cpu_quantize_pass.cc 43.3 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 16
#include "paddle/fluid/framework/ir/mkldnn/cpu_quantize_pass.h"

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

B
baoachun 已提交
21
#include "paddle/fluid/framework/ir/mkldnn/mkldnn_pass_util.h"
22
#include "paddle/fluid/platform/mkldnn_helper.h"
23 24 25 26 27 28
#include "paddle/fluid/string/pretty_log.h"

namespace paddle {
namespace framework {
namespace ir {

29
using EigenVectorArrayMap = Eigen::Map<Eigen::Array<double, Eigen::Dynamic, 1>>;
30 31
using EigenVectorArrayMapFloat =
    Eigen::Map<Eigen::Array<float, Eigen::Dynamic, 1>>;
32 33
using string::PrettyLogDetail;

34 35 36 37 38 39 40 41 42
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());
}

43
void MarkAndLogCannotQuantizeOp(Node* op, const char* details = nullptr) {
44 45 46
  std::stringstream msg_ss;
  msg_ss << "Cannot quantize operator " << op->Name()
         << " (type: " << op->Op()->Type() << ", id: " << op->id() << ").";
47
  if (details) msg_ss << " " << details;
48 49
  VLOG(2) << msg_ss.str().c_str();
  op->Op()->SetAttr("mkldnn_data_type", std::string("float32"));
50 51
}

52 53 54 55 56 57
void LogScaleIsMissingForVarName(const std::string& name) {
  VLOG(4) << "Quantization scale for the variable " << name << " is missing.";
}

void LogScaleIsMissingForVarNode(Node* node) {
  LogScaleIsMissingForVarName(node->Name());
58 59
}

60
void LogQuantizationDisabled(Node* op) {
61
  VLOG(2) << "Quantization skipped for operator " << op->Name()
62
          << " (type: " << op->Op()->Type() << ", id: " << op->id()
63
          << "). Attribute mkldnn_data_type != \"int8\".";
64 65
}

66 67
void LogQuantizedOpsCounter(const std::string& type,
                            const int counter,
68 69 70 71 72 73 74
                            const char* details = nullptr) {
  std::stringstream msg_ss;
  msg_ss << "---    quantized " << counter << " " << type << " ops";
  if (details) msg_ss << " " << details;
  PrettyLogDetail(msg_ss.str().c_str());
}

75 76 77 78
}  // namespace

enum { U8_MAX = 255, S8_MAX = 127 };

79 80 81 82 83
void CPUQuantizePass::QuantizeInput(Graph* g,
                                    Node* op,
                                    Node* input,
                                    std::string input_name,
                                    double scale_to_one,
84
                                    bool is_input_unsigned,
85 86
                                    std::string scale_attr_name,
                                    float shift,
87
                                    std::string shift_attr_name) const {
M
Michał Gallus 已提交
88 89 90
  auto inputs = op->Op()->InputNames();
  bool name_found =
      std::find(inputs.begin(), inputs.end(), input_name) != inputs.end();
91 92
  PADDLE_ENFORCE_EQ(name_found,
                    true,
93 94
                    platform::errors::InvalidArgument(
                        "Var(%s) isn't the input of the %s operator.",
95 96
                        input_name,
                        op->Op()->Type()));
97
  unsigned max = is_input_unsigned ? U8_MAX : S8_MAX;
98 99 100 101 102 103 104 105 106 107 108 109 110
  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);
111 112
  q_desc.SetAttr("Shift", shift);
  q_desc.SetAttr("is_negative_input", !is_input_unsigned);
113

Z
Zuza 已提交
114 115 116
  // fix to fc format error
  if (op->Op()->Type() == "fc" &&
      op->Op()->GetAttrIfExists<int>("in_num_col_dims") == 2) {
117 118 119
    q_desc.SetAttr(
        "output_format",
        Has("data_layout") ? Get<std::string>("data_layout") : "NCHW");
Z
Zuza 已提交
120
  } else {
121 122 123
    q_desc.SetAttr(
        "output_format",
        Has("data_layout") ? Get<std::string>("data_layout") : "NHWC");
Z
Zuza 已提交
124
  }
125 126 127 128 129 130 131 132 133 134 135 136 137
  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);
138
  if (!shift_attr_name.empty()) op->Op()->SetAttr(shift_attr_name, shift);
139 140
}

