cpu_quantize_pass.cc 41.5 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"
W
wanghuancoder 已提交
16

17
#include <sstream>
18 19
#include <utility>
#include <vector>
W
wanghuancoder 已提交
20

21
#include "paddle/fluid/platform/mkldnn_helper.h"
22 23 24 25 26 27
#include "paddle/fluid/string/pretty_log.h"

namespace paddle {
namespace framework {
namespace ir {

28
using EigenVectorArrayMap = Eigen::Map<Eigen::Array<double, Eigen::Dynamic, 1>>;
29 30
using EigenVectorArrayMapFloat =
    Eigen::Map<Eigen::Array<float, Eigen::Dynamic, 1>>;
31 32
using string::PrettyLogDetail;

33 34 35 36 37 38 39 40 41
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());
}

42
void LogCannotQuantizeOp(Node* op, const char* details = nullptr) {
43 44 45
  std::stringstream msg_ss;
  msg_ss << "Cannot quantize operator " << op->Name()
         << " (type: " << op->Op()->Type() << ", id: " << op->id() << ").";
46
  if (details) msg_ss << " " << details;
47 48 49
  PrettyLogDetail(msg_ss.str().c_str());
}

50 51 52 53 54 55
void LogScaleIsMissingForVarName(const std::string& name) {
  VLOG(4) << "Quantization scale for the variable " << name << " is missing.";
}

void LogScaleIsMissingForVarNode(Node* node) {
  LogScaleIsMissingForVarName(node->Name());
56 57
}

58 59 60 61
void LogQuantizationDisabled(Node* op) {
  std::stringstream msg_ss;
  VLOG(4) << "Qantization skipped for operator " << op->Name()
          << " (type: " << op->Op()->Type() << ", id: " << op->id()
62
          << "). Attribute mkldnn_data_type != \"int8\".";
63 64
}

65 66 67 68 69 70
}  // namespace

enum { U8_MAX = 255, S8_MAX = 127 };

void CPUQuantizePass::QuantizeInput(Graph* g, Node* op, Node* input,
                                    std::string input_name, double scale_to_one,
71 72 73
                                    bool is_input_unsigned,
                                    std::string scale_attr_name, float shift,
                                    std::string shift_attr_name) const {
M
Michał Gallus 已提交
74 75 76
  auto inputs = op->Op()->InputNames();
  bool name_found =
      std::find(inputs.begin(), inputs.end(), input_name) != inputs.end();
77 78 79 80
  PADDLE_ENFORCE_EQ(name_found, true,
                    platform::errors::InvalidArgument(
                        "Var(%s) isn't the input of the %s operator.",
                        input_name, op->Op()->Type()));
81
  unsigned max = is_input_unsigned ? U8_MAX : S8_MAX;
82 83 84 85 86 87 88 89 90 91 92 93 94
  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);
95 96
  q_desc.SetAttr("Shift", shift);
  q_desc.SetAttr("is_negative_input", !is_input_unsigned);
97

Z
Zuza 已提交
98 99 100 101 102 103 104 105 106 107 108
  // fix to fc format error
  if (op->Op()->Type() == "fc" &&
      op->Op()->GetAttrIfExists<int>("in_num_col_dims") == 2) {
    q_desc.SetAttr("output_format", Has("data_layout")
                                        ? Get<std::string>("data_layout")
                                        : "NCHW");
  } else {
    q_desc.SetAttr("output_format", Has("data_layout")
                                        ? Get<std::string>("data_layout")
                                        : "NHWC");
  }
109 110 111 112 113 114 115 116 117 118 119 120 121
  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);
122
  if (!shift_attr_name.empty()) op->Op()->SetAttr(shift_attr_name, shift);
123 124
}

125
void CPUQuantizePass::QuantizeInputs(Graph* g, Node* op, std::string input_name,
126 127 128
                                     bool are_inputs_unsigned,
                                     std::string scale_attr_name, float shift,
                                     std::string shift_attr_name) const {
129
  auto inputs = op->inputs;
130
  auto output = op->outputs[0];
131 132 133 134 135 136 137 138
  PADDLE_ENFORCE_GE(inputs.size(), 1,
                    platform::errors::InvalidArgument(
                        "OP(%s)'s inputs(%d) must be equal or greater than 1.",
                        op->Name(), inputs.size()));
  PADDLE_ENFORCE_EQ(op->outputs.size(), 1,
                    platform::errors::InvalidArgument(
                        "OP(%s)'s outputs(%d) must be equal to 1.", op->Name(),
                        op->outputs.size()));
139 140 141 142 143 144 145 146

  // 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());

147
  double scale_out = GetScaleValueForNode(output);
148
  unsigned max = are_inputs_unsigned ? U8_MAX : S8_MAX;
149
  float scale = scale_out * max;
150 151 152 153 154 155 156 157

