Add fusion_gru and multi_gru to PTQ (Post-Training Quantization) (#33749)

* Add calculation for gru op

* Correct the types

* Remove mkldnn only

* Correct mkldnn ifdef

* Remove mkldnn ifdef

* Separate mkldnn quantizer test

* Correct Windows test

* Check different cmake fix

* Revert cmake change

* Cmake change 2

* Cmake change 3

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -2249,9 +2249,9 @@ PDNode *patterns::MultipleQuantize::operator()() {
 PDNode *patterns::QuantizePlacement::operator()(
     const std::unordered_set<std::string> &quantize_enabled_op_types) {
   std::unordered_set<std::string> supported_op_types =
-      std::unordered_set<std::string>({"concat", "conv2d", "elementwise_add",
-                                       "fc", "matmul", "pool2d", "prior_box",
-                                       "reshape2", "transpose2", "fusion_gru"});
+      std::unordered_set<std::string>(
+          {"concat", "conv2d", "elementwise_add", "fc", "matmul", "pool2d",
+           "prior_box", "reshape2", "transpose2", "fusion_gru", "multi_gru"});
   if (!quantize_enabled_op_types.empty()) {
     supported_op_types = quantize_enabled_op_types;
   }

paddle/fluid/framework/ir/mkldnn/cpu_quantize_pass.cc

@@ -832,7 +832,7 @@ void CPUQuantizePass::QuantizeFusionGru(Graph* graph) const {
     GET_IR_NODE_FROM_SUBGRAPH(weight_x, weight_x, pattern);
     GET_IR_NODE_FROM_SUBGRAPH(out, out, pattern);
 
-    if (!AreScalesPresentForNodes({x, weight_h, weight_x})) {
+    if (!AreScalesPresentForNodes({x, weight_x})) {
       LogCannotQuantizeOp(op);
       return;
     }

paddle/fluid/inference/CMakeLists.txt

@@ -78,7 +78,6 @@ set(SHARED_INFERENCE_SRCS
     ${CMAKE_CURRENT_SOURCE_DIR}/api/analysis_predictor.cc
     ${CMAKE_CURRENT_SOURCE_DIR}/api/details/zero_copy_tensor.cc
     ${CMAKE_CURRENT_SOURCE_DIR}/utils/io_utils.cc
-    ${mkldnn_quantizer_src_file}
     ${PADDLE_CUSTOM_OP_SRCS})
 
 # shared inference library deps

paddle/fluid/inference/api/CMakeLists.txt

@@ -20,10 +20,9 @@ endif(APPLE)
 add_subdirectory(details)
 
 if(WITH_MKLDNN)
-  set(mkldnn_quantizer_src ${CMAKE_CURRENT_SOURCE_DIR}/mkldnn_quantizer.cc)
   set(mkldnn_quantizer_cfg mkldnn_quantizer_config)
+  set(mkldnn_quantizer_src ${CMAKE_CURRENT_SOURCE_DIR}/mkldnn_quantizer.cc)
   cc_library(${mkldnn_quantizer_cfg} SRCS mkldnn_quantizer_config.cc DEPS lod_tensor paddle_pass_builder)
-  set(mkldnn_quantizer_src_file ${mkldnn_quantizer_src} PARENT_SCOPE)
   set(mkldnn_quantizer_cfg ${mkldnn_quantizer_cfg} PARENT_SCOPE)
 endif()
 
@@ -71,6 +70,16 @@ elseif (WIN32)
   cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_predictor benchmark ${inference_deps}
           ARGS --dirname=${WORD2VEC_MODEL_DIR})
 endif()
+
+if(WITH_TESTING AND WITH_MKLDNN)
+  if (NOT APPLE AND NOT WIN32)
+    cc_test(test_mkldnn_quantizer SRCS mkldnn_quantizer_tester.cc DEPS paddle_inference_shared ARGS --dirname=${WORD2VEC_MODEL_DIR})
+  elseif (WIN32)
+    cc_test(test_mkldnn_quantizer SRCS mkldnn_quantizer_tester.cc DEPS analysis_predictor benchmark ${inference_deps}
+            ARGS --dirname=${WORD2VEC_MODEL_DIR})
+  endif()
+endif()
+
 if(WITH_TESTING AND TEST test_api_impl)
     if(NOT APPLE)
         set_tests_properties(test_api_impl PROPERTIES TIMEOUT 120)

paddle/fluid/inference/api/analysis_predictor_tester.cc

@@ -22,9 +22,6 @@
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/inference/tests/api/tester_helper.h"
 #include "paddle/fluid/platform/cpu_info.h"
-#ifdef PADDLE_WITH_MKLDNN
-#include "paddle/fluid/inference/api/mkldnn_quantizer.h"
-#endif
 
