cpu_quantize_pass.cc 20.7 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
#include <utility>
#include <vector>
#include "paddle/fluid/framework/eigen.h"
M
Michał Gallus 已提交
20
#include "paddle/fluid/platform/errors.h"
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
#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 {
M
Michał Gallus 已提交
47 48 49 50 51 52 53
  auto inputs = op->Op()->InputNames();
  bool name_found =
      std::find(inputs.begin(), inputs.end(), input_name) != inputs.end();
  PADDLE_ENFORCE_EQ(
      name_found, true,
      platform::errors::InvalidArgument("%s isn't the input of the %s operator",
                                        input_name, op->Op()->Type()));
54 55 56 57 58 59 60 61 62 63 64 65 66 67 68
  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);
69 70 71

  q_desc.SetAttr("output_format",
                 Has("data_layout") ? Get<std::string>("data_layout") : "NHWC");
72 73 74 75 76 77 78 79 80 81 82 83 84 85 86
  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);
}

87 88 89 90
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;
91
  auto output = op->outputs[0];
92
  PADDLE_ENFORCE_GE(inputs.size(), 1);
93
  PADDLE_ENFORCE_EQ(op->outputs.size(), 1);
94 95 96 97 98 99 100 101

  // 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());

102
  double scale_out = (*scales)[output->Name()].second.data<double>()[0];
103
  unsigned max = are_unsigned ? U8_MAX : S8_MAX;
104
  float scale = scale_out * max;
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

  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);
}

132 133 134 135
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 {
M
Michał Gallus 已提交
136 137 138 139 140 141 142
  auto outputs = op->Op()->OutputNames();
  bool name_found =
      std::find(outputs.begin(), outputs.end(), output_name) != outputs.end();
  PADDLE_ENFORCE_EQ(name_found, true,
                    platform::errors::InvalidArgument(
                        "%s isn't the output of the %s operator", output_name,
                        op->Op()->Type()));
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
  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
187
    if (!conv_op_desc->GetAttrIfExists<bool>("use_quantizer")) return;
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

    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");

227
    // change threshold in bounded ReLu
228 229
    if (conv_op->Op()->GetAttrIfExists<std::string>("fuse_activation") ==
        "relu6") {
230
      float scale_out = boost::get<float>(conv_op->Op()->GetAttr("Scale_out"));
231 232
      float threshold = boost::get<float>(conv_op->Op()->GetAttr("fuse_alpha"));
      conv_op->Op()->SetAttr("fuse_alpha", scale_out * threshold);
233 234
    }

235 236 237 238 239 240 241 242 243 244 245 246
    ++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());
}

M
Michał Gallus 已提交
247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306
void CPUQuantizePass::QuantizeFc(Graph* graph) const {
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
  patterns::FCMKLDNN fc_pattern{pattern, name_scope_};
  auto* fc_input = gpd.mutable_pattern()
                       ->NewNode("fc_quantizer/input")
                       ->AsInput()
                       ->assert_is_op_input("fc", "Input");
  fc_pattern(fc_input, false);

  int quantize_fc_count = 0;
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
    VLOG(4) << "Quantize fc op";
    GET_IR_NODE_FROM_SUBGRAPH(fc, fc, fc_pattern);
    auto* fc_op_desc = fc->Op();

    // skip if should not be quantized
    if (fc_op_desc->GetAttrIfExists<bool>("use_quantizer") != true ||
        fc_op_desc->GetAttrIfExists<bool>("use_mkldnn") != true)
      return;

    GET_IR_NODE_FROM_SUBGRAPH(weights, weights, fc_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(input, input, fc_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(output, output, fc_pattern);

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

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

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

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

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

    ++quantize_fc_count;
  };

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

  std::stringstream msg_ss;
  msg_ss << "---    quantized " << quantize_fc_count << " fc ops";
  PrettyLogDetail(msg_ss.str().c_str());
}

307 308 309 310 311 312 313 314 315 316 317 318 319 320
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
321
    if (!pool_op_desc->GetAttrIfExists<bool>("use_quantizer")) return;
322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346

    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);
}

347 348 349 350 351 352 353 354 355 356 357 358 359 360
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
361
    if (!concat_op_desc->GetAttrIfExists<bool>("use_quantizer")) return;
362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386

    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);
}

387 388 389 390 391 392 393 394 395 396 397 398 399 400
void CPUQuantizePass::QuantizePriorBox(Graph* graph) const {
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
  patterns::PriorBox prior_box_pattern{pattern, name_scope_};
  prior_box_pattern();

  int quantize_prior_box_count = 0;
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
    VLOG(4) << "Quantize prior_box op";
    GET_IR_NODE_FROM_SUBGRAPH(prior_box_op, prior_box_op, prior_box_pattern);
    auto* prior_box_op_desc = prior_box_op->Op();

