mkldnn_quantizer.cc 18.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
// 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.

#include "paddle/fluid/inference/api/mkldnn_quantizer.h"
#include <algorithm>
17
#include <limits>
18 19 20 21 22 23 24 25 26 27 28 29
#include <map>
#include <numeric>
#include <unordered_map>
#include <utility>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/type_defs.h"
#include "paddle/fluid/inference/analysis/analyzer.h"
#include "paddle/fluid/inference/api/analysis_predictor.h"
30
#include "paddle/fluid/platform/mkldnn_helper.h"
31 32 33 34 35 36 37 38 39 40
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/string/pretty_log.h"

namespace paddle {

using platform::CPUPlace;
using framework::LoDTensor;
using framework::ir::Graph;
using ConstEigenVectorArrayMap =
    Eigen::Map<const Eigen::Array<float, Eigen::Dynamic, 1>>;
M
Michał Gallus 已提交
41 42 43 44 45
using EigenMatrixDoubleArray =
    Eigen::Array<double, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>;
using EigenMatrixArray =
    Eigen::Array<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>;
using ConstEigenMatrixArrayMap = Eigen::Map<const EigenMatrixArray>;
46
using string::PrettyLogH1;
47
static LoDTensor CreateScaleTensor(int64_t channels_num = 1);
48 49 50 51 52 53

bool AnalysisPredictor::MkldnnQuantizer::CalculateScales() {
  PrettyLogH1("--- Calculating scales for quantization");
  using VariableNameMap = std::map<std::string, std::vector<std::string>>;
  std::map<std::string, std::map<std::string, LoDTensor>> gathered_data;
  for (const auto* op : predictor_.inference_program_->Block(0).AllOps()) {
54
    if (platform::HasOpINT8DataType(op)) {
55 56 57 58 59
      const VariableNameMap& connections_in = op->Inputs();
      const VariableNameMap& connections_out = op->Outputs();

      auto glambda = [&](const VariableNameMap& connections, bool is_output) {
        for (auto const& conn : connections) {
60 61
          for (const auto& var_name : conn.second) {
            // skip if scale already computed
62
            if (scales_.find(var_name) != scales_.end()) continue;
63

64 65 66 67 68
            auto* var = predictor_.sub_scope_->FindVar(var_name);
            PADDLE_ENFORCE(var, "%s is not in the scope", var_name);
            PADDLE_ENFORCE(var->IsType<LoDTensor>(),
                           "Only support lod tensor now.");
            LoDTensor* var_tensor = var->GetMutable<LoDTensor>();
69

70 71
            // force unsigned type if already know it
            bool is_unsigned = false;
72 73
            bool compute_scale = true;
            if (is_output) {
M
Michał Gallus 已提交
74
              if (op->Type() == "conv2d" || op->Type() == "fc") {
75
                // output of conv2d with relu must be unsigned
76
                std::string fuse_activation =
77
                    op->GetAttrIfExists<std::string>("fuse_activation");
78 79
                is_unsigned =
                    (fuse_activation == "relu" || fuse_activation == "relu6");
80 81 82 83 84
              } else if (op->Type() == "relu") {
                is_unsigned = true;
              } else if (op->Type() == "transpose2" ||
                         op->Type() == "reshape2" || op->Type() == "pool2d") {
                auto input_var_name = op->Input("X")[0];
85 86 87
                PADDLE_ENFORCE(scales_.find(input_var_name) != scales_.end(),
                               "Input scales must be calculated before the "
                               "output scales to infer if output is unsigned.");
88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109
                if (scales_.find(input_var_name) != scales_.end()) {
                  scales_[var_name] = scales_[input_var_name];
                }
                compute_scale = false;
              } else if (op->Type() == "concat") {
                // output of ops with unsigned input must be unsigned
                is_unsigned = true;
                double min_scale = std::numeric_limits<double>::max();
                for (auto input_var_name : op->Input("X")) {
                  PADDLE_ENFORCE(
                      scales_.find(input_var_name) != scales_.end(),
                      "Input scales must be calculated before the "
                      "output scales to infer if output is unsigned.");
                  is_unsigned = is_unsigned && scales_[input_var_name].first;
                  min_scale = std::min(
                      min_scale,
                      scales_[input_var_name].second.data<double>()[0]);
                }
                auto scale_tensor = CreateScaleTensor();
                scale_tensor.data<double>()[0] = min_scale;
                scales_[var_name] = {is_unsigned, scale_tensor};
                compute_scale = false;
110
              }
111
            }
112 113 114
            if (compute_scale)
              CalculateSingleScale(op->Type(), conn.first, var_name,
                                   *var_tensor, is_unsigned);
115
          }
116 117 118
        }
      };

