cinn_compiler.cc 7.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
// 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.

#include "paddle/fluid/framework/paddle2cinn/cinn_compiler.h"

17
#include <cstdint>
18
#include <iterator>
19 20 21
#include <map>
#include <memory>
#include <string>
22
#include <unordered_map>
23 24 25 26 27 28 29 30 31 32 33

#include "cinn/common/target.h"
#include "cinn/common/type.h"
#include "cinn/frontend/decomposer/use_decomposer.h"
#include "cinn/frontend/pass/use_program_pass.h"
#include "cinn/frontend/program_pass.h"
#include "cinn/frontend/syntax.h"
#include "cinn/hlir/framework/graph.h"
#include "cinn/hlir/framework/graph_compiler.h"
#include "cinn/hlir/framework/pass.h"
#include "cinn/hlir/pass/use_pass.h"
34
#include "paddle/fluid/framework/framework.pb.h"
35 36
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
37
#include "paddle/fluid/framework/ir/node.h"
38 39 40
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/paddle2cinn/cinn_graph_symbolization.h"
#include "paddle/fluid/framework/program_desc.h"
41
#include "paddle/fluid/framework/rw_lock.h"
42
#include "paddle/fluid/framework/tensor.h"
43
#include "paddle/fluid/inference/analysis/dot.h"
44
#include "paddle/fluid/platform/enforce.h"
45
#include "paddle/fluid/string/string_helper.h"
46 47 48 49 50 51

namespace paddle {
namespace framework {
namespace paddle2cinn {

using ir::Graph;
52 53
using ir::Node;
using inference::analysis::Dot;
54 55 56 57 58 59 60 61 62 63 64 65
using ::cinn::common::Target;
using ::cinn::common::Float;
using ::cinn::hlir::framework::GraphCompiler;
using ::cinn::hlir::framework::BuildScope;
using ::cinn::frontend::ProgramPass;
using ::cinn::hlir::framework::ApplyPass;

CinnCompiler* CinnCompiler::GetInstance() {
  static CinnCompiler instance;
  return &instance;
}

66 67 68 69
const CinnCompiledObject& CinnCompiler::Compile(
    const Graph& graph,
    const std::map<std::string, const LoDTensor*>& input_tensors,
    const Target& target) {
70
  VLOG(1) << "-- The graph to be compiled is:\n" << VizGraph(graph);
71 72 73 74 75 76 77
  CinnCacheKey cur_key(graph, input_tensors, target.arch_str());
  bool exist = false;
  {
    AutoRDLock r_guard{&rwlock_};
    exist = cache_.count(cur_key) != 0;
  }
  if (!exist) {
78 79 80
    std::int64_t compiled_num = real_compiled_num_.fetch_add(1);
    auto compiled_res =
        CompileGraph(graph, input_tensors, target, compiled_num);
81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98
    AutoWRLock w_guard{&rwlock_};
    if (!cache_.count(cur_key)) {
      cache_[cur_key] = std::move(compiled_res);
    }
  }
  AutoRDLock guard{&rwlock_};
  const auto& cached_boj = *cache_[cur_key];
  return cached_boj;
}

const CinnCompiledObject& CinnCompiler::Compile(
    const std::string& compilation_key,
    const std::map<std::string, const LoDTensor*>& input_tensors,
    const Target& target) {
  const auto& graph = FindGraph(compilation_key);
  return Compile(graph, input_tensors, target);
}

99 100 101 102 103
std::string CinnCompiler::AddGraph(std::unique_ptr<Graph> graph) {
  std::string graph_key;
  ProgramDesc program;
  GraphToProgram(*graph, &program);
  program.Proto()->SerializeToString(&graph_key);
104 105 106 107 108 109 110 111 112

  PADDLE_ENFORCE_EQ(
      graphs_.count(graph_key), 0,
      platform::errors::PreconditionNotMet(
          "The graph to be added is already in CinnCompiler, which is:\n",
          VizGraph(graph_key).c_str()));
  graphs_[graph_key] = std::move(graph);
  VLOG(4) << "-- Add a graph into CinnCompiler, which is:\n"
          << VizGraph(graph_key);
113 114 115 116 117 118
  return graph_key;
}

const Graph& CinnCompiler::FindGraph(const std::string& graph_key) const {
  PADDLE_ENFORCE_NE(
      graphs_.count(graph_key), 0,
119 120 121
      platform::errors::PreconditionNotMet(
          "Can not find the target graph, of which the key is:\n%s",
          ReadableKey(graph_key).c_str()));
122 123 124
  return *graphs_.at(graph_key);
}