 DEFINE_string(dirname, "", "dirname to tests.");
 
@@ -254,242 +251,6 @@ TEST(AnalysisPredictor, memory_optim) {
 }
 */
 
-#ifdef PADDLE_WITH_MKLDNN
-class MkldnnQuantizerTest : public testing::Test {
- public:
-  MkldnnQuantizerTest() {
-    AnalysisConfig config(FLAGS_dirname);
-    predictor = std::move(CreatePaddlePredictor(config));
-    auto* predictor_p = static_cast<AnalysisPredictor*>(predictor.get());
-
-    auto qconfig = new MkldnnQuantizerConfig();
-
-    mkldnn_quantizer.reset(
-        new AnalysisPredictor::MkldnnQuantizer(*predictor_p, qconfig));
-  }
-
-  std::pair<std::vector<int>, float> Histogram(
-      const framework::LoDTensor& var_tensor, float min_val, float max_val,
-      int num_bins) const {
-    return mkldnn_quantizer->Histogram(var_tensor, min_val, max_val, num_bins);
-  }
-
-  std::pair<bool, framework::LoDTensor> GetMaxScalingFactor(
-      const framework::LoDTensor& var_tensor, bool is_unsigned) const {
-    return mkldnn_quantizer->GetMaxScalingFactor(var_tensor, is_unsigned);
-  }
-
-  std::pair<bool, framework::LoDTensor> GetMaxChScalingFactor(
-      const framework::LoDTensor& var_tensor, bool is_unsigned) const {
-    return mkldnn_quantizer->GetMaxChScalingFactor(var_tensor, is_unsigned, 0);
-  }
-
-  std::pair<bool, framework::LoDTensor> GetKLScalingFactor(
-      const framework::LoDTensor& var_tensor, bool is_unsigned) const {
-    return mkldnn_quantizer->GetKLScalingFactor(var_tensor, is_unsigned);
-  }
-
- protected:
-  std::unique_ptr<PaddlePredictor> predictor;
-  std::unique_ptr<AnalysisPredictor::MkldnnQuantizer> mkldnn_quantizer;
-  float abs_error = 1e-6;
-  static const std::array<float, 10> non_negative_values;
-  static const std::array<float, 10> positive_and_negative_values;
-};
-
-const std::array<float, 10> MkldnnQuantizerTest::non_negative_values = {
-    0.0158671, 0.026459,   0.0280772,  0.00962479, 0.0131628,
-    0.016704,  0.00118407, 0.00765726, 0.0123213,  0.00944741};
-const std::array<float, 10> MkldnnQuantizerTest::positive_and_negative_values =
-    {-0.0482659, -0.0102493, -0.00794221, -0.00387115, -0.00674586,
-     -0.0495346, 0.0629528,  -0.00531285, -0.0230353,  0.0269089};
-
-TEST_F(MkldnnQuantizerTest, histogram_inverted_min_max) {
-  const auto& values = non_negative_values;
-  auto min_val = *std::min_element(values.begin(), values.end());
-  auto max_val = *std::max_element(values.begin(), values.end());
-
-  framework::LoDTensor var_tensor;
-  var_tensor.Resize(framework::make_dim(values.size()));
-  std::copy(begin(values), end(values),
-            var_tensor.mutable_data<float>(platform::CPUPlace()));
-
-  ASSERT_THROW(Histogram(var_tensor, max_val, min_val, 3),
-               platform::EnforceNotMet);
-}
-
-TEST_F(MkldnnQuantizerTest, histogram_non_negative_to_3) {
-  // all non-negative values
-  const auto& values = non_negative_values;
-  auto min_val = *std::min_element(values.begin(), values.end());
-  auto max_val = *std::max_element(values.begin(), values.end());
-
-  framework::LoDTensor var_tensor;
-  var_tensor.Resize(framework::make_dim(values.size()));
-  std::copy(begin(values), end(values),
-            var_tensor.mutable_data<float>(platform::CPUPlace()));
-
-  std::vector<int> histogram;
-  float bin_width;
-
-  std::tie(histogram, bin_width) = Histogram(var_tensor, min_val, max_val, 3);
-
-  ASSERT_NEAR(bin_width, std::abs(max_val - min_val) / 3.f, abs_error)
-      << "Improperly calculated bin_width.";
-
-  ASSERT_EQ(histogram[0], 4);
-  ASSERT_EQ(histogram[1], 4);
-  ASSERT_EQ(histogram[2], 2);
-}
-
-TEST_F(MkldnnQuantizerTest, histogram_positive_and_negative_to_3) {
-  const auto& values = positive_and_negative_values;
-  auto min_val = *std::min_element(values.begin(), values.end());
-  auto max_val = *std::max_element(values.begin(), values.end());
-
-  framework::LoDTensor var_tensor;
-  var_tensor.Resize(framework::make_dim(values.size()));
-  std::copy(begin(values), end(values),
-            var_tensor.mutable_data<float>(platform::CPUPlace()));
-
-  std::vector<int> histogram;
-  float bin_width;
-
-  std::tie(histogram, bin_width) = Histogram(var_tensor, min_val, max_val, 3);
-
-  ASSERT_NEAR(bin_width, std::abs(max_val - min_val) / 3.0f, abs_error)
-      << "Improperly calculated bin_width.";
-
-  ASSERT_EQ(histogram[0], 3);
-  ASSERT_EQ(histogram[1], 5);
-  ASSERT_EQ(histogram[2], 2);
-}
-