  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);
158
    q_desc.SetAttr("Shift", shift);
159 160 161
    q_desc.SetInput("Input", std::vector<std::string>({inputs[i]->Name()}));
    q_desc.SetOutput("Output",
                     std::vector<std::string>({quantize_out_node_names[i]}));
162
    q_desc.SetAttr("is_negative_input", !are_inputs_unsigned);
163 164 165 166 167 168 169 170 171 172 173 174 175
    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);
176
  if (!shift_attr_name.empty()) op->Op()->SetAttr(shift_attr_name, shift);
177 178
}

179 180 181 182
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
183 184 185 186 187
  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(
188 189
                        "Var(%s) isn't the output of the %s operator.",
                        output_name, op->Op()->Type()));
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
  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);
}

219 220 221 222 223 224 225 226 227 228 229 230 231
bool CPUQuantizePass::AreScalesPresentForVarNames(
    std::vector<std::string> names) const {
  auto& scales = Get<VarQuantScale>("quant_var_scales");
  bool present = true;
  for (auto name : names) {
    if (scales.find(name) == scales.end()) {
      present = false;
      LogScaleIsMissingForVarName(name);
    }
  }
  return present;
}

232
bool CPUQuantizePass::AreScalesPresentForNodes(
233
    std::initializer_list<Node*> nodes) const {
234 235 236 237 238
  auto& scales = Get<VarQuantScale>("quant_var_scales");
  bool present = true;
  for (auto node : nodes) {
    if (scales.count(node->Name()) == 0) {
      present = false;
239
      LogScaleIsMissingForVarNode(node);
240 241 242 243 244
    }
  }
  return present;
}

245 246 247 248 249 250
std::pair<bool, LoDTensor> CPUQuantizePass::GetScaleDataByName(
    const std::string& name) const {
  auto& scales = Get<VarQuantScale>("quant_var_scales");
  return scales.at(name);
}

251 252
std::pair<bool, LoDTensor> CPUQuantizePass::GetScaleDataForNode(
    const Node* node) const {
253 254 255 256 257
  return GetScaleDataByName(node->Name());
}

LoDTensor CPUQuantizePass::GetScaleTensorByName(const std::string& name) const {
  return GetScaleDataByName(name).second;
258 259 260 261 262 263 264 265 266 267 268 269 270
}

LoDTensor CPUQuantizePass::GetScaleTensorForNode(const Node* node) const {
  return GetScaleDataForNode(node).second;
}

double CPUQuantizePass::GetScaleValueForNode(const Node* node,
                                             bool* is_unsigned) const {
  auto scale_data = GetScaleDataForNode(node);
  if (is_unsigned != nullptr) *is_unsigned = scale_data.first;
  return scale_data.second.data<double>()[0];
}

271 272
bool CPUQuantizePass::IsOpDequantized(const Node* node) const {
  return node->Op()->Type() == "dequantize" ||
273
         platform::HasOpINT8DataType(node->Op());
274 275 276
}

bool CPUQuantizePass::IsOpQuantized(const Node* node) const {
277 278 279 280 281 282
  // return true only if all of outputs are ops and their are either quantize or
  // have int8 data type
  return all_of(node->outputs.begin(), node->outputs.end(), [](Node* output) {
    return (output->IsOp() && (output->Op()->Type() == "quantize" ||
                               platform::HasOpINT8DataType(output->Op())));
  });
283 284
}

285 286 287 288 289 290 291 292 293 294 295 296 297 298
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);

    // skip if should not be quantized
299
    if (!platform::HasOpINT8DataType(conv_op->Op())) {
300 301 302
      LogQuantizationDisabled(conv_op);
      return;
    }
303 304 305 306 307

    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);

308
    auto has_output_scale = AreScalesPresentForNodes({conv_output});
W
Wojciech Uss 已提交
309 310 311 312 313 314 315
    if (with_residual_data && !has_output_scale) {
      LogCannotQuantizeOp(conv_op,
                          "Conv op with ResidualData input cannot be quantized "
                          "without output scale.");
      return;
    }

316 317 318
    if (with_residual_data) {
      GET_IR_NODE_FROM_SUBGRAPH(conv_residual_data, conv_residual_data,
                                conv_pattern);
319
      if (!AreScalesPresentForNodes(
320
              {conv_input, conv_filter, conv_residual_data})) {
321
        LogCannotQuantizeOp(conv_op, "No scale available for the operator");
322
        return;
323
      }
324 325 326 327 328 329 330 331

      bool is_residual_unsigned{false};
      auto residual_scale =
          GetScaleValueForNode(conv_residual_data, &is_residual_unsigned);

      QuantizeInput(g, conv_op, conv_residual_data, "ResidualData",
                    residual_scale, is_residual_unsigned, "Scale_in_eltwise");
    } else {
332
      if (!AreScalesPresentForNodes({conv_input, conv_filter})) {
333
        LogCannotQuantizeOp(conv_op, "No scale available for the operator");
334
        return;
335
      }
336 337
    }

