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 68 69 70 71 72 73
void LogQuantizedOpsCounter(const std::string& type, const int counter,
                            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());
}

74 75 76 77 78 79
}  // 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,
80 81 82
                                    bool is_input_unsigned,
                                    std::string scale_attr_name, float shift,
                                    std::string shift_attr_name) const {
M
Michał Gallus 已提交
83 84 85
  auto inputs = op->Op()->InputNames();
  bool name_found =
      std::find(inputs.begin(), inputs.end(), input_name) != inputs.end();
86 87 88 89
  PADDLE_ENFORCE_EQ(name_found, true,
                    platform::errors::InvalidArgument(
                        "Var(%s) isn't the input of the %s operator.",
                        input_name, op->Op()->Type()));
90
  unsigned max = is_input_unsigned ? U8_MAX : S8_MAX;
91 92 93 94 95 96 97 98 99 100 101 102 103
  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);
104 105
  q_desc.SetAttr("Shift", shift);
  q_desc.SetAttr("is_negative_input", !is_input_unsigned);
106

Z
Zuza 已提交
107 108 109 110 111 112 113 114 115 116 117
  // 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");
  }
118 119 120 121 122 123 124 125 126 127 128 129 130
  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);
131
  if (!shift_attr_name.empty()) op->Op()->SetAttr(shift_attr_name, shift);
132 133
}

134
void CPUQuantizePass::QuantizeInputs(Graph* g, Node* op, std::string input_name,
135 136 137
                                     bool are_inputs_unsigned,
                                     std::string scale_attr_name, float shift,
                                     std::string shift_attr_name) const {
138
  auto inputs = op->inputs;
139
  auto output = op->outputs[0];
140 141 142 143 144 145 146 147
  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()));
148 149 150 151 152 153 154 155

  // 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());

156
  double scale_out = GetScaleValueForNode(output);
157
  unsigned max = are_inputs_unsigned ? U8_MAX : S8_MAX;
158
  float scale = scale_out * max;
159 160 161 162 163 164 165 166

  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);
167
    q_desc.SetAttr("Shift", shift);
168 169 170
    q_desc.SetInput("Input", std::vector<std::string>({inputs[i]->Name()}));
    q_desc.SetOutput("Output",
                     std::vector<std::string>({quantize_out_node_names[i]}));
171
    q_desc.SetAttr("is_negative_input", !are_inputs_unsigned);
172 173 174 175 176 177 178 179 180 181 182 183 184
    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);
185
  if (!shift_attr_name.empty()) op->Op()->SetAttr(shift_attr_name, shift);
186 187
}

188 189 190 191
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
192 193 194 195 196
  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(
197 198
                        "Var(%s) isn't the output of the %s operator.",
                        output_name, op->Op()->Type()));
199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227
  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);
}

228 229 230
bool CPUQuantizePass::AreScalesPresentForVarNames(
    std::vector<std::string> names) const {
  bool present = true;
B
baoachun 已提交
231 232 233 234 235 236 237 238 239 240 241 242 243 244
  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);
      }
245 246 247 248 249
    }
  }
  return present;
}

250
bool CPUQuantizePass::AreScalesPresentForNodes(
251
    std::initializer_list<Node*> nodes) const {
252
  bool present = true;
B
baoachun 已提交
253 254 255 256 257 258 259 260 261 262 263 264 265 266
  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);
      }
267 268 269 270 271
    }
  }
  return present;
}

272 273
std::pair<bool, LoDTensor> CPUQuantizePass::GetScaleDataByName(
    const std::string& name) const {
B
baoachun 已提交
274 275 276 277 278
  if (var_quant_scales_->empty()) {
    auto& scales = Get<VarQuantScale>("quant_var_scales");
    return scales.at(name);
  }
  return var_quant_scales_->at(name);
279 280
}

281 282
std::pair<bool, LoDTensor> CPUQuantizePass::GetScaleDataForNode(
    const Node* node) const {
283 284 285 286 287
  return GetScaleDataByName(node->Name());
}

LoDTensor CPUQuantizePass::GetScaleTensorByName(const std::string& name) const {
  return GetScaleDataByName(name).second;
288 289 290 291 292 293 294 295 296 297 298 299 300
}

