op_function_generator.h 11.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32
// Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed 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.

#pragma once

#include <map>
#include <set>
#include <string>

// NOTE(zhiqiu): Commonly, the inputs in auto-generated OP function are
// determined by the OP`s proto automatically, i.e., all the inputs registered
// in OpMaker.
// However, some OPs have dispensable inputs, which means the input can
// be none for some conditions. It is discovered that most dispensable inputs
// is not used in imperative mode, so we drop those inputs when generating OP
// functions. While, for very few OPs, the dispensable inputs are used, we
// need to manually specify them in this map.
std::map<std::string, std::set<std::string>> op_ins_map = {
    {"layer_norm", {"X", "Scale", "Bias"}},
    {"bincount", {"X", "Weights"}},
    {"fused_attention",
33 34
     {"X", "LnScale", "LnBias", "QKVW", "QKVBias", "CacheKV", "SrcMask",
      "OutLinearW", "OutLinearBias", "Ln2Scale", "Ln2Bias"}},
35 36 37 38 39 40
    {"instance_norm", {"X", "Scale", "Bias"}},
    {"gru_unit", {"Input", "HiddenPrev", "Weight", "Bias"}},
    {"label_smooth", {"X", "PriorDist"}},
    {"assign", {"X"}},
    {"reshape2", {"X", "Shape"}},
    {"expand", {"X", "ExpandTimes"}},
W
wanghuancoder 已提交
41 42 43 44 45 46 47 48 49
    {"slice",
     {"Input", "StartsTensor", "EndsTensor", "StartsTensorList",
      "EndsTensorList"}},
    {"strided_slice",
     {"Input", "StartsTensor", "EndsTensor", "StridesTensor",
      "StartsTensorList", "EndsTensorList", "StridesTensorList"}},
    {"set_value",
     {"Input", "ValueTensor", "StartsTensorList", "EndsTensorList",
      "StepsTensorList"}},
50 51 52 53 54
    {"fake_quantize_dequantize_moving_average_abs_max",
     {"X", "InScale", "InAccum", "InState"}},
    {"nll_loss", {"X", "Label", "Weight"}},
    {"bilinear_tensor_product", {"X", "Y", "Weight", "Bias"}},
    {"gather", {"X", "Index", "Axis"}},
K
kuizhiqing 已提交
55
    {"repeat_interleave", {"X", "RepeatsTensor"}},
56 57 58 59 60 61 62 63 64 65 66 67 68
    {"roi_pool", {"X", "ROIs", "RoisNum"}},
    {"roi_align", {"X", "ROIs", "RoisNum"}},
    {"psroi_pool", {"X", "ROIs", "RoisNum"}},
    {"collect_fpn_proposals",
     {"MultiLevelRois", "MultiLevelScores", "MultiLevelRoIsNum"}},
    {"distribute_fpn_proposals", {"FpnRois", "RoisNum"}},
    {"warpctc", {"Logits", "Label", "LogitsLength", "LabelLength"}},
    {"hierarchical_sigmoid",
     {"X", "W", "Label", "PathTable", "PathCode", "Bias"}},
    {"moving_average_abs_max_scale", {"X", "InAccum", "InState"}},
    {"multiclass_nms3", {"BBoxes", "Scores", "RoisNum"}},
    {"box_coder", {"PriorBox", "PriorBoxVar", "TargetBox"}},
    {"momentum", {"Param", "Grad", "Velocity", "LearningRate", "MasterParam"}},
69 70
    {"merged_momentum",
     {"Param", "Grad", "Velocity", "LearningRate", "MasterParam"}},
71 72
    {"sparse_momentum",
     {"Param", "Grad", "Velocity", "Index", "LearningRate", "MasterParam"}},
73 74 75 76 77 78 79 80 81 82
    {"rnn", {"Input", "PreState", "WeightList", "SequenceLength"}},
    {"run_program", {"X", "Params"}},
    {"fused_feedforward",
     {"Dropout1Seed", "Dropout2Seed", "Linear1Bias", "Linear2Bias", "Ln1Scale",
      "Ln1Bias", "Ln2Scale", "Ln2Bias"}},
    {"faster_tokenizer", {"Text", "Vocab", "TextPair"}},
    {"matrix_rank", {"X", "TolTensor"}},
    {"adam",
     {"Param", "Grad", "LearningRate", "Moment1", "Moment2", "Beta1Pow",
      "Beta2Pow", "MasterParam"}},
Z
zhangbo9674 已提交
83 84 85
    {"merged_adam",
     {"Param", "Grad", "LearningRate", "Moment1", "Moment2", "Beta1Pow",
      "Beta2Pow", "MasterParam"}},
86 87 88
    {"adamw",
     {"Param", "Grad", "LearningRate", "Moment1", "Moment2", "Beta1Pow",
      "Beta2Pow", "MasterParam"}},
89 90 91
    {"lamb",
     {"Param", "Grad", "LearningRate", "Moment1", "Moment2", "Beta1Pow",
      "Beta2Pow", "MasterParam"}},
92 93
    {"sparse_attention",
     {"Q", "K", "V", "Offset", "Columns", "KeyPaddingMask", "AttnMask"}},
94
    {"sgd", {"Param", "LearningRate", "Grad", "MasterParam"}},
S
Siming Dai 已提交
95
    {"graph_khop_sampler", {"Row", "Eids", "Col_Ptr", "X"}},
W
Weilong Wu 已提交
96 97 98
    {"nce",
     {"Input", "Label", "Weight", "Bias", "SampleWeight", "CustomDistProbs",
      "CustomDistAlias", "CustomDistAliasProbs"}},
F
furnace 已提交
99
    {"check_finite_and_unscale", {"X", "Scale", "FloatStatus"}},
100
    {"group_norm", {"X", "Scale", "Bias"}},
101 102 103
    {"linear_chain_crf", {"Emission", "Transition", "Label", "Length"}},
    {"crf_decoding", {"Emission", "Transition", "Label", "Length"}},
    {"chunk_eval", {"Inference", "Label", "SeqLength"}},
104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119
};