-TEST_F(MkldnnQuantizerTest, histogram_zero_bins) {
-  const auto& values = non_negative_values;
-  auto min_val = *std::min_element(values.begin(), values.end());
-  auto max_val = *std::max_element(values.begin(), values.end());
-
-  framework::LoDTensor var_tensor;
-  var_tensor.Resize(framework::make_dim(values.size()));
-  std::copy(begin(values), end(values),
-            var_tensor.mutable_data<float>(platform::CPUPlace()));
-
-  ASSERT_THROW(Histogram(var_tensor, min_val, max_val, 0),
-               platform::EnforceNotMet);
-}
-
-TEST_F(MkldnnQuantizerTest, histogram_empty) {
-  // empty tensor
-  ASSERT_THROW(Histogram({}, -1, 1, 1), platform::EnforceNotMet);
-
-  // zero tensor
-  framework::LoDTensor var_tensor;
-  var_tensor.Resize({0});
-  var_tensor.mutable_data<double>(platform::CPUPlace());
-
-  ASSERT_THROW(Histogram(var_tensor, -1, 1, 1), platform::EnforceNotMet);
-}
-
-TEST_F(MkldnnQuantizerTest, kl_scaling_factor_signed) {
-  const auto& values = positive_and_negative_values;
-
-  framework::LoDTensor var_tensor;
-  var_tensor.Resize(framework::make_dim(values.size()));
-  std::copy(begin(values), end(values),
-            var_tensor.mutable_data<float>(platform::CPUPlace()));
-
-  bool is_unsigned;
-  framework::LoDTensor lod_tensor;
-
-  std::tie(is_unsigned, lod_tensor) = GetKLScalingFactor(var_tensor, false);
-
-  ASSERT_EQ(is_unsigned, false);
-  ASSERT_EQ(lod_tensor.numel(), 1);
-  ASSERT_NEAR(lod_tensor.data<double>()[0], 1.0 / 0.0899106152344, abs_error);
-}
-
-TEST_F(MkldnnQuantizerTest, max_scaling_factor_signed) {
-  const auto& values = positive_and_negative_values;
-  auto max_val = *std::max_element(values.begin(), values.end());
-
-  framework::LoDTensor var_tensor;
-  var_tensor.Resize(framework::make_dim(values.size()));
-  std::copy(begin(values), end(values),
-            var_tensor.mutable_data<float>(platform::CPUPlace()));
-
-  bool is_unsigned;
-  framework::LoDTensor lod_tensor;
-
-  std::tie(is_unsigned, lod_tensor) = GetMaxScalingFactor(var_tensor, false);
-
-  ASSERT_EQ(is_unsigned, false);
-  ASSERT_EQ(lod_tensor.numel(), 1);
-  ASSERT_NEAR(lod_tensor.data<double>()[0], 1.0 / max_val, abs_error);
-}
-
-TEST_F(MkldnnQuantizerTest, max_scaling_factor_unsigned) {
-  const auto& values = non_negative_values;
-  auto max_val = *std::max_element(values.begin(), values.end());
-
-  framework::LoDTensor var_tensor;
-  var_tensor.Resize(framework::make_dim(values.size()));
-  std::copy(begin(values), end(values),
-            var_tensor.mutable_data<float>(platform::CPUPlace()));
-
-  bool is_unsigned;
-  framework::LoDTensor lod_tensor;
-
-  std::tie(is_unsigned, lod_tensor) = GetMaxScalingFactor(var_tensor, true);
-
-  ASSERT_EQ(is_unsigned, true);
-  ASSERT_EQ(lod_tensor.numel(), 1);
-  ASSERT_NEAR(lod_tensor.data<double>()[0], 1.0 / max_val, abs_error);
-}
-
-TEST_F(MkldnnQuantizerTest, max_scaling_factor_chwise_unsigned) {
-  const auto& values = non_negative_values;
-  auto max_val = *std::max_element(values.begin(), values.end());
-  int channels = 3;
-
-  framework::LoDTensor var_tensor;
-  var_tensor.Resize(framework::make_dim(channels, 1, 1, values.size()));
-  for (int i = 0; i < channels; i++)
-    std::copy(begin(values), end(values),
-              var_tensor.mutable_data<float>(platform::CPUPlace()) +
-                  i * values.size());
-
-  bool is_unsigned;
-  framework::LoDTensor lod_tensor;
-
-  std::tie(is_unsigned, lod_tensor) = GetMaxChScalingFactor(var_tensor, true);
-
-  ASSERT_EQ(is_unsigned, true);
-  ASSERT_EQ(lod_tensor.numel(), channels);
-  for (int i = 0; i < channels; i++) {
-    ASSERT_NEAR(lod_tensor.data<double>()[i], 1.0 / max_val, abs_error);
-  }
-}
-
-TEST_F(MkldnnQuantizerTest, kl_scaling_factor_unsigned) {
-  const auto& values = non_negative_values;
-
-  framework::LoDTensor var_tensor;
-  var_tensor.Resize(framework::make_dim(values.size()));
-  std::copy(begin(values), end(values),
-            var_tensor.mutable_data<float>(platform::CPUPlace()));
-
-  bool is_unsigned;
-  framework::LoDTensor lod_tensor;
-
-  std::tie(is_unsigned, lod_tensor) = GetKLScalingFactor(var_tensor, true);
-
-  ASSERT_EQ(is_unsigned, true);
-  ASSERT_EQ(lod_tensor.numel(), 1);
-  ASSERT_NEAR(lod_tensor.data<double>()[0], 1.0 / 0.0252845321362, abs_error);
-}
-#endif
-
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
 TEST(AnalysisPredictor, bf16_gpu_pass_strategy) {
   AnalysisConfig config;