338 339
    bool is_input_unsigned{false};
    auto input_scale = GetScaleValueForNode(conv_input, &is_input_unsigned);
340 341 342
    QuantizeInput(g, conv_op, conv_input, "Input", input_scale,
                  is_input_unsigned, "Scale_in");

343
    auto filter_scale_tensor = GetScaleTensorForNode(conv_filter);
344
    EigenVectorArrayMap eigen_tensor{filter_scale_tensor.data<double>(),
345
                                     filter_scale_tensor.numel()};
346 347 348 349 350 351 352
    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);

353
    // if quantization scale is missing for output tensor, return fp32 data
W
Wojciech Uss 已提交
354
    if (has_output_scale) {
355 356 357 358 359 360 361 362
      bool is_output_unsigned{false};
      auto output_scale =
          GetScaleValueForNode(conv_output, &is_output_unsigned);
      DequantizeOutput(g, conv_op, conv_output, "Output", output_scale,
                       is_output_unsigned, "Scale_out");
    } else {
      conv_op->Op()->SetAttr("force_fp32_output", true);
    }
363

364
    // change threshold in bounded ReLu
365 366
    if (conv_op->Op()->GetAttrIfExists<std::string>("fuse_activation") ==
        "relu6") {
367 368 369 370
      float scale_out =
          BOOST_GET_CONST(float, conv_op->Op()->GetAttr("Scale_out"));
      float threshold =
          BOOST_GET_CONST(float, conv_op->Op()->GetAttr("fuse_alpha"));
371
      conv_op->Op()->SetAttr("fuse_alpha", scale_out * threshold);
372 373
    }

374 375 376 377 378 379 380 381 382 383 384 385
    ++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
386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402
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);

    // skip if should not be quantized
403
    if (!platform::HasOpINT8DataType(fc->Op())) {
404 405 406
      LogQuantizationDisabled(fc);
      return;
    }
407
    if (!fc->Op()->GetAttrIfExists<bool>("use_mkldnn")) {
408
      LogCannotQuantizeOp(fc, "use_mkldnn attribute set to false");
M
Michał Gallus 已提交
409
      return;
410
    }
M
Michał Gallus 已提交
411 412 413 414 415

    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);

416
    if (!AreScalesPresentForNodes({input, weights})) {
417
      LogCannotQuantizeOp(fc, "No scale available for the operator");
418 419
      return;
    }
420

421 422
    bool is_input_unsigned{false};
    auto input_scale = GetScaleValueForNode(input, &is_input_unsigned);
M
Michał Gallus 已提交
423 424 425
    QuantizeInput(g, fc, input, "Input", input_scale, is_input_unsigned,
                  "Scale_in");

426
    auto weight_scale_tensor = GetScaleTensorForNode(weights);
M
Michał Gallus 已提交
427
    EigenVectorArrayMap eigen_tensor{weight_scale_tensor.data<double>(),
428
                                     weight_scale_tensor.numel()};
M
Michał Gallus 已提交
429 430 431 432 433 434 435
    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);

436
    // if quantization scale is missing for output tensor, return fp32 data
437
    if (AreScalesPresentForNodes({output})) {
438 439 440 441 442 443 444
      bool is_output_unsigned{false};
      auto output_scale = GetScaleValueForNode(output, &is_output_unsigned);
      DequantizeOutput(g, fc, output, "Out", output_scale, is_output_unsigned,
                       "Scale_out");
    } else {
      fc->Op()->SetAttr("force_fp32_output", true);
    }
M
Michał Gallus 已提交
445 446 447 448 449 450 451 452 453 454 455 456

    ++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());
}

457 458 459 460 461 462 463 464 465 466 467 468 469
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);

    // skip if should not be quantized
470
    if (!platform::HasOpINT8DataType(pool_op->Op())) {
471 472 473
      LogQuantizationDisabled(pool_op);
      return;
    }
474 475 476 477

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

478
    if (!AreScalesPresentForNodes({pool_input, pool_output})) {
479
      LogCannotQuantizeOp(pool_op, "No scale available for the operator");
480 481
      return;
    }
482

483 484
    bool is_input_unsigned{false};
    auto input_scale = GetScaleValueForNode(pool_input, &is_input_unsigned);
485 486
    QuantizeInput(g, pool_op, pool_input, "X", input_scale, is_input_unsigned);

487 488
    bool is_output_unsigned{false};
    auto output_scale = GetScaleValueForNode(pool_output, &is_output_unsigned);
489 490 491 492 493 494 495 496 497 498 499 500
    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);
}

501 502 503 504 505 506 507 508 509 510 511 512 513
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);

    // skip if should not be quantized
514
    if (!platform::HasOpINT8DataType(concat_op->Op())) {
515 516 517
      LogQuantizationDisabled(concat_op);
      return;
    }
518 519 520

