cpu_quantize_pass.cc 12.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
// 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.

15
#include "paddle/fluid/framework/ir/mkldnn/cpu_quantize_pass.h"
16
#include <limits>
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
#include <utility>
#include <vector>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/string/pretty_log.h"

namespace paddle {
namespace framework {
namespace ir {

namespace {

void UnlinkNodes(ir::Node* a, ir::Node* b) {
  a->outputs.erase(std::remove(a->outputs.begin(), a->outputs.end(), b),
                   a->outputs.end());
  b->inputs.erase(std::remove(b->inputs.begin(), b->inputs.end(), a),
                  b->inputs.end());
}

}  // namespace

enum { U8_MAX = 255, S8_MAX = 127 };

using EigenVectorArrayMap = Eigen::Map<Eigen::Array<double, Eigen::Dynamic, 1>>;
using string::PrettyLogDetail;

void CPUQuantizePass::QuantizeInput(Graph* g, Node* op, Node* input,
                                    std::string input_name, double scale_to_one,
                                    bool is_unsigned,
                                    std::string scale_attr_name) const {
  unsigned max = is_unsigned ? U8_MAX : S8_MAX;
  float scale = scale_to_one * max;

  // Create quantize output variable
  VarDesc quantize_out_desc(patterns::PDNodeName("quantize", "out"));
  auto* quantize_out_node = g->CreateVarNode(&quantize_out_desc);

  // create a quantize op node
  OpDesc q_desc;
  q_desc.SetType("quantize");
  q_desc.SetInput("Input", std::vector<std::string>({input->Name()}));
  q_desc.SetOutput("Output",
                   std::vector<std::string>({quantize_out_node->Name()}));
  q_desc.SetAttr("Scale", scale);
  q_desc.SetAttr("is_negative_input", !is_unsigned);
  auto quantize_op = g->CreateOpNode(&q_desc);  // OpDesc will be copied.

  // update op's input
  op->Op()->SetInput(input_name,
                     std::vector<std::string>({quantize_out_node->Name()}));

  // link quantize op
  UnlinkNodes(input, op);
  IR_NODE_LINK_TO(input, quantize_op);
  IR_NODE_LINK_TO(quantize_op, quantize_out_node);
  IR_NODE_LINK_TO(quantize_out_node, op);

  if (!scale_attr_name.empty()) op->Op()->SetAttr(scale_attr_name, scale);
}

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
void CPUQuantizePass::QuantizeInputs(Graph* g, Node* op, std::string input_name,
                                     VarQuantScale* scales, bool are_unsigned,
                                     std::string scale_attr_name) const {
  auto inputs = op->inputs;
  PADDLE_ENFORCE_GE(inputs.size(), 1);

  // create a quantize op desc prototype
  OpDesc q_desc;
  q_desc.SetType("quantize");

  std::vector<Node*> quantize_out_nodes(inputs.size());
  std::vector<std::string> quantize_out_node_names(inputs.size());

  double scale_min = std::numeric_limits<double>::max();
  for (const auto& input : inputs) {
    double scale = (*scales)[input->Name()].second.data<double>()[0];
    if (scale < scale_min) scale_min = scale;
  }
  unsigned max = are_unsigned ? U8_MAX : S8_MAX;
  float scale = scale_min * max;

  for (size_t i = 0; i < inputs.size(); i++) {
    // Create quantize output variable
    VarDesc quantize_out_desc(patterns::PDNodeName("quantize", "out"));
    quantize_out_nodes[i] = g->CreateVarNode(&quantize_out_desc);
    quantize_out_node_names[i] = quantize_out_nodes[i]->Name();

    q_desc.SetAttr("Scale", scale);
    q_desc.SetInput("Input", std::vector<std::string>({inputs[i]->Name()}));
    q_desc.SetOutput("Output",
                     std::vector<std::string>({quantize_out_node_names[i]}));
    q_desc.SetAttr("is_negative_input", !are_unsigned);
    auto quantize_op = g->CreateOpNode(&q_desc);  // OpDesc will be copied.