paddle/fluid/inference/api/mkldnn_quantizer.cc

@@ -35,6 +35,7 @@ namespace paddle {
 
 using platform::CPUPlace;
 using framework::LoDTensor;
+using framework::Variable;
 using framework::ir::Graph;
 using ConstEigenVectorArrayMap =
     Eigen::Map<const Eigen::Array<float, Eigen::Dynamic, 1>>;
@@ -44,90 +45,126 @@ using EigenMatrixArray =
     Eigen::Array<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>;
 using ConstEigenMatrixArrayMap = Eigen::Map<const EigenMatrixArray>;
 using string::PrettyLogH1;
+using VariableNameMap = std::map<std::string, std::vector<std::string>>;
 static LoDTensor CreateScaleTensor(int64_t channels_num = 1);
 
+static void check_var(const Variable* var, const std::string& var_name) {
+  PADDLE_ENFORCE_NOT_NULL(var, platform::errors::PreconditionNotMet(
+                                   "%s is not in the scope", var_name));
+  PADDLE_ENFORCE_EQ(
+      var->IsType<LoDTensor>(), true,
+      platform::errors::PreconditionNotMet("Only support lod tensor now."));
+}
+
+static void check_tensor(const LoDTensor& tensor) {
+  PADDLE_ENFORCE_GT(tensor.dims().size(), 0, platform::errors::InvalidArgument(
+                                                 "Tensor dimension is empty."));
+}
+
+void AnalysisPredictor::MkldnnQuantizer::CalculateScalesForGRUWeights(
+    const paddle::framework::OpDesc* op) {
+  const auto& wx_names = op->Input("WeightX");
+  const auto& wh_names = op->Input("WeightH");
+  for (size_t i = 0; i < wx_names.size(); ++i) {
+    const auto& wx_name = wx_names[i];
+    const auto& wh_name = wh_names[i];
+    auto* wx_var = predictor_.sub_scope_->FindVar(wx_name);
+    auto* wh_var = predictor_.sub_scope_->FindVar(wh_name);
+    check_var(wx_var, wx_name);
+    check_var(wh_var, wh_name);
+    LoDTensor* wx_tensor = wx_var->GetMutable<LoDTensor>();
+    LoDTensor* wh_tensor = wh_var->GetMutable<LoDTensor>();
+    scales_[wx_name] = GetMaxChGRUScalingFactor(*wx_tensor, *wh_tensor);
+  }
+}
+
+void AnalysisPredictor::MkldnnQuantizer::CalculateScalesForOpInputs(
+    const paddle::framework::OpDesc* op) {
+  if (op->Type() == "fusion_gru" || op->Type() == "multi_gru") {
+    CalculateScalesForGRUWeights(op);
+  }
+  for (auto const& input : op->Inputs()) {
+    for (const auto& var_name : input.second) {
+      // skip if scale already computed
+      if (scales_.find(var_name) != scales_.end()) continue;
+      auto* var = predictor_.sub_scope_->FindVar(var_name);
+      check_var(var, var_name);
+      LoDTensor* var_tensor = var->GetMutable<LoDTensor>();
+      // force unsigned type if already know it
+      bool is_unsigned = false;
+      CalculateSingleScale(op->Type(), input.first, var_name, *var_tensor,
+                           is_unsigned);
+    }
+  }
+}
+
+void AnalysisPredictor::MkldnnQuantizer::CalculateScalesForOpOutputs(
+    const paddle::framework::OpDesc* op) {
+  for (auto const& output : op->Outputs()) {
+    for (const auto& var_name : output.second) {
+      // skip if scale already computed
+      if (scales_.find(var_name) != scales_.end()) continue;
+      auto* var = predictor_.sub_scope_->FindVar(var_name);
+      check_var(var, var_name);
+      LoDTensor* var_tensor = var->GetMutable<LoDTensor>();
+      // force unsigned type if already know it
+      bool is_unsigned = false;
+      bool compute_scale = true;
+      if (op->Type() == "conv2d" || op->Type() == "fc") {
+        // output of conv2d with relu must be unsigned
+        std::string fuse_activation =