141 142 143
void CPUQuantizePass::QuantizeInputs(Graph* g,
                                     Node* op,
                                     std::string input_name,
144
                                     bool are_inputs_unsigned,
145 146
                                     std::string scale_attr_name,
                                     float shift,
147
                                     std::string shift_attr_name) const {
148
  auto inputs = op->inputs;
149
  auto output = op->outputs[0];
150 151
  PADDLE_ENFORCE_GE(inputs.size(),
                    1,
152 153
                    platform::errors::InvalidArgument(
                        "OP(%s)'s inputs(%d) must be equal or greater than 1.",
154 155 156 157
                        op->Name(),
                        inputs.size()));
  PADDLE_ENFORCE_EQ(op->outputs.size(),
                    1,
158
                    platform::errors::InvalidArgument(
159 160
                        "OP(%s)'s outputs(%d) must be equal to 1.",
                        op->Name(),
161
                        op->outputs.size()));
162 163 164 165 166 167 168 169

  // 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());

170
  double scale_out = GetScaleValueForNode(output);
171
  unsigned max = are_inputs_unsigned ? U8_MAX : S8_MAX;
172
  float scale = scale_out * max;
173 174 175 176 177 178 179 180

  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);
181
    q_desc.SetAttr("Shift", shift);
182 183 184
    q_desc.SetInput("Input", std::vector<std::string>({inputs[i]->Name()}));
    q_desc.SetOutput("Output",
                     std::vector<std::string>({quantize_out_node_names[i]}));
185
    q_desc.SetAttr("is_negative_input", !are_inputs_unsigned);
186 187 188 189 190 191 192 193 194 195 196 197 198
    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);
199
  if (!shift_attr_name.empty()) op->Op()->SetAttr(shift_attr_name, shift);
200 201
}

202 203 204
void CPUQuantizePass::DequantizeOutput(Graph* g,
                                       Node* op,
                                       Node* output,
205
                                       std::string output_name,
206 207
                                       double scale_to_one,
                                       bool is_unsigned,
208
                                       std::string scale_attr_name) const {
M
Michał Gallus 已提交
209 210 211
  auto outputs = op->Op()->OutputNames();
  bool name_found =
      std::find(outputs.begin(), outputs.end(), output_name) != outputs.end();
212 213
  PADDLE_ENFORCE_EQ(name_found,
                    true,
M
Michał Gallus 已提交
214
                    platform::errors::InvalidArgument(
215
                        "Var(%s) isn't the output of the %s operator.",
216 217
                        output_name,
                        op->Op()->Type()));
218 219 220 221 222 223 224 225 226 227 228 229 230 231
  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);
232
  deq_desc.SetAttr("is_negative_input", !is_unsigned);
233 234 235 236 237 238 239 240 241 242 243 244 245 246 247
  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);
}

P
Paulina Gacek 已提交
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
void CPUQuantizePass::DequantizeOutputs(Graph* g,
                                        Node* op,
                                        std::string output_name,
                                        double scale_to_one,
                                        bool is_unsigned,
                                        std::string scale_attr_name) const {
  auto outputs = op->outputs;
  PADDLE_ENFORCE_GE(outputs.size(),
                    1,
                    platform::errors::InvalidArgument(
                        "OP(%s)'s outputs(%d) must be equal or greater than 1.",
                        op->Name(),
                        outputs.size()));

  std::vector<std::string> quantize_in_node_names(outputs.size());

  unsigned max = is_unsigned ? U8_MAX : S8_MAX;
  float scale = scale_to_one * max;

  for (size_t i = 0; i < outputs.size(); i++) {
    // Create dequantize input variable
    VarDesc dequantize_in_desc(patterns::PDNodeName("dequantize", "in"));
    Node* dequantize_in_node = g->CreateVarNode(&dequantize_in_desc);
    quantize_in_node_names[i] = dequantize_in_node->Name();

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

    // link dequantize op
    UnlinkNodes(op, outputs[i]);
    IR_NODE_LINK_TO(op, dequantize_in_node);
    IR_NODE_LINK_TO(dequantize_in_node, dequantize_op);
    IR_NODE_LINK_TO(dequantize_op, outputs[i]);
  }

  // update op's output
  op->Op()->SetOutput(output_name, quantize_in_node_names);
  if (!scale_attr_name.empty()) op->Op()->SetAttr(scale_attr_name, scale);
}

296 297 298
bool CPUQuantizePass::AreScalesPresentForVarNames(
    std::vector<std::string> names) const {
  bool present = true;
B
baoachun 已提交
299 300 301 302 303 304 305 306 307 308 309 310 311 312
  if (var_quant_scales_->empty()) {
    auto& scales = Get<VarQuantScale>("quant_var_scales");
    for (auto name : names) {
      if (scales.find(name) == scales.end()) {
        present = false;
        LogScaleIsMissingForVarName(name);
      }
    }
  } else {
    for (auto name : names) {
      if (var_quant_scales_->find(name) == var_quant_scales_->end()) {
        present = false;
        LogScaleIsMissingForVarName(name);
      }
313 314 315 316 317
    }
  }
  return present;
}

