yangjunpro commented on a change in pull request #6190: URL: https://github.com/apache/incubator-tvm/pull/6190#discussion_r468939081
########## File path: include/tvm/auto_scheduler/feature.h ########## @@ -0,0 +1,122 @@ +/* + * 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. + */ + +/*! + * \file auto_scheduler/feature.h + * \brief Feature extraction for the cost model. + * We extract one feature vector per BufferStoreNode statement in a TIR Stmt, + * so we call this feature as "Per Store" feature. + * The cost model also does prediction for each BufferStoreNode statement and aggregates + * the predictions as the whole score for a TVM IR (Stmt). + * + * The feature specification is defined by `src/auto_scheduler/feature.cc:: FeatureSet` + */ + +#ifndef TVM_AUTO_SCHEDULER_FEATURE_H_ +#define TVM_AUTO_SCHEDULER_FEATURE_H_ + +#include <tvm/auto_scheduler/compute_dag.h> +#include <tvm/auto_scheduler/measure.h> + +#include <string> +#include <vector> + +namespace tvm { +namespace auto_scheduler { + +/*! + * \brief Get per-store feature from a TIR Stmt + * \param stmt The input lowered TIR statement + * \param cache_line_size The size of cache line in bytes + * \param max_n_bufs The maximum number of extracted buffers for one statement + * \param ret The returned feature vector + */ +void GetPerStoreFeature(const Stmt& stmt, int cache_line_size, int max_n_bufs, + std::vector<float>* ret); + +/* + * \brief Get the names of elements in the feature vector. Use this for debug and inspection. + * \param max_n_bufs The maximum number of extracted buffers for one statement + * \param ret The returned names. + */ +void GetPerStoreFeatureName(int max_n_bufs, std::vector<std::string>* ret); + +/*! + * \brief Get per-store feature from states of the same task + * \param states The input states + * \param task The same search task for all states + * \param skip_first_n_feature_extraction Skip feature extraction for the first n states + * \param max_n_bufs The maximum number of extracted buffers for one statement + * \param features The returned feature vector. The innermost vector contains the + * feature vectors for all BufferStoreNode statements + */ +void GetPerStoreFeaturesFromStates(const Array<State>& states, const SearchTask& task, + int skip_first_n_feature_extraction, int max_n_bufs, Review comment: "skip_fst_n_features" may be more compact? ########## File path: python/tvm/auto_scheduler/feature.py ########## @@ -0,0 +1,242 @@ +# 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. + +"""" +Python API for Feature extraction. The extracted features vector are used by cost models. + +We extract one feature vector per BufferStoreNode statement in a TIR Stmt, +so we call this feature as "Per Store" feature. +The cost model also does prediction for each BufferStoreNode statement and aggregates +the predicted score of each BufferStoreNode as the score of a TIR Stmt. + +The feature specification is defined by `src/auto_scheduler/feature.cc::FeatureSet` +""" + +from typing import List, Tuple, Union, Optional +import struct + +import numpy as np + +from .loop_state import State, StateObject +from .measure import MeasureInput, MeasureResult +from . import _ffi_api + +# The maximum number of extracted buffers for one statement +DEFAULT_MAX_N_BUFS = 5 + +# The length of the feature vector +DEFAULT_FEATURE_VEC_LEN = 164 + +# The size of int and float in bytes +SIZE_OF_INT = 4 +SIZE_OF_FLOAT = 4 + +def unpack_feature(byte_arr: bytearray) -> Tuple[np.ndarray, np.ndarray, np.ndarray]: + """Unpack the flatten feature (in byte array format) from c++ + + Parameters + ---------- + byte_arr: bytearray + The two-dimensional feature vector in serialized byte array format + + Returns + ------- + features: np.ndarray + Feature vectors + normalized_throughputs: np.ndarray + Normalized throughputs + task_ids: np.ndarray + Task ids + """ + + # The format for n records is: + # { + # int n; + # int[n+2] sizes Review comment: Personally, I think junrushao1994's suggestion is clearer than current comment convention.Since I think "float sizes[n + 1]" illustrates the format semantics more precisely than "float[sizes[0]]". ########## File path: python/tvm/auto_scheduler/feature.py ########## @@ -0,0 +1,242 @@ +# 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. + +"""" +Python API for Feature extraction. The extracted features vector are used by cost models. + +We extract one feature vector per BufferStoreNode statement in a TIR Stmt, +so we call this feature as "Per Store" feature. +The cost model also does prediction for each BufferStoreNode statement and aggregates +the predicted score of each BufferStoreNode as the score of a TIR Stmt. + +The feature specification is defined by `src/auto_scheduler/feature.cc::FeatureSet` +""" + +from typing import List, Tuple, Union, Optional +import struct + +import numpy as np + +from .loop_state import State, StateObject +from .measure import MeasureInput, MeasureResult +from . import _ffi_api + +# The maximum number of extracted buffers for one statement +DEFAULT_MAX_N_BUFS = 5 + +# The length of the feature vector +DEFAULT_FEATURE_VEC_LEN = 164 + +# The size of int and float in bytes +SIZE_OF_INT = 4 +SIZE_OF_FLOAT = 4 + +def unpack_feature(byte_arr: bytearray) -> Tuple[np.ndarray, np.ndarray, np.ndarray]: + """Unpack the flatten feature (in byte array format) from c++ + + Parameters + ---------- + byte_arr: bytearray + The two-dimensional feature vector in serialized byte array format + + Returns + ------- + features: np.ndarray + Feature vectors + normalized_throughputs: np.ndarray + Normalized throughputs + task_ids: np.ndarray + Task ids + """ + + # The format for n records is: + # { + # int n; + # int[n+2] sizes + + # float[sizes[0]] feature for record 1 + # float[sizes[1]] feature for record 2 + # ... feature for record i... + # float[sizes[n-1]] feature for record n + + # float[sizes[n]] normalized throughput for n records + # int[sizes[n+1]] task id for n records + # } + + vec_len = DEFAULT_FEATURE_VEC_LEN + + # unpack sizes + offset = 0 + n = struct.unpack_from("1i", byte_arr, offset=offset)[0] + offset += SIZE_OF_INT + + sizes = struct.unpack_from("%di" % (n+2), byte_arr, offset=offset) + offset += SIZE_OF_INT * (n+2) + + # unpack features + features = [] + for size in sizes[:-2]: + row = [] + + # Now, we need to unpack the feature for multiple statements. + # The format is: + # { + # int n_stmts + # float[n_stmt][vec_len] feature_vecs + # } + # where vec_len can be calculated by `(size - 1) / n_stmts` + + if size == 0: + # failed during lowering + features.append(np.zeros((1, vec_len))) + else: + n_stmts = struct.unpack_from("f", byte_arr, offset=offset) + offset += SIZE_OF_FLOAT + + n_stmts = int(n_stmts[0] + 0.5) Review comment: +1, how about store the float array and the n_stmts separately? By doing this we may need to add extra code, but I think the program semantic is clearer. ---------------------------------------------------------------- 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]