119
      // handle inputs first to let is_unsigned be inferred for the outputs
120
      glambda(connections_in, false /* is_output */);
121
      glambda(connections_out, true /* is_output */);
122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145
    }
  }

  return true;
}

void AnalysisPredictor::MkldnnQuantizer::CalculateSingleScale(
    const std::string& op_type_name, const std::string& conn_name,
    const std::string& var_name, const LoDTensor& var_tensor,
    bool is_unsigned) {
  auto rule = qconfig_->scale_algo(op_type_name, conn_name);
  if (rule == ScaleAlgo::NONE) return;

  PADDLE_ENFORCE(
      var_tensor.numel() > 0,
      "MkldnnQuantizer: LoDTensor of variable %s for quantization of op "
      "%s of connection %s should not be empty.",
      var_name, op_type_name, conn_name);

  switch (rule) {
    case ScaleAlgo::MAX:
      scales_[var_name] = GetMaxScalingFactor(var_tensor, is_unsigned);
      break;
    case ScaleAlgo::MAX_CH:
M
Michał Gallus 已提交
146 147 148 149 150 151
      scales_[var_name] = GetMaxChScalingFactor(var_tensor, is_unsigned,
                                                /*is_transposed*/ false);
      break;
    case ScaleAlgo::MAX_CH_T:
      scales_[var_name] = GetMaxChScalingFactor(var_tensor, is_unsigned,
                                                /*is_transposed*/ true);
152 153 154 155 156 157 158 159 160 161
      break;
    case ScaleAlgo::KL:
      scales_[var_name] = GetKLScalingFactor(var_tensor, is_unsigned);
      break;
    default:
      throw std::runtime_error(
          "MkldnnQuantizer: Unexpected ScaleAlgo specified.");
  }
}

162 163 164 165 166 167 168
static LoDTensor CreateScaleTensor(int64_t channels_num) {
  LoDTensor scale_tensor;
  scale_tensor.Resize({channels_num});
  scale_tensor.mutable_data<double>(CPUPlace());
  return scale_tensor;
}

169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304
std::vector<int> AnalysisPredictor::MkldnnQuantizer::ExpandQuantizedBins(
    std::vector<int> quantized_bins, std::vector<int> reference_bins) const {
  std::vector<int> expanded_quantized_bins(reference_bins.size(), 0);
  int num_merged_bins = reference_bins.size() / quantized_bins.size();
  int j_start = 0;
  int j_end = num_merged_bins;
  for (size_t idx = 0; idx < quantized_bins.size(); idx++) {
    int zero_count =
        std::count(&reference_bins[j_start], &reference_bins[j_end], 0);
    num_merged_bins = j_end - j_start;
    int avg_bin_ele;
    if (zero_count == num_merged_bins) {
      avg_bin_ele = 0;
    } else {
      avg_bin_ele = quantized_bins[idx] / (num_merged_bins - zero_count + 0.0);
    }
    for (int idx1 = j_start; idx1 < j_end; idx1++) {
      expanded_quantized_bins[idx1] =
          (reference_bins[idx1] == 0) ? 0 : avg_bin_ele;
    }
    j_start += num_merged_bins;
    j_end += num_merged_bins;
    if ((idx + 1) == quantized_bins.size() - 1) {
      j_end = reference_bins.size();
    }
  }
  return expanded_quantized_bins;
}

std::pair<bool, LoDTensor>
AnalysisPredictor::MkldnnQuantizer::GetKLScalingFactor(
    const LoDTensor& var_tensor, bool is_unsigned) const {
  ConstEigenVectorArrayMap eigen_tensor{var_tensor.data<float>(),
                                        var_tensor.numel(), 1};
  int precision_hist_num_bins = 2048;
  float max_val = eigen_tensor.maxCoeff();
  float min_val = eigen_tensor.minCoeff();
  bool is_positive = min_val >= 0.0f;
  if (is_unsigned)
    PADDLE_ENFORCE(
        is_positive,
        "Tensor is claimed to be unsigned, but its min value (%f) is < 0.0",
        min_val);

