merrymercy commented on a change in pull request #5962: URL: https://github.com/apache/incubator-tvm/pull/5962#discussion_r449732736
########## File path: python/tvm/ansor/loop_state.py ########## @@ -0,0 +1,221 @@ +# 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. +# pylint: disable=unused-import + +""" +The definition of the "state" in search. + +Each LoopState corresponds to a specific schedule for its target ComputeDAG. +A LoopState consists of: 1. a current loop structure; 2. a history of transformations used to +construct the loop structure. +The loop structure keeps a preview of how the schedule will finally look like after lowering the +current state (e.g. number of iterators, the extent of each iterator, the compute_at locations ...). +During the schedule search process, the loop structure can provide search policy with necessary +information on how to perform further operations with the current state. +The transform history is a sequence of TransformStep which will finally be mapped to schedule +primitives. The steps can also be used for serialization of a state. + +The LoopState can be seen as a lightweight loop structure IR specifically for schedule search. +We don't use the existing TVM IR but to extend a new structure on it is because: +1. We want fast incremental change to the loop structures, search policy needs to get the immediate +loop structures update rather than after TVM lowering; +2. We want serializable transform history for replay, backtracking, and mutation; +3. We may create some macro schedule primitives that represent the combination of several +TVM schedule primitives. + +When the search is complete, we will lower the state to TVM IR with TVM's schedule primitives. +Since we share a lot of common objects during search, the transformation is implemented in +copy on write style. All objects are immutable, which is similar to TVM IR. +""" Review comment: ```suggestion The definition of the "state" in search. Each LoopState corresponds to a schedule for its ComputeDAG. A LoopState consists of: 1. a current loop structure; 2. a list of transformation steps used to construct the loop structure. The loop structure keeps a preview of how the schedule will finally look like after lowering the current state (e.g. number of iterators, the extent of each iterator, the compute_at locations ...). During the schedule search process, the loop structure can provide search policy with necessary information on how to manipulate the current state. The transform history is a sequence of `TransformStep` which will finally be mapped to TVM schedule primitives. The steps can also be used for the serialization of a state. The LoopState can be seen as a lightweight loop structure IR specifically for schedule search. We don't use the existing TVM IR but to extend a new structure on it is because: 1. We want fast incremental change to the loop structures. The search policy needs to get the immediate loop structures update rather than after TVM lowering; 2. We want serializable transform history for replay, backtracking, and mutation; 3. We may create some macro schedule primitives that represent the combination of several TVM schedule primitives. When the search is complete, we will lower the state to TVM IR with TVM's schedule primitives. Since we share a lot of common objects during search, the transformation is implemented in copy on write style. All objects are immutable, which is similar to TVM IR. """ ``` Also, propagate the changes to c++ files. ---------------------------------------------------------------- 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]
