junrushao1994 commented on a change in pull request #5962: URL: https://github.com/apache/incubator-tvm/pull/5962#discussion_r450399570
########## File path: python/tvm/ansor/workload_registry.py ########## @@ -0,0 +1,170 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. + +""" +Workload registration and serialization. + +We use a json string to represent a workload (a computation graph). +The format of the string is `[func_name, [args...]]`. +The dag should be the return value of this `func_name(*args)`. + +Rationale: The workload is actually a compute dag defined by tvm dsl. But serializing compute dags +and matching them efficiently is not easy. Therefore, we use the above string to encode a compute +dag. +These strings are efficient for serialization/matching and wont' be too long. Review comment: ```suggestion These strings are efficient for serialization/matching and won't be too long. ``` ########## File path: python/tvm/ansor/workload_registry.py ########## @@ -0,0 +1,170 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. + +""" +Workload registration and serialization. + +We use a json string to represent a workload (a computation graph). +The format of the string is `[func_name, [args...]]`. +The dag should be the return value of this `func_name(*args)`. + +Rationale: The workload is actually a compute dag defined by tvm dsl. But serializing compute dags +and matching them efficiently is not easy. Therefore, we use the above string to encode a compute +dag. +These strings are efficient for serialization/matching and wont' be too long. +When we need the dag, we decode the string and call the function, which will return the dag. +""" + +import pickle +import json + +import tvm._ffi +from .utils import serialize_args, deserialize_args + +WORKLOAD_FUNC_REGISTRY = {} + + +def register_workload(func): + """ Register a workload by generation function. + + The input function should take hashable and jsonable arguments + (int, float, tuple of int, tvm.tensor.Tensor, ...) and return a list of tvm.tensor.Tensor. + + Parameters + ---------- + func : Function + The generation function that returns the compute declaration Tensors. + + Examples + -------- + @ansor.register_workload + def matmul(N, M, K): + A = te.placeholder((N, K), name='A') + B = te.placeholder((K, M), name='B') + k = te.reduce_axis((0, K), name='k') + C = te.compute((N, M), lambda i, j: tvm.sum(A[i][k] * B[k][j], axis=[k]), name='C') + return [A, B, C] + """ + assert callable(func) + func_name = func.__name__ + if func_name in WORKLOAD_FUNC_REGISTRY: + raise RuntimeError('%s has been registered already' % func_name) + + WORKLOAD_FUNC_REGISTRY[func_name] = func + return func + + +def make_workload_key(func, args): + """ make a workload key from function and arguments. + + Parameters + ---------- + func : Union[Function, str] + The function that returns the compute declaration Tensors. + Can be the a function or the function name. + args : Args + The args of the function. + + Returns + ------- + workload_key : Str + The workload key of the function. + """ + if callable(func): + func_name = func.__name__ Review comment: Let write a helper function to get name from a given python func. ########## File path: python/tvm/ansor/workload_registry.py ########## @@ -0,0 +1,170 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. + +""" +Workload registration and serialization. + +We use a json string to represent a workload (a computation graph). +The format of the string is `[func_name, [args...]]`. +The dag should be the return value of this `func_name(*args)`. + Review comment: I am trying to understand this...Could you check if the following understanding is correct? A. "Workload" is the workload ansor wants to tune. For exmaple, it can be a matmul, whose inputs are n, m, k (matrix dimensions). B. The workload key is used to describe the entire workload, i.e. the entire workload can be completely described using registered function and its inputs. For example, using `n`, `m`, `k`, we can reconstruct the matmul function by re-creating the two corresponding tensors, build a ComputeOp and then return its result. C. The computational dag can be created using a workload key. ########## File path: python/tvm/ansor/workload_registry.py ########## @@ -0,0 +1,170 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. + +""" +Workload registration and serialization. + +We use a json string to represent a workload (a computation graph). +The format of the string is `[func_name, [args...]]`. +The dag should be the return value of this `func_name(*args)`. + +Rationale: The workload is actually a compute dag defined by tvm dsl. But serializing compute dags +and matching them efficiently is not easy. Therefore, we use the above string to encode a compute +dag. +These strings are efficient for serialization/matching and wont' be too long. +When we need the dag, we decode the string and call the function, which will return the dag. +""" + +import pickle +import json + +import tvm._ffi +from .utils import serialize_args, deserialize_args + +WORKLOAD_FUNC_REGISTRY = {} + + +def register_workload(func): + """ Register a workload by generation function. + + The input function should take hashable and jsonable arguments + (int, float, tuple of int, tvm.tensor.Tensor, ...) and return a list of tvm.tensor.Tensor. + + Parameters + ---------- + func : Function + The generation function that returns the compute declaration Tensors. + + Examples + -------- + @ansor.register_workload + def matmul(N, M, K): + A = te.placeholder((N, K), name='A') + B = te.placeholder((K, M), name='B') + k = te.reduce_axis((0, K), name='k') + C = te.compute((N, M), lambda i, j: tvm.sum(A[i][k] * B[k][j], axis=[k]), name='C') + return [A, B, C] + """ + assert callable(func) + func_name = func.__name__ Review comment: This can be problematic when codebase grows, because of potential naming conflicts across different python files. IMHO a better solution would be 1) Use `func.__qualname__` instead 2) Allow users to provide a customized name, like what is done in `tvm._ffi.register_func` ########## File path: python/tvm/ansor/workload_registry.py ########## @@ -0,0 +1,170 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. + +""" +Workload registration and serialization. + +We use a json string to represent a workload (a compute dag). +The format of the string is `[func_name, [args...]]`. +The dag should be the return value of this `func_name(*args)`. + +Rationale: The workload is actually a compute dag defined by tvm dsl. But serializing compute dags +and matching them efficiently is not easy. Therefore, we use the above string to encode a compute +dag. +These strings are efficient for serialization/matching and wont' be too long. +When we need the dag, we decode the string and call the function, which will return the dag. +""" + +import pickle +import json + +import tvm._ffi +from .utils import serialize_args, deserialize_args + +WORKLOAD_FUNC_REGISTRY = {} + + +def register_workload_by_func(func): + """ Register a workload by generation function. Review comment: I found "generation function" confusing too...Not a native speaker who may provide professional suggestions, but I guess it would be slightly better to say "Register a function who generates a certain workload" ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