318
bool CPUQuantizePass::AreScalesPresentForNodes(
319
    std::initializer_list<Node*> nodes) const {
320
  bool present = true;
B
baoachun 已提交
321 322 323 324 325 326 327 328 329 330 331 332 333 334
  if (var_quant_scales_->empty()) {
    auto& scales = Get<VarQuantScale>("quant_var_scales");
    for (auto node : nodes) {
      if (scales.count(node->Name()) == 0) {
        present = false;
        LogScaleIsMissingForVarNode(node);
      }
    }
  } else {
    for (auto node : nodes) {
      if (var_quant_scales_->count(node->Name()) == 0) {
        present = false;
        LogScaleIsMissingForVarNode(node);
      }
335 336 337 338 339
    }
  }
  return present;
}

340
std::pair<bool, phi::DenseTensor> CPUQuantizePass::GetScaleDataByName(
341
    const std::string& name) const {
B
baoachun 已提交
342 343 344 345 346
  if (var_quant_scales_->empty()) {
    auto& scales = Get<VarQuantScale>("quant_var_scales");
    return scales.at(name);
  }
  return var_quant_scales_->at(name);
347 348
}

349
std::pair<bool, phi::DenseTensor> CPUQuantizePass::GetScaleDataForNode(
350
    const Node* node) const {
351 352 353
  return GetScaleDataByName(node->Name());
}

354 355
phi::DenseTensor CPUQuantizePass::GetScaleTensorByName(
    const std::string& name) const {
356
  return GetScaleDataByName(name).second;
357 358
}

359 360
phi::DenseTensor CPUQuantizePass::GetScaleTensorForNode(
    const Node* node) const {
361 362 363 364 365 366 367 368 369 370
  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];
}

371 372
bool CPUQuantizePass::IsOpDequantized(const Node* node) const {
  return node->Op()->Type() == "dequantize" ||
373
         platform::HasOpINT8DataType(node->Op());
374 375 376
}

bool CPUQuantizePass::IsOpQuantized(const Node* node) const {
377 378 379 380 381 382
  // 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())));
  });
383 384
}

B
baoachun 已提交
385
void CPUQuantizePass::GetQuantInfo(Graph* graph) const {
386 387
  GetInfoFromTheFirstOp(
      graph, "has_quant_info", "var_quant_scales", var_quant_scales_);
B
baoachun 已提交
388 389
}

390 391 392 393 394 395 396 397 398 399 400 401 402 403
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
404
    if (!platform::HasOpINT8DataType(conv_op->Op())) {
405 406 407
      LogQuantizationDisabled(conv_op);
      return;
    }
408 409 410 411 412

    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);

413
    auto has_output_scale = AreScalesPresentForNodes({conv_output});
W
Wojciech Uss 已提交
414
    if (with_residual_data && !has_output_scale) {
415 416 417 418
      MarkAndLogCannotQuantizeOp(
          conv_op,
          "Conv op with ResidualData input cannot be quantized "
          "without output scale.");
W
Wojciech Uss 已提交
419 420 421
      return;
    }

422
    if (with_residual_data) {
423 424
      GET_IR_NODE_FROM_SUBGRAPH(
          conv_residual_data, conv_residual_data, conv_pattern);
425
      if (!AreScalesPresentForNodes(
426
              {conv_input, conv_filter, conv_residual_data})) {
427 428
        MarkAndLogCannotQuantizeOp(conv_op,
                                   "No scale available for the operator");
429
        return;
430
      }
431 432 433 434 435

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

436 437 438 439 440 441 442
      QuantizeInput(g,
                    conv_op,
                    conv_residual_data,
                    "ResidualData",
                    residual_scale,
                    is_residual_unsigned,
                    "Scale_in_eltwise");
443
    } else {
444
      if (!AreScalesPresentForNodes({conv_input, conv_filter})) {
445 446
        MarkAndLogCannotQuantizeOp(conv_op,
                                   "No scale available for the operator");
447
        return;
448
      }
449 450
    }

451 452
    bool is_input_unsigned{false};
    auto input_scale = GetScaleValueForNode(conv_input, &is_input_unsigned);
453 454 455 456 457 458 459
    QuantizeInput(g,
                  conv_op,
                  conv_input,
                  "Input",
                  input_scale,
                  is_input_unsigned,
                  "Scale_in");
460