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];
}

301 302
bool CPUQuantizePass::IsOpDequantized(const Node* node) const {
  return node->Op()->Type() == "dequantize" ||
303
         platform::HasOpINT8DataType(node->Op());
304 305 306
}

bool CPUQuantizePass::IsOpQuantized(const Node* node) const {
307 308 309 310 311 312
  // 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())));
  });
313 314
}

B
baoachun 已提交
315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331
void CPUQuantizePass::GetQuantInfo(Graph* graph) const {
  std::unordered_map<std::string, std::vector<float>> info_map{};
  GetInfoFromTheFirstOp(graph, "has_quant_info", "var_quant_scales", &info_map);

  for (auto iter = info_map.begin(); iter != info_map.end(); iter++) {
    LoDTensor tensor;
    const int size = static_cast<int>(iter->second.size());
    auto* data = tensor.mutable_data<double>({size}, platform::CPUPlace());
    for (int i = 0; i < size; i++) {
      data[i] = static_cast<double>(iter->second[i]);
    }

    auto pair = std::make_pair(false, tensor);
    var_quant_scales_->insert(std::make_pair(iter->first, pair));
  }
}

332 333 334 335 336 337 338 339 340 341 342 343 344 345
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
346
    if (!platform::HasOpINT8DataType(conv_op->Op())) {
347 348 349
      LogQuantizationDisabled(conv_op);
      return;
    }
350 351 352 353 354

    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);

355
    auto has_output_scale = AreScalesPresentForNodes({conv_output});
W
Wojciech Uss 已提交
356
    if (with_residual_data && !has_output_scale) {
357 358 359 360
      MarkAndLogCannotQuantizeOp(
          conv_op,
          "Conv op with ResidualData input cannot be quantized "
          "without output scale.");
W
Wojciech Uss 已提交
361 362 363
      return;
    }

364 365 366
    if (with_residual_data) {
      GET_IR_NODE_FROM_SUBGRAPH(conv_residual_data, conv_residual_data,
                                conv_pattern);
367
      if (!AreScalesPresentForNodes(
368
              {conv_input, conv_filter, conv_residual_data})) {
369 370
        MarkAndLogCannotQuantizeOp(conv_op,
                                   "No scale available for the operator");
371
        return;
372
      }
373 374 375 376 377 378 379 380

      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 {
381
      if (!AreScalesPresentForNodes({conv_input, conv_filter})) {
382 383
        MarkAndLogCannotQuantizeOp(conv_op,
                                   "No scale available for the operator");
384
        return;
385
      }
386 387
    }

388 389
    bool is_input_unsigned{false};
    auto input_scale = GetScaleValueForNode(conv_input, &is_input_unsigned);
390 391 392
    QuantizeInput(g, conv_op, conv_input, "Input", input_scale,
                  is_input_unsigned, "Scale_in");

393
    auto filter_scale_tensor = GetScaleTensorForNode(conv_filter);
394
    EigenVectorArrayMap eigen_tensor{filter_scale_tensor.data<double>(),
395
                                     filter_scale_tensor.numel()};
396 397 398 399 400 401 402
    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);

403
    // if quantization scale is missing for output tensor, return fp32 data
W
Wojciech Uss 已提交
404
    if (has_output_scale) {
405 406 407 408 409 410 411 412
      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);
    }
413

414
    // change threshold in bounded ReLu
415 416
    if (conv_op->Op()->GetAttrIfExists<std::string>("fuse_activation") ==
        "relu6") {
417 418 419 420
      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"));
421
      conv_op->Op()->SetAttr("fuse_alpha", scale_out * threshold);
422 423
    }

424 425 426 427 428 429
    ++quantize_conv_count;
  };

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

430 431 432
  LogQuantizedOpsCounter(
      "conv2d", quantize_conv_count,
      ((with_residual_data) ? "with residual connection" : ""));
433 434
}

M
Michał Gallus 已提交
435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451
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
452
    if (!platform::HasOpINT8DataType(fc->Op())) {
453 454 455
      LogQuantizationDisabled(fc);
      return;
    }
456
    if (!fc->Op()->GetAttrIfExists<bool>("use_mkldnn")) {
457
      MarkAndLogCannotQuantizeOp(fc, "use_mkldnn attribute set to false");
M
Michał Gallus 已提交
458
      return;
459
    }
M
Michał Gallus 已提交
460 461 462 463 464

    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);