    GET_IR_NODE_FROM_SUBGRAPH(concat_out, concat_out, concat_pattern);

521
    if (!AreScalesPresentForNodes({concat_out})) {
522
      LogCannotQuantizeOp(concat_op, "No scale available for the operator");
523 524
      return;
    }
525

526 527
    // if all inputs were unsigned, then the output was set to unsigned
    // during the scale calculation step
528 529 530
    bool are_all_inputs_unsigned{false};
    auto output_scale =
        GetScaleValueForNode(concat_out, &are_all_inputs_unsigned);
531

532
    QuantizeInputs(g, concat_op, "X", are_all_inputs_unsigned);
533 534 535 536 537 538 539 540 541 542 543 544 545

    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);
}

546 547 548 549 550 551 552 553 554 555 556 557 558
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);

    // skip if should not be quantized
559
    if (!platform::HasOpINT8DataType(prior_box_op->Op())) {
560 561 562
      LogQuantizationDisabled(prior_box_op);
      return;
    }
563 564 565 566

    GET_IR_NODE_FROM_SUBGRAPH(prior_box_input, prior_box_input,
                              prior_box_pattern);

567
    if (!AreScalesPresentForNodes({prior_box_input})) {
568
      LogCannotQuantizeOp(prior_box_op, "No scale available for the operator");
569 570
      return;
    }
571

572 573 574
    bool is_input_unsigned{false};
    auto input_scale =
        GetScaleValueForNode(prior_box_input, &is_input_unsigned);
575 576 577 578 579 580 581 582 583 584 585 586 587
    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);
}

588 589 590 591 592 593 594 595 596 597 598 599 600
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);

    // skip if should not be quantized
601
    if (!platform::HasOpINT8DataType(transpose_op->Op())) {
602
      LogQuantizationDisabled(transpose_op);
603 604 605
      return;
    }
    GET_IR_NODE_FROM_SUBGRAPH(prev_op, prev_op, transpose_pattern);
606 607
    GET_IR_NODE_FROM_SUBGRAPH(transpose_in, transpose_in, transpose_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(transpose_out, transpose_out, transpose_pattern);
608

609
    // skip if prev op and next op is not quantized
610 611 612
    if (!(IsOpDequantized(prev_op)) && !(IsOpQuantized(transpose_out))) {
      LogCannotQuantizeOp(transpose_op,
                          "No other quantizable operators nearby");
613 614 615
      return;
    }

616
    if (!AreScalesPresentForNodes({transpose_in, transpose_out})) {
617
      LogCannotQuantizeOp(transpose_op, "No scale available for the operator");
618
      return;
619
    }
620

621 622
    bool is_input_unsigned{false};
    auto input_scale = GetScaleValueForNode(transpose_in, &is_input_unsigned);
623 624 625
    QuantizeInput(g, transpose_op, transpose_in, "X", input_scale,
                  is_input_unsigned);

626 627 628
    bool is_output_unsigned{false};
    auto output_scale =
        GetScaleValueForNode(transpose_out, &is_output_unsigned);
629 630 631 632 633 634 635 636 637 638 639 640 641
    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);
}

642 643 644 645 646 647 648 649 650 651 652 653 654
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);

    // skip if should not be quantized
655
    if (!platform::HasOpINT8DataType(reshape_op->Op())) {
656
      LogQuantizationDisabled(reshape_op);
657 658 659
      return;
    }
    GET_IR_NODE_FROM_SUBGRAPH(prev_op, prev_op, reshape_pattern);
660 661
    GET_IR_NODE_FROM_SUBGRAPH(reshape_in, reshape_in, reshape_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(reshape_out, reshape_out, reshape_pattern);
662

663 664 665
    // skip if prev op is not quantized
    if (!(IsOpDequantized(prev_op)) && !(IsOpQuantized(reshape_out))) {
      LogCannotQuantizeOp(reshape_op, "No other quantizable operators nearby");
666 667 668
      return;
    }

669
    if (!AreScalesPresentForNodes({reshape_in, reshape_out})) {
670
      LogCannotQuantizeOp(reshape_op, "No scale available for the operator");
671
      return;
672
    }
673

674 675
    bool is_input_unsigned{false};
    auto input_scale = GetScaleValueForNode(reshape_in, &is_input_unsigned);
676 677 678
    QuantizeInput(g, reshape_op, reshape_in, "X", input_scale,
                  is_input_unsigned);

679 680
    bool is_output_unsigned{false};
    auto output_scale = GetScaleValueForNode(reshape_out, &is_output_unsigned);
681 682 683 684 685 686 687 688 689 690 691 692
    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);
}

Z
Zuza 已提交
693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710
void CPUQuantizePass::QuantizeSlice(Graph* graph) const {
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
  patterns::Slice slice_pattern{pattern, name_scope_};
  slice_pattern();

  int quantize_slice_count = 0;
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
    VLOG(4) << "Quantize slice op";
    GET_IR_NODE_FROM_SUBGRAPH(slice_op, slice_op, slice_pattern);