461
    auto filter_scale_tensor = GetScaleTensorForNode(conv_filter);
462
    EigenVectorArrayMap eigen_tensor{filter_scale_tensor.data<double>(),
463
                                     filter_scale_tensor.numel()};
464 465 466 467 468 469 470 471 472 473

    // If the scale value of a weight is already multiplied by S8_MAX, it does
    // not need to be multiplied again
    if (std::find(change_weight_->begin(),
                  change_weight_->end(),
                  conv_filter->Name()) == change_weight_->end()) {
      eigen_tensor *= static_cast<double>(S8_MAX);
      change_weight_->push_back(conv_filter->Name());
    }

474 475 476 477 478 479
    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);

480
    // if quantization scale is missing for output tensor, return fp32 data
W
Wojciech Uss 已提交
481
    if (has_output_scale) {
482 483 484
      bool is_output_unsigned{false};
      auto output_scale =
          GetScaleValueForNode(conv_output, &is_output_unsigned);
485 486 487 488 489 490 491
      DequantizeOutput(g,
                       conv_op,
                       conv_output,
                       "Output",
                       output_scale,
                       is_output_unsigned,
                       "Scale_out");
492 493 494
    } else {
      conv_op->Op()->SetAttr("force_fp32_output", true);
    }
495

496
    // change threshold in bounded ReLu
497 498
    if (conv_op->Op()->GetAttrIfExists<std::string>("fuse_activation") ==
        "relu6") {
499
      float scale_out =
R
Ruibiao Chen 已提交
500
          PADDLE_GET_CONST(float, conv_op->Op()->GetAttr("Scale_out"));
501
      float threshold =
R
Ruibiao Chen 已提交
502
          PADDLE_GET_CONST(float, conv_op->Op()->GetAttr("fuse_alpha"));
503
      conv_op->Op()->SetAttr("fuse_alpha", scale_out * threshold);
504 505
    }

506 507 508 509 510 511
    ++quantize_conv_count;
  };

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

512
  LogQuantizedOpsCounter(
513 514
      "conv2d",
      quantize_conv_count,
515
      ((with_residual_data) ? "with residual connection" : ""));
516 517
}

M
Michał Gallus 已提交
518 519 520 521
void CPUQuantizePass::QuantizeFc(Graph* graph) const {
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
  patterns::FCMKLDNN fc_pattern{pattern, name_scope_};
522
  fc_pattern(false /* with_residual */);
M
Michał Gallus 已提交
523 524 525 526 527 528 529 530

  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
531
    if (!platform::HasOpINT8DataType(fc->Op())) {
532 533 534
      LogQuantizationDisabled(fc);
      return;
    }
535
    if (!fc->Op()->GetAttrIfExists<bool>("use_mkldnn")) {
536
      MarkAndLogCannotQuantizeOp(fc, "use_mkldnn attribute set to false");
M
Michał Gallus 已提交
537
      return;
538
    }
M
Michał Gallus 已提交
539 540 541 542 543

    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);

544
    if (!AreScalesPresentForNodes({input, weights})) {
545
      MarkAndLogCannotQuantizeOp(fc, "No scale available for the operator");
546 547
      return;
    }
548

549 550
    bool is_input_unsigned{false};
    auto input_scale = GetScaleValueForNode(input, &is_input_unsigned);
551 552
    QuantizeInput(
        g, fc, input, "Input", input_scale, is_input_unsigned, "Scale_in");
M
Michał Gallus 已提交
553

554
    auto weight_scale_tensor = GetScaleTensorForNode(weights);
M
Michał Gallus 已提交
555
    EigenVectorArrayMap eigen_tensor{weight_scale_tensor.data<double>(),
556
                                     weight_scale_tensor.numel()};
M
Michał Gallus 已提交
557 558 559 560 561 562 563
    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);

564
    // if quantization scale is missing for output tensor, return fp32 data
565
    if (AreScalesPresentForNodes({output})) {
566 567
      bool is_output_unsigned{false};
      auto output_scale = GetScaleValueForNode(output, &is_output_unsigned);
568 569
      DequantizeOutput(
          g, fc, output, "Out", output_scale, is_output_unsigned, "Scale_out");
570 571 572
    } else {
      fc->Op()->SetAttr("force_fp32_output", true);
    }
M
Michał Gallus 已提交
573 574 575 576 577 578

    ++quantize_fc_count;
  };

  gpd(graph, handler);
  AddStatis(quantize_fc_count);
579
  LogQuantizedOpsCounter("fc", quantize_fc_count);
M
Michał Gallus 已提交
580 581
}

582 583 584 585 586 587 588 589 590 591 592 593 594
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
595
    if (!platform::HasOpINT8DataType(pool_op->Op())) {
596 597 598
      LogQuantizationDisabled(pool_op);
      return;
    }
599 600 601 602