// NOTE(zhiqiu): Like op_ins_map.
// Commonly, the outputs in auto-generated OP function are determined by the
// OP`s proto automatically, i.e., all the outputs registered in OpMaker.
// However, some OPs have dispensable outputs, which means the output can
// be none for some conditions. It is discovered that most dispensable outputs
// is not used in imperative mode, so we drop those outputs when generating OP
// functions. While, for very few OPs, the dispensable outputs are used, we
// need to manually specify them in this map.
std::map<std::string, std::set<std::string>> op_outs_map = {
    {"fake_quantize_dequantize_moving_average_abs_max",
     {"Out", "OutScale", "OutAccum", "OutState"}},
    {"batch_norm",
     {"Y", "MeanOut", "VarianceOut", "SavedMean", "SavedVariance",
      "ReserveSpace"}},
120 121 122 123 124 125 126 127 128 129
    {"fused_attention", {"LnMean",         "LnVariance",
                         "LnOut",          "QKVOut",
                         "QKVBiasOut",     "TransposeOut2",
                         "QKOut",          "QKTVOut",
                         "SoftmaxOut",     "AttnDropoutMaskOut",
                         "AttnDropoutOut", "SrcMaskOut",
                         "FMHAOut",        "OutLinearOut",
                         "DropoutMaskOut", "Ln2Mean",
                         "Ln2Variance",    "BiasDropoutResidualOut",
                         "CacheKVOut",     "Y"}},
130 131 132 133 134 135 136 137 138 139 140 141 142 143 144
    {"sync_batch_norm",
     {"Y", "MeanOut", "VarianceOut", "SavedMean", "SavedVariance",
      "ReserveSpace"}},
    {"unique", {"Out", "Index", "Indices", "Counts"}},
    {"unique_consecutive", {"Out", "Index", "Counts"}},
    {"generate_proposals", {"RpnRois", "RpnRoiProbs", "RpnRoisNum"}},
    {"collect_fpn_proposals", {"FpnRois", "RoisNum"}},
    {"matrix_nms", {"Out", "Index", "RoisNum"}},
    {"distribute_fpn_proposals",
     {"MultiFpnRois", "RestoreIndex", "MultiLevelRoIsNum"}},
    {"moving_average_abs_max_scale",
     {"Out", "OutScale", "OutAccum", "OutState"}},
    {"multiclass_nms3", {"Out", "NmsRoisNum"}},
    {"generate_proposals_v2", {"RpnRois", "RpnRoiProbs", "RpnRoisNum"}},
    {"momentum", {"ParamOut", "VelocityOut", "MasterParamOut"}},
145
    {"merged_momentum", {"ParamOut", "VelocityOut", "MasterParamOut"}},
146
    {"sparse_momentum", {"ParamOut", "VelocityOut", "MasterParamOut"}},
147 148 149 150 151
    {"rnn", {"DropoutState", "Reserve", "Out", "State"}},
    {"run_program", {"DOut"}},
    {"adam",
     {"ParamOut", "Moment1Out", "Moment2Out", "Beta1PowOut", "Beta2PowOut",
      "MasterParamOut"}},
Z
zhangbo9674 已提交
152 153 154
    {"merged_adam",
     {"ParamOut", "Moment1Out", "Moment2Out", "Beta1PowOut", "Beta2PowOut",
      "MasterParamOut"}},
155 156 157
    {"adamw",
     {"ParamOut", "Moment1Out", "Moment2Out", "Beta1PowOut", "Beta2PowOut",
      "MasterParamOut"}},
158
    {"sgd", {"ParamOut", "MasterParamOut"}},
159 160 161
    {"lamb",
     {"ParamOut", "Moment1Out", "Moment2Out", "Beta1PowOut", "Beta2PowOut",
      "MasterParamOut"}},
162
};
163 164 165 166 167 168 169 170 171 172 173 174 175