    // skip if should not be quantized
    if (!platform::HasOpINT8DataType(slice_op->Op())) {
      LogQuantizationDisabled(slice_op);
      return;
    }
    GET_IR_NODE_FROM_SUBGRAPH(prev_op, prev_op, slice_pattern);
711 712
    GET_IR_NODE_FROM_SUBGRAPH(slice_in, slice_in, slice_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(slice_out, slice_out, slice_pattern);
Z
Zuza 已提交
713 714

    // skip if prev op and next op is not quantized
715 716
    if (!IsOpDequantized(prev_op) && !IsOpQuantized(slice_out)) {
      LogCannotQuantizeOp(slice_op, "No other quantizable operators nearby");
Z
Zuza 已提交
717 718 719 720
      return;
    }

    if (!AreScalesPresentForNodes({slice_out})) {
721
      LogCannotQuantizeOp(slice_op, "No scale available for the operator");
Z
Zuza 已提交
722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743
      return;
    }

    bool is_input_unsigned{false};
    auto input_scale = GetScaleValueForNode(slice_out, &is_input_unsigned);
    QuantizeInput(g, slice_op, slice_in, "Input", input_scale,
                  is_input_unsigned);

    bool is_output_unsigned{false};
    auto output_scale = GetScaleValueForNode(slice_out, &is_output_unsigned);
    DequantizeOutput(g, slice_op, slice_out, "Out", output_scale,
                     is_output_unsigned);

    ++quantize_slice_count;
  };

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

  PrettyLogDetail("---    quantized %d slice ops", quantize_slice_count);
}

744 745 746
void CPUQuantizePass::QuantizeMatmul(Graph* graph) const {
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
747
  patterns::MatmulWithInputOps matmul_pattern{pattern, name_scope_};
748 749 750 751 752 753 754 755 756
  matmul_pattern();

  int quantize_matmul_count = 0;
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
    VLOG(4) << "Quantize matmul op";
    GET_IR_NODE_FROM_SUBGRAPH(matmul_op, matmul_op, matmul_pattern);

    // skip if should not be quantized
757
    if (!platform::HasOpINT8DataType(matmul_op->Op())) {
758
      LogQuantizationDisabled(matmul_op);
759 760 761 762 763 764 765
      return;
    }
    GET_IR_NODE_FROM_SUBGRAPH(prev_op_x, prev_op_x, matmul_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(prev_op_y, prev_op_y, matmul_pattern);

    // skip if prev ops are not quantized
    if (!IsOpDequantized(prev_op_x) || !IsOpDequantized(prev_op_y)) {
766
      LogCannotQuantizeOp(matmul_op, "No other quantizable operators nearby");
767 768 769 770 771 772
      return;
    }
    GET_IR_NODE_FROM_SUBGRAPH(matmul_in_x, matmul_in_x, matmul_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(matmul_in_y, matmul_in_y, matmul_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(matmul_out, matmul_out, matmul_pattern);

773
    if (!AreScalesPresentForNodes({matmul_in_x, matmul_in_y})) {
774
      LogCannotQuantizeOp(matmul_op, "No scale available for the operator");
775
      return;
776
    }
777

778 779 780
    bool is_x_unsigned{false}, is_y_unsigned{false};
    auto input_x_scale = GetScaleValueForNode(matmul_in_x, &is_x_unsigned);
    auto input_y_scale = GetScaleValueForNode(matmul_in_y, &is_y_unsigned);
781 782 783 784 785 786
    PADDLE_ENFORCE_EQ(is_x_unsigned, is_y_unsigned,
                      platform::errors::InvalidArgument(
                          "Matmul inputs should have the same "
                          "attribute of signed/unsigned, but they "
                          "are different: x(%d), y(%d).",
                          is_x_unsigned, is_y_unsigned));
787 788 789 790 791
    QuantizeInput(g, matmul_op, matmul_in_x, "X", input_x_scale, is_x_unsigned,
                  "Scale_x");
    QuantizeInput(g, matmul_op, matmul_in_y, "Y", input_y_scale, is_y_unsigned,
                  "Scale_y");

792
    // if quantization scale is missing for output tensor, return fp32 data
793
    if (AreScalesPresentForNodes({matmul_out})) {
794 795 796 797 798 799 800
      bool is_output_unsigned{false};
      auto output_scale = GetScaleValueForNode(matmul_out, &is_output_unsigned);
      DequantizeOutput(g, matmul_op, matmul_out, "Out", output_scale,
                       is_output_unsigned, "Scale_out");
    } else {
      matmul_op->Op()->SetAttr("force_fp32_output", true);
    }
801 802 803 804 805 806 807 808 809

    ++quantize_matmul_count;
  };
  gpd(graph, handler);
  AddStatis(quantize_matmul_count);