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

603
    if (!AreScalesPresentForNodes({pool_input, pool_output})) {
604 605
      MarkAndLogCannotQuantizeOp(pool_op,
                                 "No scale available for the operator");
606 607
      return;
    }
608

609 610
    bool is_input_unsigned{false};
    auto input_scale = GetScaleValueForNode(pool_input, &is_input_unsigned);
611 612
    QuantizeInput(g, pool_op, pool_input, "X", input_scale, is_input_unsigned);

613 614
    bool is_output_unsigned{false};
    auto output_scale = GetScaleValueForNode(pool_output, &is_output_unsigned);
615 616
    DequantizeOutput(
        g, pool_op, pool_output, "Out", output_scale, is_output_unsigned);
617 618 619 620 621 622

    ++quantize_pool_count;
  };

  gpd(graph, handler);
  AddStatis(quantize_pool_count);
623
  LogQuantizedOpsCounter("pool2d", quantize_pool_count);
624 625
}

626 627 628 629 630 631 632 633 634 635 636 637 638
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
639
    if (!platform::HasOpINT8DataType(concat_op->Op())) {
640 641 642
      LogQuantizationDisabled(concat_op);
      return;
    }
643

644 645 646 647 648 649 650 651 652 653 654 655 656 657
    bool are_all_inputs_unsigned{true};
    // if all inputs were unsigned, then the output was set to unsigned
    // during the scale calculation step
    auto inputs = concat_op->inputs;
    for (size_t i = 0; i < inputs.size(); i++) {
      if (AreScalesPresentForVarNames({inputs[i]->Name()})) {
        auto scale_data = GetScaleDataByName(inputs[i]->Name());
        if (scale_data.first == false) {
          are_all_inputs_unsigned = false;
          break;
        }
      }
    }

658 659
    GET_IR_NODE_FROM_SUBGRAPH(concat_out, concat_out, concat_pattern);

660
    if (!AreScalesPresentForNodes({concat_out})) {
661 662
      MarkAndLogCannotQuantizeOp(concat_op,
                                 "No scale available for the operator");
663 664
      return;
    }
665

666
    auto output_scale = GetScaleValueForNode(concat_out);
667

668
    QuantizeInputs(g, concat_op, "X", are_all_inputs_unsigned);
669

670 671
    DequantizeOutput(
        g, concat_op, concat_out, "Out", output_scale, are_all_inputs_unsigned);
672 673 674 675 676
    ++quantize_concat_count;
  };

  gpd(graph, handler);
  AddStatis(quantize_concat_count);
677
  LogQuantizedOpsCounter("concat", quantize_concat_count);
678 679
}

680 681 682 683 684 685 686 687 688 689 690 691 692
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
693
    if (!platform::HasOpINT8DataType(prior_box_op->Op())) {
694 695 696
      LogQuantizationDisabled(prior_box_op);
      return;
    }
697

698 699
    GET_IR_NODE_FROM_SUBGRAPH(
        prior_box_input, prior_box_input, prior_box_pattern);
700

701
    if (!AreScalesPresentForNodes({prior_box_input})) {
702 703
      MarkAndLogCannotQuantizeOp(prior_box_op,
                                 "No scale available for the operator");
704 705
      return;
    }
706

707 708 709
    bool is_input_unsigned{false};
    auto input_scale =
        GetScaleValueForNode(prior_box_input, &is_input_unsigned);
710 711 712 713 714
    QuantizeInput(g,
                  prior_box_op,
                  prior_box_input,
                  "Input",
                  input_scale,
715 716 717 718 719 720 721
                  is_input_unsigned);

    ++quantize_prior_box_count;
  };

  gpd(graph, handler);
  AddStatis(quantize_prior_box_count);
722
  LogQuantizedOpsCounter("prior_box", quantize_prior_box_count);
723 724
}

725 726 727
void CPUQuantizePass::QuantizeImmutable(Graph* graph,
                                        const std::string& immutable_type,
                                        const std::string& input_name) const {
728 729
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
730 731
  patterns::Immutable immutable_pattern{pattern, name_scope_};
  immutable_pattern(immutable_type, input_name);
732

733
  int quantize_immutable_count = 0;
734 735
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
736 737
    VLOG(4) << "Quantize " + immutable_type + " op";
    GET_IR_NODE_FROM_SUBGRAPH(immutable_op, immutable_op, immutable_pattern);
738 739