// NOTE(zhiqiu): Commonly, the outputs in auto-generated OP function are
// generated in C++ automatically.
// However, some OPs need to pass the outputs from Python instead of generating
// them in C++. There are mainly 2 reasons for that,
// (1) Optimizer OPs need to update the input param in-place, like sgd.
//     So they need to pass the output which is same as input param.
// (2) Very few python APIs has out in their arguments, like fill_constant.
//     So they need to pass the python output to C++.
//     Actually, this is not a good design, since it may break the SSA graph,
//     especially in declarative mode.
// For those OPs, we need to manually specify the outs need to pass in this map.
std::map<std::string, std::set<std::string>> op_passing_outs_map = {
176
    {"sgd", {"ParamOut", "MasterParamOut"}},
177 178 179 180 181 182 183 184 185
    {"rmsprop", {"ParamOut", "MomentOut", "MeanSquareOut", "MeanGradOut"}},
    {"ftrl", {"ParamOut", "SquaredAccumOut", "LinearAccumOut"}},
    {"adadelta", {"ParamOut", "AvgSquaredGradOut", "AvgSquaredUpdateOut"}},
    {"adagrad", {"ParamOut", "MomentOut"}},
    {"adamax", {"ParamOut", "MomentOut", "InfNormOut"}},
    {"dpsgd", {"ParamOut"}},
    {"decayed_adagrad", {"ParamOut", "MomentOut"}},
    {"lars_momentum", {"ParamOut", "VelocityOut"}},
    {"coalesce_tensor", {"Output", "FusedOutput"}},
186 187 188
    {"adam",
     {"ParamOut", "Moment1Out", "Moment2Out", "Beta1PowOut", "Beta2PowOut",
      "MasterParamOut"}},
Z
zhangbo9674 已提交
189 190 191
    {"merged_adam",
     {"ParamOut", "Moment1Out", "Moment2Out", "Beta1PowOut", "Beta2PowOut",
      "MasterParamOut"}},
192 193 194
    {"adamw",
     {"ParamOut", "Moment1Out", "Moment2Out", "Beta1PowOut", "Beta2PowOut",
      "MasterParamOut"}},
195 196 197
    {"lamb",
     {"ParamOut", "Moment1Out", "Moment2Out", "Beta1PowOut", "Beta2PowOut",
      "MasterParamOut"}},
198 199 200 201
    {"average_accumulates",
     {"out_sum_1", "out_sum_2", "out_sum_3", "out_num_accumulates",
      "out_old_num_accumulates", "out_num_updates"}},
    {"momentum", {"ParamOut", "VelocityOut", "MasterParamOut"}},
202
    {"merged_momentum", {"ParamOut", "VelocityOut", "MasterParamOut"}},
203
    {"sparse_momentum", {"ParamOut", "VelocityOut", "MasterParamOut"}},
204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247
    {"batch_norm", {"MeanOut", "VarianceOut"}},
    {"sync_batch_norm", {"MeanOut", "VarianceOut"}},
    {"accuracy", {"Correct", "Total"}},
    {"fill_constant", {"Out"}},
    {"recv_v2", {"Out"}},
    {"partial_recv", {"Out"}},
    {"matmul", {"Out"}},
    {"c_broadcast", {"Out"}},
    {"c_sync_calc_stream", {"Out"}},
    {"c_sync_comm_stream", {"Out"}},
    {"c_reduce_sum", {"Out"}},
    {"c_reduce_max", {"Out"}},
    {"c_reduce_min", {"Out"}},
    {"c_reduce_prod", {"Out"}},
    {"c_reduce", {"Out"}},
    {"c_scatter", {"Out"}},
    {"barrier", {"Out"}},
    {"fake_quantize_dequantize_moving_average_abs_max",
     {"Out", "OutScale", "OutAccum", "OutState"}},
    {"fake_quantize_dequantize_abs_max", {"Out", "OutScale"}},
    {"fake_channel_wise_quantize_dequantize_abs_max", {"Out", "OutScale"}},
    {"check_finite_and_unscale", {"Out", "FoundInfinite"}},
    {"update_loss_scaling",
     {"Out", "LossScaling", "OutGoodSteps", "OutBadSteps"}},
    {"moving_average_abs_max_scale",
     {"Out", "OutScale", "OutAccum", "OutState"}},
    {"rnn", {"DropoutState"}},
    {"run_program", {"Out", "DOut", "OutScope"}},
    {"clear_float_status", {"FloatStatusOut"}},
    {"get_float_status", {"FloatStatusOut"}},
};

// NOTE(pangyoki): Tensor View Strategy.
// In this case, a new output varbase will be created, and this varbase will
// reuse the input varbase's allocation.
// It's a map. The key of outer map is the view op name, the value is
// a pair which implies the mapping relationship between the input and
// output varbase.
std::map<std::string, std::pair<std::string, std::string>> view_op_map = {
    {"squeeze2", {"X", "Out"}},  // "X" -> "Out"
    {"unsqueeze2", {"X", "Out"}},
    {"reshape2", {"X", "Out"}},
    {"flatten_contiguous_range", {"X", "Out"}},
};