  PrettyLogDetail("---    quantized %d matmul ops", quantize_matmul_count);
}

810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826
void CPUQuantizePass::QuantizeElementwiseAdd(Graph* graph) const {
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
  patterns::ElementwiseAdd elementwise_add_pattern{pattern, name_scope_};

  elementwise_add_pattern(
      pattern->NewNode(elementwise_add_pattern.elementwise_add_x_repr()),
      pattern->NewNode(elementwise_add_pattern.elementwise_add_y_repr()));

  int quantize_elementwise_add_count = 0;
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
    VLOG(4) << "Quantize elementwise_add op";
    GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_op, elementwise_add_op,
                              elementwise_add_pattern);

    // skip if should not be quantized
827
    if (!platform::HasOpINT8DataType(elementwise_add_op->Op())) {
828 829 830 831 832 833 834 835 836 837 838
      LogQuantizationDisabled(elementwise_add_op);
      return;
    }

    GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_x, elementwise_add_x,
                              elementwise_add_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_y, elementwise_add_y,
                              elementwise_add_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_out, elementwise_add_out,
                              elementwise_add_pattern);

839 840
    if (!AreScalesPresentForNodes(
            {elementwise_add_x, elementwise_add_y, elementwise_add_out})) {
841 842
      LogCannotQuantizeOp(elementwise_add_op,
                          "No scale available for the operator");
843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863
      return;
    }

    bool is_x_unsigned{false}, is_y_unsigned{false};
    auto input_x_scale =
        GetScaleValueForNode(elementwise_add_x, &is_x_unsigned);
    auto input_y_scale =
        GetScaleValueForNode(elementwise_add_y, &is_y_unsigned);

    // TODO(sfraczek): add support for different signness
    if (is_x_unsigned != is_y_unsigned) {
      LogCannotQuantizeOp(elementwise_add_op,
                          "ElementwiseAdd inputs must be of the same type.");
      return;
    }

    QuantizeInput(g, elementwise_add_op, elementwise_add_x, "X", input_x_scale,
                  is_x_unsigned, "Scale_x");
    QuantizeInput(g, elementwise_add_op, elementwise_add_y, "Y", input_y_scale,
                  is_y_unsigned, "Scale_y");

864 865 866 867 868 869
    bool is_output_unsigned{false};
    auto output_scale =
        GetScaleValueForNode(elementwise_add_out, &is_output_unsigned);

    DequantizeOutput(g, elementwise_add_op, elementwise_add_out, "Out",
                     output_scale, is_output_unsigned, "Scale_out");
870 871 872 873 874 875 876 877 878 879

    ++quantize_elementwise_add_count;
  };
  gpd(graph, handler);
  AddStatis(quantize_elementwise_add_count);

  PrettyLogDetail("---    quantized %d elementwise_add ops",
                  quantize_elementwise_add_count);
}

880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901
void CPUQuantizePass::QuantizeFusionGru(Graph* graph) const {
  GraphPatternDetector gpd;
  patterns::FusionGru pattern{gpd.mutable_pattern(), name_scope_};
  pattern();

  int quantize_count = 0;
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
    VLOG(4) << "Quantize fusion_gru op";
    GET_IR_NODE_FROM_SUBGRAPH(op, op, pattern);

    // skip if should not be quantized
    if (!platform::HasOpINT8DataType(op->Op())) {
      LogQuantizationDisabled(op);
      return;
    }

    GET_IR_NODE_FROM_SUBGRAPH(x, x, pattern);
    GET_IR_NODE_FROM_SUBGRAPH(weight_h, weight_h, pattern);
    GET_IR_NODE_FROM_SUBGRAPH(weight_x, weight_x, pattern);
    GET_IR_NODE_FROM_SUBGRAPH(out, out, pattern);

902
    if (!AreScalesPresentForNodes({x, weight_x})) {
903
      LogCannotQuantizeOp(op, "No scale available for the operator");
904 905 906 907 908 909 910 911 912 913 914 915 916 917
      return;
    }

    bool is_x_unsigned{false};
    auto input_x_scale = GetScaleValueForNode(x, &is_x_unsigned);

    double input_x_shift{128.};
    if (is_x_unsigned) input_x_shift = 0.;

    QuantizeInput(g, op, x, "X", input_x_scale, is_x_unsigned, "Scale_data",
                  input_x_shift, "Shift_data");

    auto weight_scale_tensor = GetScaleTensorForNode(weight_x);
    EigenVectorArrayMap eigen_tensor{weight_scale_tensor.data<double>(),
918
                                     weight_scale_tensor.numel()};
919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935
    eigen_tensor *= static_cast<double>(S8_MAX);
    std::vector<float> scale_weights{
        weight_scale_tensor.data<double>(),
        weight_scale_tensor.data<double>() + weight_scale_tensor.numel()};

    op->Op()->SetAttr("Scale_weights", scale_weights);
    // return fp32 data
    op->Op()->SetAttr("force_fp32_output", true);