    // skip if should not be quantized
740 741
    if (!platform::HasOpINT8DataType(immutable_op->Op())) {
      LogQuantizationDisabled(immutable_op);
742 743
      return;
    }
744 745 746
    GET_IR_NODE_FROM_SUBGRAPH(prev_op, prev_op, immutable_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(immutable_in, immutable_in, immutable_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(immutable_out, immutable_out, immutable_pattern);
747

748
    // skip if prev op and next op is not quantized
749 750
    if (!IsOpDequantized(prev_op) && !IsOpQuantized(immutable_out)) {
      MarkAndLogCannotQuantizeOp(immutable_op,
751
                                 "No other quantizable operators nearby");
752 753 754
      return;
    }

755 756 757 758 759 760 761
    // skip if the dtype of immutable_in is not float32
    auto dtype = immutable_in->Var()->GetDataType();
    if (dtype != proto::VarType::FP32) {
      MarkAndLogCannotQuantizeOp(immutable_op, "The input dtype is not float.");
      return;
    }

762 763
    if (!AreScalesPresentForNodes({immutable_out})) {
      MarkAndLogCannotQuantizeOp(immutable_op,
764
                                 "No scale available for the operator");
765
      return;
766
    }
767

768
    bool is_input_unsigned{false};
769 770 771 772 773 774 775 776
    auto input_scale = GetScaleValueForNode(immutable_out, &is_input_unsigned);

    QuantizeInput(g,
                  immutable_op,
                  immutable_in,
                  input_name,
                  input_scale,
                  is_input_unsigned);
777

778 779
    bool is_output_unsigned{false};
    auto output_scale =
780
        GetScaleValueForNode(immutable_out, &is_output_unsigned);
P
Paulina Gacek 已提交
781 782 783 784 785 786 787 788 789 790 791
    if (immutable_type == "split") {  // ops with multiple outputs
      DequantizeOutputs(
          g, immutable_op, "Out", output_scale, is_output_unsigned);
    } else {
      DequantizeOutput(g,
                       immutable_op,
                       immutable_out,
                       "Out",
                       output_scale,
                       is_output_unsigned);
    }
792
    ++quantize_immutable_count;
793 794 795
  };

  gpd(graph, handler);
796 797
  AddStatis(quantize_immutable_count);
  LogQuantizedOpsCounter(immutable_type, quantize_immutable_count);
Z
Zuza 已提交
798 799
}

800
void CPUQuantizePass::QuantizeMatmul(Graph* graph, bool with_residual) const {
801 802
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
803
  patterns::MatmulWithInputOps matmul_pattern{pattern, name_scope_};
804
  matmul_pattern(with_residual);
805 806 807 808 809 810 811 812

  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
813
    if (!platform::HasOpINT8DataType(matmul_op->Op())) {
814
      LogQuantizationDisabled(matmul_op);
815 816 817 818 819 820
      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
821
    if (!IsOpDequantized(prev_op_x) && !IsOpDequantized(prev_op_y)) {
822 823
      MarkAndLogCannotQuantizeOp(matmul_op,
                                 "No other quantizable operators nearby");
824 825 826 827 828 829
      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);

830 831 832 833 834 835 836 837 838
    auto has_output_scale = AreScalesPresentForNodes({matmul_out});
    if (with_residual && !has_output_scale) {
      MarkAndLogCannotQuantizeOp(
          matmul_op,
          "Matmul op with ResidualData input cannot be quantized "
          "without output scale.");
      return;
    }

839
    if (!AreScalesPresentForNodes({matmul_in_x, matmul_in_y})) {
840 841
      MarkAndLogCannotQuantizeOp(matmul_op,
                                 "No scale available for the operator");
842
      return;
843
    }
844

845 846 847
    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);
848 849
    PADDLE_ENFORCE_EQ(is_x_unsigned,
                      is_y_unsigned,
850 851 852 853
                      platform::errors::InvalidArgument(
                          "Matmul inputs should have the same "
                          "attribute of signed/unsigned, but they "
                          "are different: x(%d), y(%d).",
854 855
                          is_x_unsigned,
                          is_y_unsigned));
856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877

    if (with_residual) {
      GET_IR_NODE_FROM_SUBGRAPH(
          matmul_residual_data, matmul_residual_data, matmul_pattern);
      if (!AreScalesPresentForNodes({matmul_residual_data})) {
        MarkAndLogCannotQuantizeOp(matmul_op,
                                   "No scale available for the operator");
        return;
      }
      bool is_residual_unsigned{false};
      auto residual_scale =
          GetScaleValueForNode(matmul_residual_data, &is_residual_unsigned);

      QuantizeInput(g,
                    matmul_op,
                    matmul_residual_data,
                    "ResidualData",
                    residual_scale,
                    is_residual_unsigned,
                    "Scale_in_eltwise");
    }