465
    if (!AreScalesPresentForNodes({input, weights})) {
466
      MarkAndLogCannotQuantizeOp(fc, "No scale available for the operator");
467 468
      return;
    }
469

470 471
    bool is_input_unsigned{false};
    auto input_scale = GetScaleValueForNode(input, &is_input_unsigned);
M
Michał Gallus 已提交
472 473 474
    QuantizeInput(g, fc, input, "Input", input_scale, is_input_unsigned,
                  "Scale_in");

475
    auto weight_scale_tensor = GetScaleTensorForNode(weights);
M
Michał Gallus 已提交
476
    EigenVectorArrayMap eigen_tensor{weight_scale_tensor.data<double>(),
477
                                     weight_scale_tensor.numel()};
M
Michał Gallus 已提交
478 479 480 481 482 483 484
    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);

485
    // if quantization scale is missing for output tensor, return fp32 data
486
    if (AreScalesPresentForNodes({output})) {
487 488 489 490 491 492 493
      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
494 495 496 497 498 499

    ++quantize_fc_count;
  };

  gpd(graph, handler);
  AddStatis(quantize_fc_count);
500
  LogQuantizedOpsCounter("fc", quantize_fc_count);
M
Michał Gallus 已提交
501 502
}

503 504 505 506 507 508 509 510 511 512 513 514 515
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
516
    if (!platform::HasOpINT8DataType(pool_op->Op())) {
517 518 519
      LogQuantizationDisabled(pool_op);
      return;
    }
520 521 522 523

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

524
    if (!AreScalesPresentForNodes({pool_input, pool_output})) {
525 526
      MarkAndLogCannotQuantizeOp(pool_op,
                                 "No scale available for the operator");
527 528
      return;
    }
529

530 531
    bool is_input_unsigned{false};
    auto input_scale = GetScaleValueForNode(pool_input, &is_input_unsigned);
532 533
    QuantizeInput(g, pool_op, pool_input, "X", input_scale, is_input_unsigned);

534 535
    bool is_output_unsigned{false};
    auto output_scale = GetScaleValueForNode(pool_output, &is_output_unsigned);
536 537 538 539 540 541 542 543
    DequantizeOutput(g, pool_op, pool_output, "Out", output_scale,
                     is_output_unsigned);

    ++quantize_pool_count;
  };

  gpd(graph, handler);
  AddStatis(quantize_pool_count);
544
  LogQuantizedOpsCounter("pool2d", quantize_pool_count);
545 546
}

547 548 549 550 551 552 553 554 555 556 557 558 559
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
560
    if (!platform::HasOpINT8DataType(concat_op->Op())) {
561 562 563
      LogQuantizationDisabled(concat_op);
      return;
    }
564 565 566

    GET_IR_NODE_FROM_SUBGRAPH(concat_out, concat_out, concat_pattern);

567
    if (!AreScalesPresentForNodes({concat_out})) {
568 569
      MarkAndLogCannotQuantizeOp(concat_op,
                                 "No scale available for the operator");
570 571
      return;
    }
572

573 574
    // if all inputs were unsigned, then the output was set to unsigned
    // during the scale calculation step
575 576 577
    bool are_all_inputs_unsigned{false};
    auto output_scale =
        GetScaleValueForNode(concat_out, &are_all_inputs_unsigned);
578

579
    QuantizeInputs(g, concat_op, "X", are_all_inputs_unsigned);
580 581 582 583 584 585 586 587 588

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

    ++quantize_concat_count;
  };

  gpd(graph, handler);
  AddStatis(quantize_concat_count);
589
  LogQuantizedOpsCounter("concat", quantize_concat_count);
590 591
}

592 593 594 595 596 597 598 599 600 601 602 603 604
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
605
    if (!platform::HasOpINT8DataType(prior_box_op->Op())) {
606 607 608
      LogQuantizationDisabled(prior_box_op);
      return;
    }
609 610 611 612

    GET_IR_NODE_FROM_SUBGRAPH(prior_box_input, prior_box_input,
                              prior_box_pattern);

613
    if (!AreScalesPresentForNodes({prior_box_input})) {
614 615
      MarkAndLogCannotQuantizeOp(prior_box_op,
                                 "No scale available for the operator");
616 617
      return;
    }
618