    ++quantize_count;
  };
  gpd(graph, handler);
  AddStatis(quantize_count);

  PrettyLogDetail("---    quantized %d fusion_gru ops", quantize_count);
}

936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959
void CPUQuantizePass::QuantizeMultiGru(Graph* graph) const {
  GraphPatternDetector gpd;
  patterns::MultiGru pattern{gpd.mutable_pattern(), name_scope_};
  pattern();

  int quantize_count = 0;
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
    VLOG(4) << "Quantize multi_gru op";
    GET_IR_NODE_FROM_SUBGRAPH(gru, gru, pattern);

    // skip if should not be quantized
    if (!platform::HasOpINT8DataType(gru->Op())) {
      LogQuantizationDisabled(gru);
      return;
    }

    GET_IR_NODE_FROM_SUBGRAPH(x, x, pattern);
    GET_IR_NODE_FROM_SUBGRAPH(wx, wx, pattern);
    GET_IR_NODE_FROM_SUBGRAPH(h, h, pattern);

    auto wx_names = gru->Op()->Input("WeightX");
    if (!AreScalesPresentForNodes({x}) ||
        !AreScalesPresentForVarNames(wx_names)) {
960
      LogCannotQuantizeOp(gru, "No scale available for the operator");
961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013
      return;
    }

    bool is_x_unsigned{false};
    auto input_x_scale = GetScaleValueForNode(x, &is_x_unsigned);

    double input_x_shift{128.};
    if (is_x_unsigned) input_x_shift = 0.;

    QuantizeInput(g, gru, x, "X", input_x_scale, is_x_unsigned, "Scale_data",
                  input_x_shift, "Shift_data");

    auto* scope = param_scope();
    int wx_size = wx_names.size();
    std::vector<std::string> w_scale_var_names;
    for (int i = 0; i < wx_size; ++i) {
      auto scale_tensor_src = GetScaleTensorByName(wx_names[i]);
      EigenVectorArrayMap eigen_tensor_src{scale_tensor_src.data<double>(),
                                           scale_tensor_src.numel()};

      VarDesc scale_var_desc(patterns::PDNodeName("multi_gru", "w_scale"));

      scale_var_desc.SetShape(framework::vectorize(scale_tensor_src.dims()));
      scale_var_desc.SetDataType(proto::VarType::FP32);
      scale_var_desc.SetLoDLevel(scale_tensor_src.lod().size());
      scale_var_desc.SetPersistable(true);
      auto* w_scale_node = g->CreateVarNode(&scale_var_desc);

      auto* w_scale_tensor_dst =
          scope->Var(w_scale_node->Name())->GetMutable<LoDTensor>();
      w_scale_tensor_dst->Resize(scale_tensor_src.dims());
      auto* dst_data =
          w_scale_tensor_dst->mutable_data<float>(platform::CPUPlace());
      EigenVectorArrayMapFloat eigen_tensor_dst{dst_data,
                                                w_scale_tensor_dst->numel()};
      eigen_tensor_dst =
          eigen_tensor_src.cast<float>() * static_cast<float>(S8_MAX);
      w_scale_var_names.push_back(w_scale_node->Name());
      IR_NODE_LINK_TO(w_scale_node, gru);
    }

    gru->Op()->SetInput("Scale_weights", w_scale_var_names);
    // return fp32 data
    gru->Op()->SetAttr("force_fp32_output", true);

    ++quantize_count;
  };
  gpd(graph, handler);
  AddStatis(quantize_count);

  PrettyLogDetail("---    quantized %d multi_gru ops", quantize_count);
}

1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038
void CPUQuantizePass::QuantizeFusionLSTM(Graph* graph) const {
  GraphPatternDetector gpd;
  patterns::FusionLSTM pattern{gpd.mutable_pattern(), name_scope_};
  pattern();

  int quantize_count = 0;
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
    VLOG(4) << "Quantize fusion_lstm op";
    GET_IR_NODE_FROM_SUBGRAPH(op, op, pattern);

    // skip if should not be quantized
    if (!platform::HasOpINT8DataType(op->Op())) {
      LogQuantizationDisabled(op);
      return;
    }

    GET_IR_NODE_FROM_SUBGRAPH(x, x, pattern);
    GET_IR_NODE_FROM_SUBGRAPH(weight_h, weight_h, pattern);
    GET_IR_NODE_FROM_SUBGRAPH(weight_x, weight_x, pattern);
    GET_IR_NODE_FROM_SUBGRAPH(hidden, hidden, pattern);
    GET_IR_NODE_FROM_SUBGRAPH(cell, cell, pattern);