+            op->GetAttrIfExists<std::string>("fuse_activation");
+        is_unsigned = (fuse_activation == "relu" || fuse_activation == "relu6");
+      } 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];
+        PADDLE_ENFORCE_NE(scales_.find(input_var_name), scales_.end(),
+                          platform::errors::PreconditionNotMet(
+                              "Input scales must be calculated before the "
+                              "output scales to infer if output is unsigned."));
+        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_NE(
+              scales_.find(input_var_name), scales_.end(),
+              platform::errors::PreconditionNotMet(
+                  "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;
+      }
+      if (compute_scale) {
+        CalculateSingleScale(op->Type(), output.first, var_name, *var_tensor,
+                             is_unsigned);
+      }
+    }
+  }
+}
+
 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()) {
     if (platform::HasOpINT8DataType(op)) {
-      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) {
-          for (const auto& var_name : conn.second) {
-            // skip if scale already computed
-            if (scales_.find(var_name) != scales_.end()) continue;
-
-            auto* var = predictor_.sub_scope_->FindVar(var_name);
-            PADDLE_ENFORCE_NOT_NULL(var,
-                                    platform::errors::PreconditionNotMet(
-                                        "%s is not in the scope", var_name));
-            PADDLE_ENFORCE_EQ(var->IsType<LoDTensor>(), true,
-                              platform::errors::PreconditionNotMet(
-                                  "Only support lod tensor now."));
-            LoDTensor* var_tensor = var->GetMutable<LoDTensor>();
-
-            // force unsigned type if already know it
-            bool is_unsigned = false;
-            bool compute_scale = true;
-            if (is_output) {
-              if (op->Type() == "conv2d" || op->Type() == "fc") {
-                // output of conv2d with relu must be unsigned
-                std::string fuse_activation =
-                    op->GetAttrIfExists<std::string>("fuse_activation");
-                is_unsigned =
-                    (fuse_activation == "relu" || fuse_activation == "relu6");
-              } 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];
-                PADDLE_ENFORCE_NE(
-                    scales_.find(input_var_name), scales_.end(),
-                    platform::errors::PreconditionNotMet(
-                        "Input scales must be calculated before the "
-                        "output scales to infer if output is unsigned."));
-                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_NE(
-                      scales_.find(input_var_name), scales_.end(),
-                      platform::errors::PreconditionNotMet(
-                          "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;
-              }
-            }
-            if (compute_scale)
-              CalculateSingleScale(op->Type(), conn.first, var_name,
-                                   *var_tensor, is_unsigned);
-          }
-        }
-      };
-
       // handle inputs first to let is_unsigned be inferred for the outputs
-      glambda(connections_in, false /* is_output */);
-      glambda(connections_out, true /* is_output */);
+      CalculateScalesForOpInputs(op);
+      CalculateScalesForOpOutputs(op);
     }
   }
-
   return true;
 }
 
