optimizer.h

// Copyright (c) 2019 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 <memory>
#include <set>
#include <string>
#include <utility>
#include <vector>
#include "lite/core/mir/generate_program_pass.h"
#include "lite/core/mir/pass_manager.h"
#include "lite/core/mir/pass_utils.h"
#include "lite/core/mir/ssa_graph.h"
#include "lite/core/mir/static_kernel_pick_pass.h"
#include "lite/core/mir/type_target_cast_pass.h"
#include "lite/core/program.h"
#include "lite/core/types.h"
#include "lite/model_parser/model_parser.h"

namespace paddle {
namespace lite {

/*
 * lite::Optimizer optimize a program. It utilize the mir passes to analysis the
 * program and export an optimized program.
 */
class Optimizer {
 public:
  Optimizer() {}

  Optimizer(Program&& program, const std::vector<Place>& valid_places) {
    program_ = &program;
    valid_places_ = valid_places;
    CHECK(!valid_places.empty()) << "At least one valid_place should be set";

    core::KernelPickFactor factor;
    factor.ConsiderTarget();
    factor.ConsiderPrecision();
    factor.ConsiderDataLayout();

    Run(std::move(program), valid_places, factor, {});
  }

  void Run(Program&& program,
           const std::vector<Place>& valid_places,
           core::KernelPickFactor kernel_pick_factor,
           const std::vector<std::string>& passes = {}) {
    program_ = &program;
    valid_places_ = valid_places;
    CHECK(!valid_places.empty()) << "At least one valid_place should be set";
    CHECK(!graph_) << "duplicate optimize found";

    graph_.reset(new mir::SSAGraph);
    graph_->Build(program, valid_places);
    graph_->SetValidPlaces(valid_places);

    SpecifyKernelPickTactic(kernel_pick_factor);
    InitTargetTypeTransformPass();

    if (passes.empty() || passes.size() == 1) {
      std::vector<std::string> passes_local{
          {"lite_quant_dequant_fuse_pass",         //
           "weight_quantization_preprocess_pass",  //
           "lite_conv_elementwise_fuse_pass",      // conv-elemwise-bn
           "lite_conv_bn_fuse_pass",               //
           "lite_conv_elementwise_fuse_pass",      // conv-bn-elemwise
           "lite_conv_conv_fuse_pass",             //
           // TODO(Superjomn) Refine the fusion related design to select fusion
           // kernels for devices automatically.
           "lite_conv_activation_fuse_pass",              //
           "lite_var_conv_2d_activation_fuse_pass",       //
           "lite_fc_fuse_pass",                           //
           "lite_shuffle_channel_fuse_pass",              //
           "lite_transpose_softmax_transpose_fuse_pass",  //
           "lite_interpolate_fuse_pass",                  //
           "identity_scale_eliminate_pass",               //
           "elementwise_mul_constant_eliminate_pass",     //
           "lite_sequence_pool_concat_fuse_pass",         //
           "lite_scale_activation_fuse_pass",             //
#if (defined LITE_WITH_LIGHT_WEIGHT_FRAMEWORK) || (defined LITE_WITH_CUDA) || \
    (defined LITE_WITH_ARM)
           "lite_elementwise_activation_fuse_pass",  //
#endif
           "identity_dropout_eliminate_pass",
           "__xpu__resnet_fuse_pass",
           "__xpu__resnet_cbam_fuse_pass",
           "__xpu__mmdnn_fuse_pass",
           "__xpu__multi_encoder_fuse_pass",
           "__xpu__embedding_with_eltwise_add_fuse_pass",
           "__xpu__fc_fuse_pass",
           "quantized_op_attributes_inference_pass",  // Only for fully
                                                      // quantized model, infer
                                                      // the output scale and
                                                      // fix the attribute
                                                      // 'enable_int8' for all
                                                      // of the quantized ops.
           "npu_subgraph_pass",
           "xpu_subgraph_pass",
           "bm_subgraph_pass",
           "apu_subgraph_pass",
           "rknpu_subgraph_pass",
           "mlu_subgraph_pass",
           "static_kernel_pick_pass",  // pick original kernel from graph

           "remove_tf_redundant_ops_pass",
           "variable_place_inference_pass",  // inference arg/var's

           "mlu_postprocess_pass",
           // info(target/precision/layout/device)
           // using kernel info
           "argument_type_display_pass",  // debug pass: show arg-type-node's
                                          // info
                                          // (target/precision/layout/device)

           "type_target_cast_pass",  // add io_copy/io_copy_once if meet
                                     // different targets when last and next
                                     // node
           "variable_place_inference_pass",  //
           "argument_type_display_pass",     //

           "io_copy_kernel_pick_pass",    //
           "argument_type_display_pass",  //

           "variable_place_inference_pass",  //
           "argument_type_display_pass",     //

           "type_precision_cast_pass",       //
           "variable_place_inference_pass",  //
           "argument_type_display_pass",     //

           "type_layout_cast_pass",  // add layout/layout_once op if meet
                                     // different layout when last and next node
           "argument_type_display_pass",  //