619 620 621
    bool is_input_unsigned{false};
    auto input_scale =
        GetScaleValueForNode(prior_box_input, &is_input_unsigned);
622 623 624 625 626 627 628 629
    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);
630
  LogQuantizedOpsCounter("prior_box", quantize_prior_box_count);
631 632
}

633 634 635 636 637 638 639 640 641 642 643 644 645
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
646
    if (!platform::HasOpINT8DataType(transpose_op->Op())) {
647
      LogQuantizationDisabled(transpose_op);
648 649 650
      return;
    }
    GET_IR_NODE_FROM_SUBGRAPH(prev_op, prev_op, transpose_pattern);
651 652
    GET_IR_NODE_FROM_SUBGRAPH(transpose_in, transpose_in, transpose_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(transpose_out, transpose_out, transpose_pattern);
653

654
    // skip if prev op and next op is not quantized
655
    if (!(IsOpDequantized(prev_op)) && !(IsOpQuantized(transpose_out))) {
656 657
      MarkAndLogCannotQuantizeOp(transpose_op,
                                 "No other quantizable operators nearby");
658 659 660
      return;
    }

661
    if (!AreScalesPresentForNodes({transpose_in, transpose_out})) {
662 663
      MarkAndLogCannotQuantizeOp(transpose_op,
                                 "No scale available for the operator");
664
      return;
665
    }
666

667 668
    bool is_input_unsigned{false};
    auto input_scale = GetScaleValueForNode(transpose_in, &is_input_unsigned);
669 670 671
    QuantizeInput(g, transpose_op, transpose_in, "X", input_scale,
                  is_input_unsigned);

672 673 674
    bool is_output_unsigned{false};
    auto output_scale =
        GetScaleValueForNode(transpose_out, &is_output_unsigned);
675 676 677 678 679 680 681 682
    DequantizeOutput(g, transpose_op, transpose_out, "Out", output_scale,
                     is_output_unsigned);

    ++quantize_transpose_count;
  };

  gpd(graph, handler);
  AddStatis(quantize_transpose_count);
683
  LogQuantizedOpsCounter("transpose2", quantize_transpose_count);
684 685
}

686 687 688 689 690 691 692 693 694 695 696 697 698
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
699
    if (!platform::HasOpINT8DataType(reshape_op->Op())) {
700
      LogQuantizationDisabled(reshape_op);
701 702 703
      return;
    }
    GET_IR_NODE_FROM_SUBGRAPH(prev_op, prev_op, reshape_pattern);
704 705
    GET_IR_NODE_FROM_SUBGRAPH(reshape_in, reshape_in, reshape_pattern);
    GET_IR_NODE_FROM_SUBGRAPH(reshape_out, reshape_out, reshape_pattern);
706

707 708
    // skip if prev op is not quantized
    if (!(IsOpDequantized(prev_op)) && !(IsOpQuantized(reshape_out))) {
709 710
      MarkAndLogCannotQuantizeOp(reshape_op,
                                 "No other quantizable operators nearby");
711 712 713
      return;
    }

714
    if (!AreScalesPresentForNodes({reshape_in, reshape_out})) {
715 716
      MarkAndLogCannotQuantizeOp(reshape_op,
                                 "No scale available for the operator");
717
      return;
718
    }
719

720 721
    bool is_input_unsigned{false};
    auto input_scale = GetScaleValueForNode(reshape_in, &is_input_unsigned);
722 723 724
    QuantizeInput(g, reshape_op, reshape_in, "X", input_scale,
                  is_input_unsigned);

725 726
    bool is_output_unsigned{false};
    auto output_scale = GetScaleValueForNode(reshape_out, &is_output_unsigned);
727 728 729 730 731 732 733 734
    DequantizeOutput(g, reshape_op, reshape_out, "Out", output_scale,
                     is_output_unsigned);

    ++quantize_reshape_count;
  };

  gpd(graph, handler);
  AddStatis(quantize_reshape_count);
735
  LogQuantizedOpsCounter("reshape2", quantize_reshape_count);
736 737
}

Z
Zuza 已提交
738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755
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);
756 757
    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
758 759