@@ -339,9 +376,7 @@ AnalysisPredictor::MkldnnQuantizer::GetMaxScalingFactor(
 std::pair<bool, LoDTensor>
 AnalysisPredictor::MkldnnQuantizer::GetMaxChScalingFactor(
     const LoDTensor& var_tensor, bool is_unsigned, bool is_transposed) const {
-  PADDLE_ENFORCE_GT(
-      var_tensor.dims().size(), 0,
-      platform::errors::InvalidArgument("Tensor dimension is empty."));
+  check_tensor(var_tensor);
 
   ConstEigenVectorArrayMap eigen_tensor{var_tensor.data<float>(),
                                         var_tensor.numel(), 1};
@@ -374,6 +409,61 @@ AnalysisPredictor::MkldnnQuantizer::GetMaxChScalingFactor(
   return std::make_pair(is_unsigned, scale_tensor);
 }
 
+std::pair<bool, LoDTensor>
+AnalysisPredictor::MkldnnQuantizer::GetMaxChGRUScalingFactor(
+    const LoDTensor& wx_tensor, const LoDTensor& wh_tensor) const {
+  check_tensor(wx_tensor);
+  check_tensor(wh_tensor);
+
+  int OC = wh_tensor.dims()[0];
+  std::vector<float> scale_ur(2 * OC);
+  std::vector<float> scale_o(OC);
+
+  for (int row_id = 0; row_id < wx_tensor.dims()[0]; row_id++) {
+    for (int col_id = 0; col_id < 2 * OC; col_id++) {
+      int idx = (row_id * wx_tensor.dims()[1]) + col_id;
+      auto abs_value = std::abs(wx_tensor.data<float>()[idx]);
+      if (row_id == 0) {
+        scale_ur[col_id] = abs_value;
+      } else {
+        if (abs_value > scale_ur[col_id]) scale_ur[col_id] = abs_value;
+      }
+    }
+  }
+
+  for (int i = 0; i < 2 * OC * OC; i++) {
+    int col_id = i % (2 * OC);
+    auto abs_value = std::abs(wh_tensor.data<float>()[i]);
+    if (abs_value > scale_ur[col_id]) scale_ur[col_id] = abs_value;
+  }
+
+  for (int row_id = 0; row_id < wx_tensor.dims()[0]; row_id++) {
+    for (int col_id = 2 * OC; col_id < wx_tensor.dims()[1]; col_id++) {
+      int idx = (row_id * wx_tensor.dims()[1]) + col_id;
+      auto abs_value = std::abs(wx_tensor.data<float>()[idx]);
+      if (row_id == 0) {
+        scale_o[col_id % OC] = abs_value;
+      } else {
+        if (abs_value > scale_o[col_id]) scale_o[col_id % OC] = abs_value;
+      }
+    }
+  }
+
+  for (int i = 2 * OC * OC; i < OC * wh_tensor.dims()[1]; i++) {
+    int col_id = i % OC;
+    auto abs_value = std::abs(wh_tensor.data<float>()[i]);
+    if (abs_value > scale_o[col_id]) scale_o[col_id] = abs_value;
+  }
+  scale_ur.insert(scale_ur.end(), scale_o.begin(), scale_o.end());
+  transform(scale_ur.begin(), scale_ur.end(), scale_ur.begin(),
+            [](float& c) { return 1 / c; });
+  LoDTensor scale_tensor = CreateScaleTensor(scale_ur.size());
+  auto* scale_ptr = scale_tensor.mutable_data<double>(CPUPlace());
+  std::copy(scale_ur.begin(), scale_ur.end(), scale_ptr);
+  bool is_unsigned = false;
+  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,

paddle/fluid/inference/api/mkldnn_quantizer.h

@@ -52,7 +52,7 @@ class AnalysisPredictor::MkldnnQuantizer {
   // Execute full quantization procedure.
   bool Quantize();
 