125 126 127 128 129 130
std::string CinnCompiler::VizGraph(const std::string& graph_key) const {
  const Graph& graph = FindGraph(graph_key);
  return VizGraph(graph);
}

std::string CinnCompiler::VizGraph(const Graph& graph) const {
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
  Dot dot;
  std::unordered_map<const Node*, std::string> node2dot;
  int id = 0;
  // Create nodes
  for (const Node* n : graph.Nodes()) {
    std::string node_id = "Node" + std::to_string(id++);
    if (n->IsOp()) {
      dot.AddNode(
          node_id,
          {Dot::Attr("shape", "box"), Dot::Attr("style", "rounded,filled,bold"),
           Dot::Attr("color", "#303A3A"), Dot::Attr("fontcolor", "#ffffff")},
          n->Name());
    } else if (n->IsVar()) {
      auto label = n->Name();
      if (n->Var() && n->Var()->GetType() == proto::VarType::LOD_TENSOR) {
        auto shape = n->Var()->GetShape();
        std::vector<std::string> shape_str(shape.size());
        std::transform(shape.begin(), shape.end(), shape_str.begin(),
                       [](const auto& val) { return std::to_string(val); });
        label += "\n" + string::join_strings(shape_str, ',');
      }
      dot.AddNode(
          node_id,
          {Dot::Attr("shape", "box"), Dot::Attr("style", "rounded,filled,bold"),
           Dot::Attr("color", n->Var()->IsParameter() ? "#148b97" : "#dddddd"),
           Dot::Attr("fontcolor",
                     n->Var()->IsParameter() ? "#ffffff" : "#000000")},
          label);
    }
    node2dot[n] = node_id;
  }
  // Create edges
  for (const Node* n : graph.Nodes()) {
    const auto& src_id = node2dot.at(n);
    for (auto* out : n->outputs) {
      const auto& dest_id = node2dot.at(out);
      dot.AddEdge(src_id, dest_id, {});
    }
169
  }
170
  return dot.Build();
171 172
}

173 174
std::string CinnCompiler::ReadableKey(
    const std::string& compilation_key) const {
175
  proto::ProgramDesc desc;
176
  desc.ParseFromString(compilation_key);
177 178 179 180 181 182 183 184 185
  return desc.DebugString();
}

void CinnCompiler::Clear() {
  {
    AutoWRLock guard{&rwlock_};
    graphs_.clear();
    cache_.clear();
  }
186
  real_compiled_num_.store(1);
187 188 189 190 191
}

std::unique_ptr<CinnCompiledObject> CinnCompiler::CompileGraph(
    const ir::Graph& graph,
    const std::map<std::string, const LoDTensor*>& input_tensors,
192 193
    const Target& target, std::int64_t compiled_num) const {
  CinnGraphSymbolization symbol{compiled_num, graph, target, input_tensors};
194 195 196 197
  auto frontend_program = symbol();
  ProgramPass::Apply(&frontend_program, target, {"Decomposer"});
  auto cinn_graph = std::make_shared<::cinn::hlir::framework::Graph>(
      frontend_program, target);
198
  VLOG(1) << "-- The " << compiled_num << "-th compilation ("
199 200 201 202
          << target.arch_str() << "), and its related graph:\n"
          << cinn_graph->Visualize();
  ApplyPass(cinn_graph.get(), "OpFusion");
  auto scope = BuildScope(target, cinn_graph);
203 204 205

  auto graph_compiler =
      std::make_unique<GraphCompiler>(target, scope, cinn_graph);
206 207
  GraphCompiler::CompileOptions options;
  options.with_instantiate_variables = false;
208
  auto compiled_res = graph_compiler->Build(options);
209
  auto compiled_obj = std::make_unique<CinnCompiledObject>();
210 211
  *compiled_obj = {std::move(graph_compiler),
                   std::move(compiled_res.runtime_program), scope,
212 213 214 215 216 217 218
                   symbol.var_model_to_program_map()};
  return compiled_obj;
}

}  // namespace paddle2cinn
}  // namespace framework
}  // namespace paddle