combine_parallel_op_batch.cc 6.79 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 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207
/*
 * 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.
 */

/*!
 * Copyright (c) 2019 by Contributors
 *
 * \file combine_parallel_op_batch.cc
 * \brief Combine parallel ops into a single batch op.
 * 
 * This pass replaces ops that share the same input node and same shape
 * with a single op that takes in batched input. The inputs of the new
 * batched op are the stack of the original inputs. Elementwise and
 * broadcast ops following the original op are also stacked
 * and fused if possible. For example:
 * 
 *            data
 *         /         \
 *    add (2,2)     add (2,2)
 *      |            |
 * elemwise (2,2)   elemwise (2,2)
 *      |            |
 *
 * Would become:
 * 
 *            data
 *              |
 *       add+elemwise (2,2,2)
 *          /       \
 *
 */

#include <tvm/relay/analysis.h>
#include <tvm/relay/expr_functor.h>
#include <tvm/relay/attrs/nn.h>
#include <tvm/relay/attrs/transform.h>
#include <tvm/relay/op_attr_types.h>
#include <tvm/relay/transform.h>
#include <unordered_map>
#include <unordered_set>
#include "./expr_subst.h"
#include "./pattern_util.h"
#include "./combine_parallel_op.h"
#include "./combine_parallel_op_batch.h"

namespace tvm {
namespace relay {

ParallelOpBatchCombiner::ParallelOpBatchCombiner(const std::string& op_name,
                                                 const std::string& batch_op_name,
                                                 uint64_t min_num_branches)
  : ParallelOpCombiner(op_name, min_num_branches),
    batch_op_name_(batch_op_name) {
}

bool ParallelOpBatchCombiner::IsSupportedOp(const CallNode* n) {
  return true;
}

bool ParallelOpBatchCombiner::CanOpsBeCombined(const CallNode* a, const CallNode* b) {
  if (a->args.size() != b->args.size()) {
    return false;
  }

  AttrsEqual eq;
  for (size_t i = 0; i < a->args.size(); i++) {
    auto ta = a->args[i]->type_as<TensorTypeNode>();
    auto tb = b->args[i]->type_as<TensorTypeNode>();
    if (ta->shape.size() != tb->shape.size() || !eq(ta->dtype, tb->dtype)) {
      return false;
    }

    for (size_t j = 0; j < ta->shape.size(); j++) {
      if (!eq(ta->shape[j], tb->shape[j])) {
        return false;
      }
    }
  }

  return true;
}

Call ParallelOpBatchCombiner::MakeCombinedOp(const Group& branches) {
  const Op& batch_op = Op::Get(batch_op_name_);

  Array<Expr> new_args;
  size_t num_args = branches[0][0]->args.size();
  for (size_t i = 0; i < num_args; i++) {
    Array<Expr> arg_from_all_branches;
    for (const auto& branch : branches) {
      arg_from_all_branches.push_back(branch[0]->args[i]);
    }

    new_args.push_back(MakeStack(TupleNode::make(arg_from_all_branches), 0));
  }

  return CallNode::make(batch_op, new_args, Attrs(), {});
}

bool ParallelOpBatchCombiner::IsArgCompatible(const CallNode* a, const CallNode* b, size_t index) {
  AttrsEqual eq;
  auto ta = a->args[index]->type_as<TensorTypeNode>();
  auto tb = b->args[index]->type_as<TensorTypeNode>();

  if (!eq(ta->dtype, tb->dtype) || ta->shape.size() != tb->shape.size())
    return false;

  for (size_t i = 0; i < ta->shape.size(); i++) {
    if (!eq(ta->shape[i], tb->shape[i]))
      return false;
  }
  return true;
}

Call ParallelOpBatchCombiner::MakeCombinedCallFromFollowingOps(const Expr& data,
                                                               const Group& branches,
                                                               size_t depth,
                                                               size_t parent_index) {
  Array<Expr> new_args;
  const CallNode* call = branches[0][depth];

  for (size_t i = 0; i < call->args.size(); i++) {
    if (i == parent_index) {
      new_args.push_back(data);
      continue;
    }

    Array<Expr> tuple;
    for (const auto& branch : branches) {
      // if the shape of the arg is of shape (j,),
      // expand it to (1,j) so it can be properly broadcasted.
      Expr arg = branch[depth]->args[i];
      const TensorTypeNode* arg_tensor = arg->type_as<TensorTypeNode>();
      if (arg_tensor->shape.size() == 1) {
        Expr expanded_arg = MakeExpandDims(arg, 0, 1);
        tuple.push_back(expanded_arg);
      } else {
        tuple.push_back(arg);
      }
    }

    auto stack = MakeStack(TupleNode::make(tuple), 0);
    new_args.push_back(std::move(stack));
  }

  return CallNode::make(call->op, new_args, call->attrs, {});
}

void ParallelOpBatchCombiner::UpdateGroupOutput(const Expr& data,
                        const Group& branches,
                        size_t depth,
                        ExprSubstMap* subst_map) {
  int index = 0;
  auto split = MakeSplit(data, Integer(branches.size()), 0);
  for (const auto& branch : branches) {
    auto split_data = TupleGetItemNode::make(split, index++);
    auto squeezed_data = MakeSqueeze(split_data, {0});
    subst_map->insert({GetRef<Expr>(branch[depth]), squeezed_data});
  }
}

/*! \brief Combine parallel op into batched op if number of branches >= min_num_branches */
Expr CombineParallelOpBatch(const Expr& expr,
                            const std::string& op_name,
                            const std::string& batch_op_name,
                            uint64_t min_num_branches) {
  return ParallelOpBatchCombiner(op_name, batch_op_name, min_num_branches).Combine(expr);
}

namespace transform {

Pass CombineParallelOpBatch(const std::string& op_name,
                            const std::string& batch_op_name,
                            uint64_t min_num_branches) {
  runtime::TypedPackedFunc<Function(Function, Module, PassContext)> pass_func =
    [=](Function f, Module m, PassContext pc) {
      return Downcast<Function>(CombineParallelOpBatch(f,
                                                       op_name,
                                                       batch_op_name,
                                                       min_num_branches));
  };
  return CreateFunctionPass(pass_func, 4, "CombineParallelOpBatch",
                            {ir::StringImm::make("InferType")});
}

TVM_REGISTER_API("relay._transform.CombineParallelOpBatch")
.set_body_typed(CombineParallelOpBatch);

}  // namespace transform

}  // namespace relay
}  // namespace tvm