    // Starting from here there maybe issues
    if (!AreScalesPresentForNodes({x, weight_x})) {
1039
      LogCannotQuantizeOp(op, "No scale available for the operator");
1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071
      return;
    }

    bool is_x_unsigned{false};
    auto input_x_scale = GetScaleValueForNode(x, &is_x_unsigned);

    double input_x_shift{128.};
    if (is_x_unsigned) input_x_shift = 0.;

    QuantizeInput(g, op, x, "X", input_x_scale, is_x_unsigned, "Scale_data",
                  input_x_shift, "Shift_data");

    auto weight_scale_tensor = GetScaleTensorForNode(weight_x);
    EigenVectorArrayMap eigen_tensor{weight_scale_tensor.data<double>(),
                                     weight_scale_tensor.numel()};
    eigen_tensor *= static_cast<double>(S8_MAX);
    std::vector<float> scale_weights{
        weight_scale_tensor.data<double>(),
        weight_scale_tensor.data<double>() + weight_scale_tensor.numel()};

    op->Op()->SetAttr("Scale_weights", scale_weights);
    // return fp32 data
    op->Op()->SetAttr("force_fp32_output", true);

    ++quantize_count;
  };
  gpd(graph, handler);
  AddStatis(quantize_count);

  PrettyLogDetail("---    quantized %d fusion_lstm ops", quantize_count);
}

1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090
void CPUQuantizePass::QuantizeNearestInterp(Graph* graph) const {
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
  patterns::NearestInterp nearest_interp_pattern{pattern, name_scope_};
  nearest_interp_pattern();

  int quantize_nearest_interp_count = 0;
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
    VLOG(4) << "Quantize nearest_interp op";
    GET_IR_NODE_FROM_SUBGRAPH(nearest_interp_op, nearest_interp_op,
                              nearest_interp_pattern);

    // skip if should not be quantized
    if (!platform::HasOpINT8DataType(nearest_interp_op->Op())) {
      LogQuantizationDisabled(nearest_interp_op);
      return;
    }
    GET_IR_NODE_FROM_SUBGRAPH(prev_op, prev_op, nearest_interp_pattern);
1091 1092 1093 1094
    GET_IR_NODE_FROM_SUBGRAPH(nearest_interp_in, nearest_interp_in,
                              nearest_interp_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(nearest_interp_out, nearest_interp_out,
                              nearest_interp_pattern);
1095 1096

    // skip if prev op and next op is not quantized
1097
    if (!(IsOpDequantized(prev_op)) && !(IsOpQuantized(nearest_interp_out))) {
1098
      LogCannotQuantizeOp(nearest_interp_op,
1099
                          "No other quantizable operators nearby");
1100 1101 1102 1103
      return;
    }

    if (!AreScalesPresentForNodes({nearest_interp_in, nearest_interp_out})) {
1104 1105
      LogCannotQuantizeOp(nearest_interp_op,
                          "No scale available for the operator");
1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130
      return;
    }

    bool is_input_unsigned{false};
    auto input_scale =
        GetScaleValueForNode(nearest_interp_in, &is_input_unsigned);
    QuantizeInput(g, nearest_interp_op, nearest_interp_in, "X", input_scale,
                  is_input_unsigned);

    bool is_output_unsigned{false};
    auto output_scale =
        GetScaleValueForNode(nearest_interp_out, &is_output_unsigned);
    DequantizeOutput(g, nearest_interp_op, nearest_interp_out, "Out",
                     output_scale, is_output_unsigned);

    ++quantize_nearest_interp_count;
  };

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

  PrettyLogDetail("---    quantized %d nearest_interp ops",
                  quantize_nearest_interp_count);
}

1131
void CPUQuantizePass::ApplyImpl(ir::Graph* graph) const {
1132
  VLOG(3) << "Quantizing the graph.";
1133 1134
  PADDLE_ENFORCE_NOT_NULL(
      graph, platform::errors::InvalidArgument("Graph cannot be nullptr."));
1135
  FusePassBase::Init(name_scope_, graph);
1136

1137 1138
  PADDLE_ENFORCE_NOT_NULL(param_scope(), platform::errors::InvalidArgument(
                                             "Scope cannot be nullptr."));
1139

1140 1141 1142
  QuantizeConv(graph, false /* with_residual_data */);
  QuantizeConv(graph, true /* with_residual_data */);
  QuantizePool(graph);
1143
  QuantizeConcat(graph);
1144
  QuantizePriorBox(graph);
1145
  QuantizeTranspose(graph);
M
Michał Gallus 已提交
1146
  QuantizeFc(graph);
1147
  QuantizeReshape(graph);
1148
  QuantizeMatmul(graph);
1149
  QuantizeElementwiseAdd(graph);
1150
  QuantizeFusionGru(graph);
1151
  QuantizeMultiGru(graph);
1152
  QuantizeFusionLSTM(graph);
Z
Zuza 已提交
1153
  QuantizeSlice(graph);
1154
  QuantizeNearestInterp(graph);
1155 1156 1157 1158 1159 1160 1161 1162
}

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

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