.db.kernel_id(north, south, east, west,
max_wg_size, source)
dataset = self.db.datasets_id(data_width, data_height,
type_in, type_out)
scenario = self.db.scenario_id(device, kernel, dataset)
params = self.db.params_id(wg_c, wg_r)
# Add entry into runtimes table.
self.db.add_runtime(scenario, params, runtime)
self.db.commit()
io.debug(("AddStencilRuntime({scenario}, {params}, {runtime})"
.format(scenario=scenario[:8], params=params,
runtime=runtime)))
def RequestTrainingStencilParams(
self,
device_name,
device_count,
north,
south,
east,
west,
data_width,
data_height,
type_in,
type_out,
source,
max_wg_size,
):
"""
Request training parameter values for a SkelCL stencil operation.
Determines the parameter values to use for a SkelCL stencil
operation by iterating over the space of parameter values.
Args:
device_name (str): The name of the execution device.
device_count (int): The number of execution devices.
north (int): The stencil shape north direction.
south (int): The stencil shape south direction.
east (int): The stencil shape east direction.
west (int): The stencil shape west direction.
data_width (int): The number of columns of data.
data_height (int): The number of rows of data.
type_in (str): The input data type.
type_out (str): The output data type.
max_wg_size (int): The maximum kernel workgroup size.
source (str): The stencil kernel source code.
Returns:
A tuple of work group size values, e.g.
(16,32)
"""
# Parse arguments.
device_name = util.parse_str(device_name)
device_count = int(device_count)
north = int(north)
south = int(south)
east = int(east)
west = int(west)
data_width = int(data_width)
data_height = int(data_height)
type_in = util.parse_str(type_in)
type_out = util.parse_str(type_out)
source = util.parse_str(source)
max_wg_size = int(max_wg_size)
# Get the next scenario ID to train on.
wg = training.random_wg_value(max_wg_size)
return wg
def RequestStencilParams(self, device_name, device_count, north, south,
east, west, data_width, data_height, type_in,
type_out, source, max_wg_size):
"""
Request parameter values for a SkelCL stencil operation.
Determines the parameter values to use for a SkelCL stencil
operation, using a machine learning classifier to predict the
optimal parameter values given a set of features determined
from the arguments.
Args:
device_name (str): The name of the execution device.
device_count (int): The number of execution devices.
north (int): The stencil shape north direction.
south (int): The stencil shape south direction.
east (int): The stencil shape east direction.
west (int): The stencil shape west direction.
data_width (int): The number of columns of data.
data_height (int): The number of rows of data.
type_in (str): The input data type.
type_out (str): The output data type.
max_wg_size (int): The maximum kernel workgroup size.
source (str): The stencil kernel source code.
Returns:
A tuple of work group size values, e.g.
(16,32)
"""
start_time = time.time()
# Parse arguments.
device_name = util.parse_str(device_name)
device_count = int(device_count)
north = int(north)
south = int(south)
east = int(east)
west = int(west)
data_width = int(data_width)
data_height = int(data_height)
source = util.parse_str(source)
max_wg_size = int(max_wg_size)
# TODO: Perform feature extraction & classification
wg = (64, 32)
end_time = time.time()
io.debug(("RequestStencilParams() -> "
"({c}, {r}) [{t:.3f}s]".format(c=wg[0],
r=wg[1],
t=end_time - start_time)))
return wg
def RequestStencilParams(self, device_name, device_count,
north, south, east, west, data_width,
data_height, type_in, type_out, source,
max_wg_size):
"""
Request parameter values for a SkelCL stencil operation.
Determines the parameter values to use for a SkelCL stencil
operation, using a machine learning classifier to predict the
optimal parameter values given a set of features determined
from the arguments.
Args:
device_name (str): The name of the execution device.
device_count (int): The number of execution devices.
north (int): The stencil shape north direction.
south (int): The stencil shape south direction.
east (int): The stencil shape east direction.
west (int): The stencil shape west direction.
data_width (int): The number of columns of data.
data_height (int): The number of rows of data.
type_in (str): The input data type.
type_out (str): The output data type.
max_wg_size (int): The maximum kernel workgroup size.
source (str): The stencil kernel source code.
Returns:
A tuple of work group size values, e.g.
(16,32)
"""
start_time = time.time()
# Parse arguments.
device_name = util.parse_str(device_name)
device_count = int(device_count)
north = int(north)
south = int(south)
east = int(east)
west = int(west)
data_width = int(data_width)
data_height = int(data_height)
source = util.parse_str(source)
max_wg_size = int(max_wg_size)
# TODO: Perform feature extraction & classification
wg = (64, 32)
end_time = time.time()
io.debug(("RequestStencilParams() -> "
"({c}, {r}) [{t:.3f}s]"
.format(c=wg[0], r=wg[1], t=end_time - start_time)))
return wg
def RequestTrainingStencilParams(self, device_name, device_count,
north, south, east, west, data_width,
data_height, type_in, type_out, source,
max_wg_size):
"""
Request training parameter values for a SkelCL stencil operation.
Determines the parameter values to use for a SkelCL stencil
operation by iterating over the space of parameter values.
Args:
device_name (str): The name of the execution device.
device_count (int): The number of execution devices.
north (int): The stencil shape north direction.
south (int): The stencil shape south direction.
east (int): The stencil shape east direction.
west (int): The stencil shape west direction.
data_width (int): The number of columns of data.
data_height (int): The number of rows of data.
type_in (str): The input data type.
type_out (str): The output data type.
max_wg_size (int): The maximum kernel workgroup size.
source (str): The stencil kernel source code.
Returns:
A tuple of work group size values, e.g.
(16,32)
"""
# Parse arguments.
device_name = util.parse_str(device_name)
device_count = int(device_count)
north = int(north)
south = int(south)
east = int(east)
west = int(west)
data_width = int(data_width)
data_height = int(data_height)
type_in = util.parse_str(type_in)
type_out = util.parse_str(type_out)
source = util.parse_str(source)
max_wg_size = int(max_wg_size)
# Get the next scenario ID to train on.
wg = training.random_wg_value(max_wg_size)
return wg
def test_parse_str(self):
self._test("abc", util.parse_str("abc"))
self._test("a\nc", util.parse_str("a\nc"))
self._test("123", util.parse_str("123"))