-#if PADDLE_WITH_TESTING
+#ifdef PADDLE_WITH_TESTING
   friend class MkldnnQuantizerTest;
 #endif
 
@@ -67,6 +67,11 @@ class AnalysisPredictor::MkldnnQuantizer {
                             const std::string& var_name,
                             const framework::LoDTensor& var_tensor,
                             bool is_unsigned);
+  void CalculateSingleGRUWeightsScale(const std::string& var_name,
+                                      const framework::LoDTensor& var_tensor);
+  void CalculateScalesForGRUWeights(const paddle::framework::OpDesc* op);
+  void CalculateScalesForOpOutputs(const paddle::framework::OpDesc* op);
+  void CalculateScalesForOpInputs(const paddle::framework::OpDesc* op);
   void PrepareArgument() const;
   void ClearDeviceContext() const;
   bool RunQuantizePasses() const;
@@ -82,6 +87,10 @@ class AnalysisPredictor::MkldnnQuantizer {
       const framework::LoDTensor& var_tensor, bool is_unsigned,
       bool is_transposed) const;
 
+  std::pair<bool, framework::LoDTensor> GetMaxChGRUScalingFactor(
+      const framework::LoDTensor& wx_tensor,
+      const framework::LoDTensor& wh_tensor) const;
+
   std::pair<bool, framework::LoDTensor> GetMaxScalingFactor(
       const framework::LoDTensor& var_tensor, bool is_unsigned) const;
 

paddle/fluid/inference/api/mkldnn_quantizer_config.cc

@@ -55,6 +55,10 @@ MkldnnQuantizerConfig::MkldnnQuantizerConfig() {
   rules_["elementwise_add"]["Y"] = ScaleAlgo::KL;
   rules_["elementwise_add"]["Out"] = ScaleAlgo::KL;
 
+  rules_["elementwise_mul"]["X"] = ScaleAlgo::KL;
+  rules_["elementwise_mul"]["Y"] = ScaleAlgo::KL;
+  rules_["elementwise_mul"]["Out"] = ScaleAlgo::KL;
+
   // Reshape2 does not perform calculation on the data and shapes are not
   // changed. Scale is calculated on input data and assign to Quantize and
   // Dequantize scale.
@@ -63,6 +67,24 @@ MkldnnQuantizerConfig::MkldnnQuantizerConfig() {
   rules_["reshape2"]["ShapeTensor"] = ScaleAlgo::NONE;
   rules_["reshape2"]["XShape"] = ScaleAlgo::NONE;
   rules_["reshape2"]["Out"] = ScaleAlgo::NONE;
+
+  rules_["fusion_gru"]["X"] = ScaleAlgo::KL;
+  rules_["fusion_gru"]["H0"] = ScaleAlgo::NONE;
+  rules_["fusion_gru"]["Bias"] = ScaleAlgo::NONE;
+  rules_["fusion_gru"]["WeightX"] = ScaleAlgo::NONE;  // Weights will be handled
+  rules_["fusion_gru"]["WeightH"] = ScaleAlgo::NONE;  // separately
+  rules_["fusion_gru"]["ReorderedH0"] = ScaleAlgo::NONE;
+  rules_["fusion_gru"]["XX"] = ScaleAlgo::NONE;
+  rules_["fusion_gru"]["BatchedInput"] = ScaleAlgo::NONE;
+  rules_["fusion_gru"]["BatchedOut"] = ScaleAlgo::NONE;
+  rules_["fusion_gru"]["Hidden"] = ScaleAlgo::KL;
+
+  rules_["multi_gru"]["X"] = ScaleAlgo::KL;
+  rules_["multi_gru"]["Bias"] = ScaleAlgo::NONE;
+  rules_["multi_gru"]["WeightX"] = ScaleAlgo::NONE;  // Weights will be handled
+  rules_["multi_gru"]["WeightH"] = ScaleAlgo::NONE;  // separately
+  rules_["multi_gru"]["Scale_weights"] = ScaleAlgo::NONE;
+  rules_["multi_gru"]["Hidden"] = ScaleAlgo::KL;
 }
 