  int num_quantized_bins = 255;

  std::vector<int> hist;
  float bin_width;
  int starting_iter;
  int ending_iter = precision_hist_num_bins - 1;
  if (is_positive) {
    std::tie(hist, bin_width) =
        Histogram(var_tensor, min_val, max_val, precision_hist_num_bins);
    starting_iter = static_cast<int>(ending_iter * 0.7);
  } else {
    float th = std::max(std::abs(max_val), std::abs(min_val));
    std::tie(hist, bin_width) =
        Histogram(var_tensor, -th, th, precision_hist_num_bins);
    starting_iter = 0;
    if (std::abs(max_val) > std::abs(min_val)) {
      while (starting_iter < ending_iter) {
        if (hist[starting_iter] == 0) {
          ++starting_iter;
          continue;
        } else {
          break;
        }
      }
      starting_iter += static_cast<int>((ending_iter - starting_iter) * 0.6);
    } else {
      while (ending_iter > 0) {
        if (hist[ending_iter] == 0) {
          --ending_iter;
          continue;
        } else {
          break;
        }
      }
      starting_iter = static_cast<int>(0.6 * ending_iter);
    }
  }
  auto P_sum = eigen_tensor.size();
  int min_kl_divergence = 0;
  int min_kl_index = 0;
  bool kl_inited = false;
  for (int i = starting_iter; i <= ending_iter; i++) {
    std::vector<int> reference_distr_P(&hist[0], &hist[i]);
    auto outliers_count =
        std::accumulate(&hist[i], &hist[precision_hist_num_bins], 0);
    if (reference_distr_P[i - 1] == 0) {
      continue;
    }
    reference_distr_P[i - 1] += outliers_count;
    auto reference_distr_bins = reference_distr_P;
    std::vector<int> candidate_distr_Q(&hist[0], &hist[i]);
    int num_merged_bins = i / num_quantized_bins;
    std::vector<int> candidate_distr_Q_quantized(num_quantized_bins, 0);
    int j_start = 0;
    int j_end = num_merged_bins;
    for (int idx = 0; idx < num_quantized_bins; idx++) {
      candidate_distr_Q_quantized[idx] = std::accumulate(
          &candidate_distr_Q[j_start], &candidate_distr_Q[j_end], 0);
      j_start += num_merged_bins;
      j_end += num_merged_bins;
      if ((idx + 1) == num_quantized_bins - 1) {
        j_end = i;
      }
    }
    candidate_distr_Q =
        ExpandQuantizedBins(candidate_distr_Q_quantized, reference_distr_bins);
    int Q_sum =
        std::accumulate(candidate_distr_Q.begin(), candidate_distr_Q.end(), 0);
    auto kl_divergence =
        SafeEntropy(reference_distr_P, P_sum, candidate_distr_Q, Q_sum);
    if (!kl_inited) {
      min_kl_divergence = kl_divergence;
      min_kl_index = i;
      kl_inited = true;
    } else if (kl_divergence < min_kl_divergence) {
      min_kl_divergence = kl_divergence;
      min_kl_index = i;
    } else {
    }
  }
  if (min_kl_index == 0) {
    while (starting_iter > 0) {
      if (hist[starting_iter] == 0) {
        starting_iter -= 1;
        continue;
      } else {
        break;
      }
    }
    min_kl_index = starting_iter;
  }

305 306
  LoDTensor scale_tensor = CreateScaleTensor();
  scale_tensor.data<double>()[0] = 1.0 / ((min_kl_index + 0.5) * bin_width);
307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323

  return std::make_pair(is_unsigned, scale_tensor);
}

std::pair<bool, LoDTensor>
AnalysisPredictor::MkldnnQuantizer::GetMaxScalingFactor(
    const LoDTensor& var_tensor, bool is_unsigned) const {
  ConstEigenVectorArrayMap eigen_tensor{var_tensor.data<float>(),
                                        var_tensor.numel(), 1};
  float max_abs = eigen_tensor.abs().maxCoeff();
  float min_val = eigen_tensor.minCoeff();
  if (is_unsigned)
    PADDLE_ENFORCE(
        min_val >= 0.0f,
        "Tensor is claimed to be unsigned, but its min value (%f) is < 0.0",
        min_val);