878 879 880 881 882 883
    QuantizeInput(g,
                  matmul_op,
                  matmul_in_x,
                  "X",
                  input_x_scale,
                  is_x_unsigned,
884
                  "Scale_x");
885 886 887 888 889 890
    QuantizeInput(g,
                  matmul_op,
                  matmul_in_y,
                  "Y",
                  input_y_scale,
                  is_y_unsigned,
891 892
                  "Scale_y");

893
    // if quantization scale is missing for output tensor, return fp32 data
894
    if (AreScalesPresentForNodes({matmul_out})) {
895 896
      bool is_output_unsigned{false};
      auto output_scale = GetScaleValueForNode(matmul_out, &is_output_unsigned);
897 898 899 900 901 902 903
      DequantizeOutput(g,
                       matmul_op,
                       matmul_out,
                       "Out",
                       output_scale,
                       is_output_unsigned,
                       "Scale_out");
904 905 906
    } else {
      matmul_op->Op()->SetAttr("force_fp32_output", true);
    }
907 908 909 910 911

    ++quantize_matmul_count;
  };
  gpd(graph, handler);
  AddStatis(quantize_matmul_count);
912 913 914
  LogQuantizedOpsCounter("matmul",
                         quantize_matmul_count,
                         (with_residual ? "with residual connection" : ""));
915 916
}

Z
Zuza 已提交
917
void CPUQuantizePass::QuantizeElementwise(
918
    Graph* graph, const std::string& elementwise_type) const {
919 920
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
921
  patterns::ElementwiseOp elementwise_pattern{pattern, name_scope_};
922

923
  elementwise_pattern(elementwise_type);
924

Z
Zuza 已提交
925
  int quantize_elementwise_count = 0;
926 927
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
Z
Zuza 已提交
928
    VLOG(4) << "Quantize " + elementwise_type + " op";
929 930
    GET_IR_NODE_FROM_SUBGRAPH(
        elementwise_op, elementwise_op, elementwise_pattern);
931 932

    // skip if should not be quantized
Z
Zuza 已提交
933 934
    if (!platform::HasOpINT8DataType(elementwise_op->Op())) {
      LogQuantizationDisabled(elementwise_op);
935 936 937
      return;
    }

938 939
    auto x_name = elementwise_op->Op()->Input("X");
    auto y_name = elementwise_op->Op()->Input("Y");
940
    Node *elementwise_x{nullptr}, *elementwise_y{nullptr};
941 942 943 944 945 946 947 948 949

    for (auto& input : elementwise_op->inputs) {
      if (input->Name() == x_name[0]) elementwise_x = input;
      if (input->Name() == y_name[0]) elementwise_y = input;
    }
    if (!elementwise_x || !elementwise_y) {
      return;
    }

950 951
    GET_IR_NODE_FROM_SUBGRAPH(
        elementwise_out, elementwise_out, elementwise_pattern);
952

953
    if (!AreScalesPresentForNodes(
Z
Zuza 已提交
954
            {elementwise_x, elementwise_y, elementwise_out})) {
955 956
      MarkAndLogCannotQuantizeOp(elementwise_op,
                                 "No scale available for the operator");
957 958 959 960
      return;
    }

    bool is_x_unsigned{false}, is_y_unsigned{false};
Z
Zuza 已提交
961 962
    auto input_x_scale = GetScaleValueForNode(elementwise_x, &is_x_unsigned);
    auto input_y_scale = GetScaleValueForNode(elementwise_y, &is_y_unsigned);
963 964 965

    // TODO(sfraczek): add support for different signness
    if (is_x_unsigned != is_y_unsigned) {
966 967
      MarkAndLogCannotQuantizeOp(
          elementwise_op, "Elementwise inputs must be of the same type.");
968 969 970
      return;
    }

971 972 973 974 975 976 977 978 979 980 981 982 983 984
    QuantizeInput(g,
                  elementwise_op,
                  elementwise_x,
                  "X",
                  input_x_scale,
                  is_x_unsigned,
                  "Scale_x");
    QuantizeInput(g,
                  elementwise_op,
                  elementwise_y,
                  "Y",
                  input_y_scale,
                  is_y_unsigned,
                  "Scale_y");
985

986 987
    bool is_output_unsigned{false};
    auto output_scale =
Z
Zuza 已提交
988
        GetScaleValueForNode(elementwise_out, &is_output_unsigned);
989

990 991 992 993 994 995 996
    DequantizeOutput(g,
                     elementwise_op,
                     elementwise_out,
                     "Out",
                     output_scale,
                     is_output_unsigned,
                     "Scale_out");
997