    // skip if prev op and next op is not quantized
760
    if (!IsOpDequantized(prev_op) && !IsOpQuantized(slice_out)) {
761 762
      MarkAndLogCannotQuantizeOp(slice_op,
                                 "No other quantizable operators nearby");
Z
Zuza 已提交
763 764 765 766
      return;
    }

    if (!AreScalesPresentForNodes({slice_out})) {
767 768
      MarkAndLogCannotQuantizeOp(slice_op,
                                 "No scale available for the operator");
Z
Zuza 已提交
769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786
      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);
787
  LogQuantizedOpsCounter("slice", quantize_slice_count);
Z
Zuza 已提交
788 789
}

790 791 792
void CPUQuantizePass::QuantizeMatmul(Graph* graph) const {
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
793
  patterns::MatmulWithInputOps matmul_pattern{pattern, name_scope_};
794 795 796 797 798 799 800 801 802
  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
803
    if (!platform::HasOpINT8DataType(matmul_op->Op())) {
804
      LogQuantizationDisabled(matmul_op);
805 806 807 808 809 810 811
      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)) {
812 813
      MarkAndLogCannotQuantizeOp(matmul_op,
                                 "No other quantizable operators nearby");
814 815 816 817 818 819
      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);

820
    if (!AreScalesPresentForNodes({matmul_in_x, matmul_in_y})) {
821 822
      MarkAndLogCannotQuantizeOp(matmul_op,
                                 "No scale available for the operator");
823
      return;
824
    }
825

826 827 828
    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);
829 830 831 832 833 834
    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));
835 836 837 838 839
    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");

840
    // if quantization scale is missing for output tensor, return fp32 data
841
    if (AreScalesPresentForNodes({matmul_out})) {
842 843 844 845 846 847 848
      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);
    }
849 850 851 852 853

    ++quantize_matmul_count;
  };
  gpd(graph, handler);
  AddStatis(quantize_matmul_count);
854
  LogQuantizedOpsCounter("matmul", quantize_matmul_count);
855 856
}

Z
Zuza 已提交
857 858
void CPUQuantizePass::QuantizeElementwise(
    Graph* graph, const std::string elementwise_type) const {
859 860
  GraphPatternDetector gpd;
  auto pattern = gpd.mutable_pattern();
861
  patterns::ElementwiseOp elementwise_pattern{pattern, name_scope_};
862

863
  elementwise_pattern(elementwise_type);
864

Z
Zuza 已提交
865
  int quantize_elementwise_count = 0;
866 867
  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                     Graph* g) {
Z
Zuza 已提交
868 869 870
    VLOG(4) << "Quantize " + elementwise_type + " op";
    GET_IR_NODE_FROM_SUBGRAPH(elementwise_op, elementwise_op,
                              elementwise_pattern);
871 872

    // skip if should not be quantized
Z
Zuza 已提交
873 874
    if (!platform::HasOpINT8DataType(elementwise_op->Op())) {
      LogQuantizationDisabled(elementwise_op);
875 876 877
      return;
    }

878 879 880 881 882 883 884 885 886 887 888 889
    auto x_name = elementwise_op->Op()->Input("X");
    auto y_name = elementwise_op->Op()->Input("Y");
    Node *elementwise_x, *elementwise_y;

    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;
    }

Z
Zuza 已提交
890 891
    GET_IR_NODE_FROM_SUBGRAPH(elementwise_out, elementwise_out,
                              elementwise_pattern);
892

893
    if (!AreScalesPresentForNodes(
Z
Zuza 已提交
894
            {elementwise_x, elementwise_y, elementwise_out})) {
895 896
      MarkAndLogCannotQuantizeOp(elementwise_op,
                                 "No scale available for the operator");
897 898 899 900
      return;
    }

    bool is_x_unsigned{false}, is_y_unsigned{false};
Z
Zuza 已提交
901 902
    auto input_x_scale = GetScaleValueForNode(elementwise_x, &is_x_unsigned);
    auto input_y_scale = GetScaleValueForNode(elementwise_y, &is_y_unsigned);
903 904 905

    // TODO(sfraczek): add support for different signness
    if (is_x_unsigned != is_y_unsigned) {
906 907
      MarkAndLogCannotQuantizeOp(
          elementwise_op, "Elementwise inputs must be of the same type.");
908 909 910
      return;
    }

Z
Zuza 已提交
911
    QuantizeInput(g, elementwise_op, elementwise_x, "X", input_x_scale,
912
                  is_x_unsigned, "Scale_x");
Z
Zuza 已提交
913
    QuantizeInput(g, elementwise_op, elementwise_y, "Y", input_y_scale,
914 915
                  is_y_unsigned, "Scale_y");