    // link quantize op
    UnlinkNodes(inputs[i], op);
    IR_NODE_LINK_TO(inputs[i], quantize_op);
    IR_NODE_LINK_TO(quantize_op, quantize_out_nodes[i]);
    IR_NODE_LINK_TO(quantize_out_nodes[i], op);
  }

  // update op's input
  op->Op()->SetInput(input_name, quantize_out_node_names);

  if (!scale_attr_name.empty()) op->Op()->SetAttr(scale_attr_name, scale);
}

123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266
void CPUQuantizePass::DequantizeOutput(Graph* g, Node* op, Node* output,
                                       std::string output_name,
                                       double scale_to_one, bool is_unsigned,
                                       std::string scale_attr_name) const {
  unsigned max = is_unsigned ? U8_MAX : S8_MAX;
  float scale = scale_to_one * max;

  // Create dequantize input variable
  VarDesc dequantize_in_desc(patterns::PDNodeName("dequantize", "in"));
  auto* dequantize_in_node = g->CreateVarNode(&dequantize_in_desc);

  // create a dequantize op node for output.
  OpDesc deq_desc;
  deq_desc.SetType("dequantize");
  deq_desc.SetInput("Input",
                    std::vector<std::string>({dequantize_in_node->Name()}));
  deq_desc.SetOutput("Output", std::vector<std::string>({output->Name()}));
  deq_desc.SetAttr("Scale", scale);
  auto dequantize_op = g->CreateOpNode(&deq_desc);  // OpDesc will be copied.

  // update op's output
  op->Op()->SetOutput(output_name,
                      std::vector<std::string>({dequantize_in_node->Name()}));

  // link dequantize op
  UnlinkNodes(op, output);
  IR_NODE_LINK_TO(op, dequantize_in_node);
  IR_NODE_LINK_TO(dequantize_in_node, dequantize_op);
  IR_NODE_LINK_TO(dequantize_op, output);

  if (!scale_attr_name.empty()) op->Op()->SetAttr(scale_attr_name, scale);
}

void CPUQuantizePass::QuantizeConv(Graph* graph,
                                   bool with_residual_data) const {
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
  patterns::ConvResidual conv_pattern{pattern, name_scope_};
  conv_pattern(with_residual_data);

  int quantize_conv_count = 0;
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
    VLOG(4) << "Quantize conv2d op";
    GET_IR_NODE_FROM_SUBGRAPH(conv_op, conv_op, conv_pattern);
    auto* conv_op_desc = conv_op->Op();

    // skip if should not be quantized
    if (!conv_op_desc->HasAttr("use_quantizer") ||
        !boost::get<bool>(conv_op_desc->GetAttr("use_quantizer")))
      return;

    GET_IR_NODE_FROM_SUBGRAPH(conv_filter, conv_filter, conv_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(conv_input, conv_input, conv_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(conv_output, conv_output, conv_pattern);

    // get scales calculated after warmup, they scale variables to MAX=1.0
    auto scales = Get<VarQuantScale>("quant_var_scales");

    auto input_scale = scales[conv_input->Name()].second.data<double>()[0];
    bool is_input_unsigned = scales[conv_input->Name()].first;
    QuantizeInput(g, conv_op, conv_input, "Input", input_scale,
                  is_input_unsigned, "Scale_in");

    auto filter_scale_tensor = scales[conv_filter->Name()].second;
    EigenVectorArrayMap eigen_tensor{filter_scale_tensor.data<double>(),
                                     filter_scale_tensor.numel(), 1};
    eigen_tensor *= static_cast<double>(S8_MAX);
    std::vector<float> filter_scale{
        filter_scale_tensor.data<double>(),
        filter_scale_tensor.data<double>() + filter_scale_tensor.numel()};

    conv_op->Op()->SetAttr("Scale_weights", filter_scale);

    if (with_residual_data) {
      GET_IR_NODE_FROM_SUBGRAPH(conv_residual_data, conv_residual_data,
                                conv_pattern);
      auto residual_scale =
          scales[conv_residual_data->Name()].second.data<double>()[0];
      bool is_residual_unsigned = scales[conv_residual_data->Name()].first;

      QuantizeInput(g, conv_op, conv_residual_data, "ResidualData",
                    residual_scale, is_residual_unsigned, "Scale_in_eltwise");
    }

    auto output_scale = scales[conv_output->Name()].second.data<double>()[0];
    bool is_output_unsigned = scales[conv_output->Name()].first;
    DequantizeOutput(g, conv_op, conv_output, "Output", output_scale,
                     is_output_unsigned, "Scale_out");