           "variable_place_inference_pass",  //
           "argument_type_display_pass",

           "runtime_context_assign_pass",
           "argument_type_display_pass",

           "memory_optimize_pass"}};

      if (passes.size() == 1) {
        // multi_stream_analysis_pass must be in the front of
        // runtime_context_assign_pass
        const std::string msa_pass{"multi_stream_analysis_pass"};
        const std::string depend_pass{"runtime_context_assign_pass"};
        if (passes[0] == msa_pass) {
          auto iter =
              std::find(passes_local.begin(), passes_local.end(), depend_pass);
          if (iter != passes_local.end()) {
            passes_local.insert(iter, msa_pass);
          } else {
            CHECK(false) << "Not find " << depend_pass;
          }
        } else {
          passes_local.push_back(passes[0]);
        }
      }
      RunPasses(passes_local);
    } else {
      RunPasses(passes);
    }
    exec_scope_ = program.exec_scope();
  }

  const lite::Scope* exec_scope() const { return exec_scope_; }

  // Set shape(dims) infos of var descs to scope var.
  //  developer can write pass using input / output tensor dims of op.
  //
  // Example: If you have node `Node* softmax_node`,
  //   you can get dims of output tensor in passes:
  //
  //   auto* scope = softmax_node->AsStmt().op()->scope();
  //   auto softmax_out_arg_name =
  //             softmax_node->outlinks.front()->AsArg().name;
  //   auto softmax_out_tensor =
  //             scope->FindVar(softmax_out_arg_name)->Get<lite::Tensor>();
  //   softmax_out_dims = softmax_out_tensor.dims();
  void SetVarDescShapeToScopeVar() {
    auto dims_to_str_func = [](std::vector<int64_t> shape) -> std::string {
      std::string str_res;
      for (size_t i = 0; i < shape.size(); ++i) {
        str_res += std::to_string(shape[i]);
        if (i != shape.size() - 1) {
          str_res += "x";
        }
      }
      return str_res;
    };

    auto* program_desc = program_->program_desc();
    VLOG(5) << "program_desc->BlocksSize():" << program_desc->BlocksSize();
    auto blocks_desc = program_desc->GetBlocks();
    for (size_t bidx = 0; bidx < blocks_desc.size(); ++bidx) {
      auto block_desc = blocks_desc[bidx];
      auto vars_desc = block_desc.GetVars();
      for (size_t vidx = 0; vidx < vars_desc.size(); ++vidx) {
        auto var_desc = vars_desc[vidx];
        VLOG(5) << var_desc.Name() << " "
                << dims_to_str_func(var_desc.GetShape());
        if (var_desc.Name() == "feed" || var_desc.Name() == "fetch") continue;
        auto* var = program_->exec_scope()->FindVar(var_desc.Name());
        auto tensor = var->GetMutable<lite::Tensor>();
        if (tensor->dims().size() == 0 && var_desc.GetShape().size() != 0) {
          VLOG(5) << "var_desc.Name():" << var_desc.Name()
                  << " shape:" << dims_to_str_func(var_desc.GetShape());
          tensor->Resize(var_desc.GetShape());
        }
        VLOG(5) << "var_desc.Name():" << var_desc.Name()
                << " shape:" << dims_to_str_func(var_desc.GetShape())
                << " tensor:" << tensor->dims();
      }
    }
  }

  // Generate a new program based on the mir graph.
  std::unique_ptr<RuntimeProgram> GenRuntimeProgram() {
    auto pass = mir::PassManager::Global().LookUp<mir::GenerateProgramPass>(
        "generate_program_pass");
    pass->Apply(graph_);
    auto program = pass->GenProgram();
    CHECK(exec_scope_);
    program->set_exec_scope(exec_scope_);
    return program;
  }

  void InitTargetTypeTransformPass() {
    auto* pass =
        mir::PassManager::Global().LookUp<mir::TypeTargetTransformPass>(
            "type_target_cast_pass");
    CHECK(pass);
    CHECK(!valid_places_.empty());
    pass->SetValidPlaces(valid_places_);
  }

  // Generate C++ code which combines the inference program, model and weights.
  void GenCode(const std::string& code_dir);

  const mir::SSAGraph& ssa_graph() const {
    CHECK(graph_);
    return *graph_;
  }

  mir::SSAGraph* mutable_ssa_graph() {
    CHECK(graph_);
    return graph_.get();
  }

  lite::Scope* exec_scope() { return exec_scope_; }

 protected:
  void SpecifyKernelPickTactic(core::KernelPickFactor factor);

  // Specify the passes and run them.
  void RunPasses(const std::vector<std::string>& passes) {
    SetVarDescShapeToScopeVar();
    for (auto& x : passes) {
      LOG(INFO) << "== Running pass: " << x;
      mir::Pass* pass = mir::PassManager::Global().LookUp(x);
      if (!pass) {
        LOG(INFO) << "   - Skip " << x << " because the pass isn't found.";
        continue;
      }
      std::set<TargetType> targets;
      for (const auto& place : valid_places_) {
        targets.insert(place.target);
      }
      bool matched =
          PassMatchesTarget(*pass, targets) && PassMatchesKernels(*pass);
      if (!matched) {
        LOG(INFO) << "   - Skip " << x
                  << " because the target or kernel does not match.";
      } else {
        pass->Apply(graph_);
        LOG(INFO) << "== Finished running: " << x;
      }
    }
  }

 private:
  std::unique_ptr<mir::SSAGraph> graph_;
  std::vector<Place> valid_places_;
  lite::Scope* exec_scope_{};
  Program* program_{};
};

}  // namespace lite
}  // namespace paddle