916 917
    bool is_output_unsigned{false};
    auto output_scale =
Z
Zuza 已提交
918
        GetScaleValueForNode(elementwise_out, &is_output_unsigned);
919

Z
Zuza 已提交
920 921
    DequantizeOutput(g, elementwise_op, elementwise_out, "Out", output_scale,
                     is_output_unsigned, "Scale_out");
922

Z
Zuza 已提交
923
    ++quantize_elementwise_count;
924 925
  };
  gpd(graph, handler);
Z
Zuza 已提交
926
  AddStatis(quantize_elementwise_count);
927
  LogQuantizedOpsCounter(elementwise_type, quantize_elementwise_count);
928 929
}

930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951
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);

952
    if (!AreScalesPresentForNodes({x, weight_x})) {
953
      MarkAndLogCannotQuantizeOp(op, "No scale available for the operator");
954 955 956 957 958 959 960 961 962 963 964 965 966 967
      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>(),
968
                                     weight_scale_tensor.numel()};
969 970 971 972 973 974 975 976 977 978 979 980 981
    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);
982
  LogQuantizedOpsCounter("fusion_gru", quantize_count);
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
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)) {
1009
      MarkAndLogCannotQuantizeOp(gru, "No scale available for the operator");
1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031
      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"));

1032
      scale_var_desc.SetShape(phi::vectorize(scale_tensor_src.dims()));
1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058
      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);
1059
  LogQuantizedOpsCounter("multi_gru", quantize_count);
1060 1061
}

1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086
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})) {
1087
      MarkAndLogCannotQuantizeOp(op, "No scale available for the operator");
1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115
      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);
1116
  LogQuantizedOpsCounter("fusion_lstm", quantize_count);
1117 1118
}

1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137
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);
1138 1139 1140 1141
    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);
1142 1143

    // skip if prev op and next op is not quantized
1144
    if (!(IsOpDequantized(prev_op)) && !(IsOpQuantized(nearest_interp_out))) {
1145 1146
      MarkAndLogCannotQuantizeOp(nearest_interp_op,
                                 "No other quantizable operators nearby");
1147 1148 1149 1150
      return;
    }

    if (!AreScalesPresentForNodes({nearest_interp_in, nearest_interp_out})) {
1151 1152
      MarkAndLogCannotQuantizeOp(nearest_interp_op,
                                 "No scale available for the operator");
1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172
      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);
1173
  LogQuantizedOpsCounter("nearest_interp", quantize_nearest_interp_count);
1174 1175
}

1176
void CPUQuantizePass::ApplyImpl(ir::Graph* graph) const {
1177
  VLOG(3) << "Quantizing the graph.";
1178 1179
  PADDLE_ENFORCE_NOT_NULL(
      graph, platform::errors::InvalidArgument("Graph cannot be nullptr."));
1180
  FusePassBase::Init(name_scope_, graph);
1181

1182 1183
  PADDLE_ENFORCE_NOT_NULL(param_scope(), platform::errors::InvalidArgument(
                                             "Scope cannot be nullptr."));
1184

B
baoachun 已提交
1185
  GetQuantInfo(graph);
1186 1187 1188
  QuantizeConv(graph, false /* with_residual_data */);
  QuantizeConv(graph, true /* with_residual_data */);
  QuantizePool(graph);
1189
  QuantizeConcat(graph);
1190
  QuantizePriorBox(graph);
1191
  QuantizeTranspose(graph);
M
Michał Gallus 已提交
1192
  QuantizeFc(graph);
1193
  QuantizeReshape(graph);
1194
  QuantizeMatmul(graph);
Z
Zuza 已提交
1195 1196
  QuantizeElementwise(graph, "elementwise_add");
  QuantizeElementwise(graph, "elementwise_mul");
1197
  QuantizeElementwise(graph, "elementwise_sub");
1198
  QuantizeFusionGru(graph);
1199
  QuantizeMultiGru(graph);
1200
  QuantizeFusionLSTM(graph);
Z
Zuza 已提交
1201
  QuantizeSlice(graph);
1202
  QuantizeNearestInterp(graph);
1203 1204 1205 1206 1207 1208 1209 1210
}

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

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