324 325
  LoDTensor scale_tensor = CreateScaleTensor();
  scale_tensor.data<double>()[0] = 1.0 / max_abs;
326 327 328 329 330 331

  return std::make_pair(is_unsigned, scale_tensor);
}

std::pair<bool, LoDTensor>
AnalysisPredictor::MkldnnQuantizer::GetMaxChScalingFactor(
M
Michał Gallus 已提交
332
    const LoDTensor& var_tensor, bool is_unsigned, bool is_transposed) const {
333 334 335 336 337 338 339 340 341 342 343
  PADDLE_ENFORCE(var_tensor.dims().size() > 0, "Tensor dimension is empty.");

  ConstEigenVectorArrayMap eigen_tensor{var_tensor.data<float>(),
                                        var_tensor.numel(), 1};
  float min_val = eigen_tensor.minCoeff();
  if (is_unsigned)
    PADDLE_ENFORCE(
        min_val >= 0.0f,
        "Tensor is claimed to be unsigned, but its min value (%f) is < 0.0",
        min_val);

M
Michał Gallus 已提交
344 345 346 347 348
  auto dims = var_tensor.dims();
  constexpr int num_col_dims = 1;
  auto flattened_dims = framework::flatten_to_2d(dims, num_col_dims);
  ConstEigenMatrixArrayMap eigen_tensor_mat{
      var_tensor.data<float>(), flattened_dims[0], flattened_dims[1]};
349

M
Michał Gallus 已提交
350 351 352 353 354
  EigenMatrixDoubleArray scales;
  if (is_transposed) {
    scales = 1.0 / eigen_tensor_mat.cast<double>().abs().colwise().maxCoeff();
  } else {
    scales = 1.0 / eigen_tensor_mat.cast<double>().abs().rowwise().maxCoeff();
355
  }
M
Michał Gallus 已提交
356 357 358 359 360
  int output_channel_axis = is_transposed;
  int channels = dims[output_channel_axis];
  LoDTensor scale_tensor = CreateScaleTensor(channels);
  auto* scale_ptr = scale_tensor.mutable_data<double>(CPUPlace());
  std::copy(scales.data(), scales.data() + scales.size(), scale_ptr);
361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395

  return std::make_pair(is_unsigned, scale_tensor);
}

std::pair<std::vector<int>, float>
AnalysisPredictor::MkldnnQuantizer::Histogram(
    const framework::LoDTensor& var_tensor, float min_val, float max_val,
    size_t num_bins) const {
  PADDLE_ENFORCE_GT(num_bins, 0,
                    "MkldnnQuantizer: To calculate Histogram, num_bins (" +
                        std::to_string(num_bins) + ") must be positive.");
  PADDLE_ENFORCE_GT(
      var_tensor.numel(), 0,
      "MkldnnQuantizer: To calculate Histogram, the tensor must not be empty.");
  PADDLE_ENFORCE(max_val >= min_val,
                 "MkldnnQuantizer: To calculate Histogram, max_val (" +
                     std::to_string(max_val) +
                     ") must be greater or equal"
                     "to min_val (" +
                     std::to_string(min_val) + ").");
  ConstEigenVectorArrayMap eigen_tensor{var_tensor.data<float>(),
                                        var_tensor.numel(), 1};
  auto bin_width = std::abs(max_val - min_val) / num_bins;
  std::vector<int> hist(num_bins);

  for (int i = 0; i < eigen_tensor.size(); i++) {
    int bin = std::min(
        num_bins - 1,
        static_cast<size_t>(floor((eigen_tensor[i] - min_val) / bin_width)));
    ++hist[bin];
  }

  return std::make_pair(std::move(hist), std::move(bin_width));
}

396 397 398 399 400 401 402
void AnalysisPredictor::MkldnnQuantizer::ClearDeviceContext() const {
  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
  platform::MKLDNNDeviceContext* dev_ctx =
      (platform::MKLDNNDeviceContext*)pool.Get(predictor_.place_);
  dev_ctx->ResetBlobMap();
}