    // skip if should not be quantized
401
    if (!prior_box_op_desc->GetAttrIfExists<bool>("use_quantizer")) return;
402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423

    GET_IR_NODE_FROM_SUBGRAPH(prior_box_input, prior_box_input,
                              prior_box_pattern);

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

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

    ++quantize_prior_box_count;
  };

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

  PrettyLogDetail("---    quantized %d prior_box ops",
                  quantize_prior_box_count);
}

424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443
void CPUQuantizePass::QuantizeTranspose(Graph* graph) const {
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
  patterns::Transpose transpose_pattern{pattern, name_scope_};
  transpose_pattern();

  int quantize_transpose_count = 0;
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
    VLOG(4) << "Quantize transpose op";
    GET_IR_NODE_FROM_SUBGRAPH(transpose_op, transpose_op, transpose_pattern);
    auto* transpose_op_desc = transpose_op->Op();

    // skip if should not be quantized
    if (!transpose_op_desc->GetAttrIfExists<bool>("use_quantizer")) {
      return;
    }
    GET_IR_NODE_FROM_SUBGRAPH(prev_op, prev_op, transpose_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(next_op, next_op, transpose_pattern);

444
    // skip if prev op and next op are not quantized
445
    if (!(prev_op->Op()->Type() == "dequantize" ||
446 447 448
          (prev_op->Op()->GetAttrIfExists<bool>("use_quantizer"))) &&
        !(next_op->Op()->Type() == "quantize" ||
          (next_op->Op()->GetAttrIfExists<bool>("use_quantizer")))) {
449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475
      return;
    }
    GET_IR_NODE_FROM_SUBGRAPH(transpose_in, transpose_in, transpose_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(transpose_out, transpose_out, transpose_pattern);

    // get scales calculated after warmup, they scale variables to MAX=1.0
    auto scales = Get<VarQuantScale>("quant_var_scales");
    auto input_scale = scales[transpose_in->Name()].second.data<double>()[0];
    bool is_input_unsigned = scales[transpose_in->Name()].first;
    QuantizeInput(g, transpose_op, transpose_in, "X", input_scale,
                  is_input_unsigned);

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

    ++quantize_transpose_count;
  };

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

  PrettyLogDetail("---    quantized %d transpose ops",
                  quantize_transpose_count);
}

476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527
void CPUQuantizePass::QuantizeReshape(Graph* graph) const {
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
  patterns::Reshape reshape_pattern{pattern, name_scope_};
  reshape_pattern();

  int quantize_reshape_count = 0;
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
    VLOG(4) << "Quantize reshape op";
    GET_IR_NODE_FROM_SUBGRAPH(reshape_op, reshape_op, reshape_pattern);
    auto* reshape_op_desc = reshape_op->Op();

    // skip if should not be quantized
    if (!reshape_op_desc->GetAttrIfExists<bool>("use_quantizer")) {
      return;
    }
    GET_IR_NODE_FROM_SUBGRAPH(prev_op, prev_op, reshape_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(next_op, next_op, reshape_pattern);

    // skip if prev op  and next op is not quantized
    if (!(prev_op->Op()->Type() == "dequantize" ||
          (prev_op->Op()->GetAttrIfExists<bool>("use_quantizer"))) &&
        !(next_op->Op()->Type() == "quantize" ||
          (next_op->Op()->GetAttrIfExists<bool>("use_quantizer")))) {
      return;
    }

    GET_IR_NODE_FROM_SUBGRAPH(reshape_in, reshape_in, reshape_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(reshape_out, reshape_out, reshape_pattern);

    // get scales calculated after warmup, they scale variables to MAX=1.0
    auto scales = Get<VarQuantScale>("quant_var_scales");
    auto input_scale = scales[reshape_in->Name()].second.data<double>()[0];
    bool is_input_unsigned = scales[reshape_in->Name()].first;
    QuantizeInput(g, reshape_op, reshape_in, "X", input_scale,
                  is_input_unsigned);

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

    ++quantize_reshape_count;
  };

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

  PrettyLogDetail("---    quantized %d reshape ops", quantize_reshape_count);
}

528
void CPUQuantizePass::ApplyImpl(ir::Graph* graph) const {
529
  VLOG(3) << "Quantizing the graph.";
530 531
  PADDLE_ENFORCE(graph);
  FusePassBase::Init(name_scope_, graph);
532 533 534

  PADDLE_ENFORCE(param_scope());

535 536 537
  QuantizeConv(graph, false /* with_residual_data */);
  QuantizeConv(graph, true /* with_residual_data */);
  QuantizePool(graph);
538
  QuantizeConcat(graph);
539
  QuantizePriorBox(graph);
540
  QuantizeTranspose(graph);
M
Michał Gallus 已提交
541
  QuantizeFc(graph);
542
  QuantizeReshape(graph);
543 544 545 546 547 548 549 550
}

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

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