    ++quantize_conv_count;
  };

  gpd(graph, handler);
  AddStatis(quantize_conv_count);

  std::stringstream msg_ss;
  msg_ss << "---    quantized " << quantize_conv_count << " conv2d ops";
  if (with_residual_data) msg_ss << " with residual connection";
  PrettyLogDetail(msg_ss.str().c_str());
}

void CPUQuantizePass::QuantizePool(Graph* graph) const {
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
  patterns::Pool pool_pattern{pattern, name_scope_};
  pool_pattern();

  int quantize_pool_count = 0;
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
    VLOG(4) << "Quantize pool2d op";
    GET_IR_NODE_FROM_SUBGRAPH(pool_op, pool_op, pool_pattern);
    auto* pool_op_desc = pool_op->Op();

    // skip if should not be quantized
    if (!pool_op_desc->HasAttr("use_quantizer") ||
        !boost::get<bool>(pool_op_desc->GetAttr("use_quantizer")))
      return;

    GET_IR_NODE_FROM_SUBGRAPH(pool_input, pool_input, pool_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(pool_output, pool_output, pool_pattern);

    // get scales calculated after warmup, they scale variables to MAX=1.0
    auto scales = Get<VarQuantScale>("quant_var_scales");

    auto input_scale = scales[pool_input->Name()].second.data<double>()[0];
    bool is_input_unsigned = scales[pool_input->Name()].first;
    QuantizeInput(g, pool_op, pool_input, "X", input_scale, is_input_unsigned);

    auto output_scale = scales[pool_output->Name()].second.data<double>()[0];
    bool is_output_unsigned = scales[pool_output->Name()].first;
    DequantizeOutput(g, pool_op, pool_output, "Out", output_scale,
                     is_output_unsigned);

    ++quantize_pool_count;
  };

  gpd(graph, handler);
  AddStatis(quantize_pool_count);

  PrettyLogDetail("---    quantized %d pool2d ops", quantize_pool_count);
}

267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308
void CPUQuantizePass::QuantizeConcat(Graph* graph) const {
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
  patterns::Concat concat_pattern{pattern, name_scope_};
  concat_pattern();

  int quantize_concat_count = 0;
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
    VLOG(4) << "Quantize concat op";
    GET_IR_NODE_FROM_SUBGRAPH(concat_op, concat_op, concat_pattern);
    auto* concat_op_desc = concat_op->Op();

    // skip if should not be quantized
    if (!concat_op_desc->HasAttr("use_quantizer") ||
        !boost::get<bool>(concat_op_desc->GetAttr("use_quantizer")))
      return;

    GET_IR_NODE_FROM_SUBGRAPH(concat_out, concat_out, concat_pattern);

    // get scales calculated after warmup, they scale variables to MAX=1.0
    auto scales = Get<VarQuantScale>("quant_var_scales");

    // if all inputs were unsigned, then the output was set to unsigned
    // during the scale calculation step
    bool are_all_inputs_unsigned = scales[concat_out->Name()].first;
    QuantizeInputs(g, concat_op, "X", &scales, are_all_inputs_unsigned);

    auto output_scale = scales[concat_out->Name()].second.data<double>()[0];

    DequantizeOutput(g, concat_op, concat_out, "Out", output_scale,
                     are_all_inputs_unsigned);

    ++quantize_concat_count;
  };

  gpd(graph, handler);
  AddStatis(quantize_concat_count);

  PrettyLogDetail("---    quantized %d concat ops", quantize_concat_count);
}

309
void CPUQuantizePass::ApplyImpl(ir::Graph* graph) const {
310
  VLOG(3) << "Quantizing the graph.";
311 312
  PADDLE_ENFORCE(graph);
  FusePassBase::Init(name_scope_, graph);
313 314 315

  PADDLE_ENFORCE(param_scope());

316 317 318
  QuantizeConv(graph, false /* with_residual_data */);
  QuantizeConv(graph, true /* with_residual_data */);
  QuantizePool(graph);
319
  QuantizeConcat(graph);
320 321 322 323 324 325 326 327
}

}  // namespace ir
}  // namespace framework
}  // namespace paddle

REGISTER_PASS(cpu_quantize_pass, paddle::framework::ir::CPUQuantizePass)
    .RequirePassAttr("quant_var_scales");