403 404 405 406 407 408
void AnalysisPredictor::MkldnnQuantizer::PrepareArgument() const {
  auto& arg = predictor_.argument_;
  if (!arg.scope_valid()) arg.SetScope(new framework::Scope);
  arg.SetMainProgramNotOwned(predictor_.inference_program_.get());
  auto graph = std::unique_ptr<Graph>(new Graph(arg.main_program()));
  arg.SetMainGraph(graph.release());
409 410 411
  auto* scope_ptr = arg.scope_ptr();
  PADDLE_ENFORCE(scope_ptr);
  arg.main_graph().SetNotOwned(framework::ir::kParamScopeAttr, scope_ptr);
412 413 414

  auto* builder = predictor_.config_.pass_builder();
  builder->SetPasses({
415
      "cpu_quantize_pass", "cpu_quantize_squash_pass",
416 417 418 419 420
  });
  if (predictor_.config_.ir_debug_) builder->TurnOnDebug();
  auto passes = builder->AllPasses();
  predictor_.argument_.SetIrAnalysisPasses(passes);
  predictor_.argument_.SetAnalysisPasses(
421 422
      {"ir_graph_clean_pass", "ir_analysis_pass", "memory_optimize_pass",
       "ir_graph_to_program_pass"});
423 424 425 426 427 428
  predictor_.argument_.SetQuantVarScales(scales_);
}

bool AnalysisPredictor::MkldnnQuantizer::Quantize() {
  if (!RunWarmup()) return false;
  if (!CalculateScales()) return false;
429
  ClearDeviceContext();
430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493
  predictor_.PrepareScope(predictor_.scope_);
  predictor_.CreateExecutor();
  if (!RunQuantizePasses()) return false;
  predictor_.PrepareExecutor();
  predictor_.PrepareFeedFetch();
  return true;
}

bool AnalysisPredictor::MkldnnQuantizer::RunQuantizePasses() const {
  predictor_.executor_->CreateVariables(*predictor_.inference_program_, 0, true,
                                        predictor_.sub_scope_);
  PrepareArgument();
  auto& arg = predictor_.argument_;
  Analyzer().Run(&arg);
  PADDLE_ENFORCE(arg.scope_valid());
  VLOG(5) << "to prepare executor";
  ARGUMENT_CHECK_FIELD((&arg), ir_analyzed_program);
  predictor_.inference_program_.reset(
      new framework::ProgramDesc(arg.ir_analyzed_program()));
  LOG(INFO) << "== optimize 2 end ==";
  predictor_.executor_->CreateVariables(*predictor_.inference_program_, 0,
                                        false, predictor_.sub_scope_);
  return true;
}

bool AnalysisPredictor::MkldnnQuantizer::RunWarmup() const {
  VLOG(3) << "Predictor: run a quantization warmup iteration";
  auto warmup_data = qconfig_->warmup_data();
  PADDLE_ENFORCE_NOT_NULL(warmup_data,
                          "Warmup data cannot be NULL in the config.");
  PrettyLogH1("--- Running warmup iteration for quantization");

  // Run the inference program
  std::vector<PaddleTensor> output_slots;
  predictor_.Run(*warmup_data, &output_slots, qconfig_->warmup_batch_size());

  return true;
}

float AnalysisPredictor::MkldnnQuantizer::SafeEntropy(
    std::vector<int> reference_distr_P, int P_sum,
    std::vector<int> candidate_distr_Q, int Q_sum) const {
  PADDLE_ENFORCE_EQ(reference_distr_P.size(), candidate_distr_Q.size());
  float tmp_sum1 = 0;
  float tmp_sum2 = 0;
  for (size_t idx = 0; idx < reference_distr_P.size(); idx++) {
    int p_idx = reference_distr_P[idx];
    int q_idx = candidate_distr_Q[idx];
    if (p_idx == 0) {
      tmp_sum1 += 0;
      tmp_sum2 += 0;
    } else {
      PADDLE_ENFORCE(q_idx != 0, "MkldnnQuantizer: Fatal error!, idx = " +
                                     std::to_string(idx) +
                                     " qindex = 0! p_idx = " +
                                     std::to_string(p_idx));
    }
    tmp_sum1 += p_idx * (log(Q_sum * p_idx));
    tmp_sum2 += p_idx * (log(P_sum * q_idx));
  }
  return (tmp_sum1 - tmp_sum2) / P_sum;
}

}  // namespace paddle