Z
Zuza 已提交
998
    ++quantize_elementwise_count;
999 1000
  };
  gpd(graph, handler);
Z
Zuza 已提交
1001
  AddStatis(quantize_elementwise_count);
1002
  LogQuantizedOpsCounter(elementwise_type, quantize_elementwise_count);
1003 1004
}

1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026
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);

1027
    if (!AreScalesPresentForNodes({x, weight_x})) {
1028
      MarkAndLogCannotQuantizeOp(op, "No scale available for the operator");
1029 1030 1031 1032 1033 1034 1035 1036 1037
      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.;

1038 1039 1040 1041 1042 1043 1044 1045 1046
    QuantizeInput(g,
                  op,
                  x,
                  "X",
                  input_x_scale,
                  is_x_unsigned,
                  "Scale_data",
                  input_x_shift,
                  "Shift_data");
1047 1048 1049

    auto weight_scale_tensor = GetScaleTensorForNode(weight_x);
    EigenVectorArrayMap eigen_tensor{weight_scale_tensor.data<double>(),
1050
                                     weight_scale_tensor.numel()};
1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063
    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);
1064
  LogQuantizedOpsCounter("fusion_gru", quantize_count);
1065 1066
}

1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090
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)) {
1091
      MarkAndLogCannotQuantizeOp(gru, "No scale available for the operator");
1092 1093 1094 1095 1096 1097 1098 1099 1100
      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.;

1101 1102 1103 1104 1105 1106 1107 1108 1109
    QuantizeInput(g,
                  gru,
                  x,
                  "X",
                  input_x_scale,
                  is_x_unsigned,
                  "Scale_data",
                  input_x_shift,
                  "Shift_data");
1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120

    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"));

1121
      scale_var_desc.SetShape(phi::vectorize(scale_tensor_src.dims()));
1122 1123 1124 1125 1126 1127
      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 =
1128
          scope->Var(w_scale_node->Name())->GetMutable<phi::DenseTensor>();
1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147
      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);
1148
  LogQuantizedOpsCounter("multi_gru", quantize_count);
1149 1150
}

1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175
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})) {
1176
      MarkAndLogCannotQuantizeOp(op, "No scale available for the operator");
1177 1178 1179 1180 1181 1182 1183 1184 1185
      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.;

1186 1187 1188 1189 1190 1191 1192 1193 1194
    QuantizeInput(g,
                  op,
                  x,
                  "X",
                  input_x_scale,
                  is_x_unsigned,
                  "Scale_data",
                  input_x_shift,
                  "Shift_data");
1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211

    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);
1212
  LogQuantizedOpsCounter("fusion_lstm", quantize_count);
1213 1214
}

1215
void CPUQuantizePass::ApplyImpl(ir::Graph* graph) const {
1216
  VLOG(3) << "Quantizing the graph.";
1217 1218
  PADDLE_ENFORCE_NOT_NULL(
      graph, platform::errors::InvalidArgument("Graph cannot be nullptr."));
1219
  FusePassBase::Init(name_scope_, graph);
1220

1221 1222 1223
  PADDLE_ENFORCE_NOT_NULL(
      param_scope(),
      platform::errors::InvalidArgument("Scope cannot be nullptr."));
1224

B
baoachun 已提交
1225
  GetQuantInfo(graph);
1226 1227 1228
  QuantizeConv(graph, false /* with_residual_data */);
  QuantizeConv(graph, true /* with_residual_data */);
  QuantizePool(graph);
1229
  QuantizeConcat(graph);
1230
  QuantizePriorBox(graph);
M
Michał Gallus 已提交
1231
  QuantizeFc(graph);
1232 1233
  QuantizeMatmul(graph, false /* with_residual_data */);
  QuantizeMatmul(graph, true /* with_residual_data */);
1234 1235 1236 1237 1238
  QuantizeImmutable(graph, "reshape2", "X");
  QuantizeImmutable(graph, "transpose2", "X");
  QuantizeImmutable(graph, "slice", "Input");
  QuantizeImmutable(graph, "nearest_interp", "X");
  QuantizeImmutable(graph, "nearest_interp_v2", "X");
P
Paulina Gacek 已提交
1239
  QuantizeImmutable(graph, "split", "X");
Z
Zuza 已提交
1240 1241
  QuantizeElementwise(graph, "elementwise_add");
  QuantizeElementwise(graph, "elementwise_mul");
1242
  QuantizeElementwise(graph, "elementwise_sub");
1243
  QuantizeFusionGru(graph);
1244
  QuantizeMultiGru(graph);
1245
  QuantizeFusionLSTM(graph);
1246 1247 1248 1249 1250 1251 1252 1253
}

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

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