 ScaleAlgo MkldnnQuantizerConfig::scale_algo(

paddle/fluid/inference/api/mkldnn_quantizer_tester.cc

+// Copyright (c) 2021 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 <gtest/gtest.h>
+#include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/inference/api/analysis_predictor.h"
+#include "paddle/fluid/inference/api/paddle_inference_api.h"
+
+DEFINE_string(dirname, "", "dirname to tests.");
+
+namespace paddle {
+
+class MkldnnQuantizerTest : public testing::Test {
+ public:
+  MkldnnQuantizerTest() {
+    AnalysisConfig config(FLAGS_dirname);
+    predictor = std::move(CreatePaddlePredictor(config));
+    auto* predictor_p = static_cast<AnalysisPredictor*>(predictor.get());
+
+    auto qconfig = new MkldnnQuantizerConfig();
+
+    mkldnn_quantizer.reset(
+        new AnalysisPredictor::MkldnnQuantizer(*predictor_p, qconfig));
+  }
+
+  std::pair<std::vector<int>, float> Histogram(
+      const framework::LoDTensor& var_tensor, float min_val, float max_val,
+      int num_bins) const {
+    return mkldnn_quantizer->Histogram(var_tensor, min_val, max_val, num_bins);
+  }
+
+  std::pair<bool, framework::LoDTensor> GetMaxScalingFactor(
+      const framework::LoDTensor& var_tensor, bool is_unsigned) const {
+    return mkldnn_quantizer->GetMaxScalingFactor(var_tensor, is_unsigned);
+  }
+
+  std::pair<bool, framework::LoDTensor> GetMaxChScalingFactor(
+      const framework::LoDTensor& var_tensor, bool is_unsigned) const {
+    return mkldnn_quantizer->GetMaxChScalingFactor(var_tensor, is_unsigned, 0);
+  }
+
+  std::pair<bool, framework::LoDTensor> GetKLScalingFactor(
+      const framework::LoDTensor& var_tensor, bool is_unsigned) const {
+    return mkldnn_quantizer->GetKLScalingFactor(var_tensor, is_unsigned);
+  }
+
+  std::pair<bool, framework::LoDTensor> GetMaxChGRUScalingFactor(
+      const framework::LoDTensor& wx_tensor,
+      const framework::LoDTensor& wh_tensor) const {
+    return mkldnn_quantizer->GetMaxChGRUScalingFactor(wx_tensor, wh_tensor);
+  }
+
+ protected:
+  std::unique_ptr<PaddlePredictor> predictor;
+  std::unique_ptr<AnalysisPredictor::MkldnnQuantizer> mkldnn_quantizer;
+  float abs_error = 1e-6;
+  static const std::array<float, 10> non_negative_values;
+  static const std::array<float, 10> positive_and_negative_values;
+};
+
+const std::array<float, 10> MkldnnQuantizerTest::non_negative_values = {
+    0.0158671, 0.026459,   0.0280772,  0.00962479, 0.0131628,
+    0.016704,  0.00118407, 0.00765726, 0.0123213,  0.00944741};
+const std::array<float, 10> MkldnnQuantizerTest::positive_and_negative_values =
+    {-0.0482659, -0.0102493, -0.00794221, -0.00387115, -0.00674586,
+     -0.0495346, 0.0629528,  -0.00531285, -0.0230353,  0.0269089};
+
+TEST_F(MkldnnQuantizerTest, histogram_inverted_min_max) {
+  const auto& values = non_negative_values;
+  auto min_val = *std::min_element(values.begin(), values.end());
+  auto max_val = *std::max_element(values.begin(), values.end());
+
+  framework::LoDTensor var_tensor;
+  var_tensor.Resize(framework::make_dim(values.size()));
+  std::copy(begin(values), end(values),
+            var_tensor.mutable_data<float>(platform::CPUPlace()));
+
+  ASSERT_THROW(Histogram(var_tensor, max_val, min_val, 3),
+               platform::EnforceNotMet);
+}
+
+TEST_F(MkldnnQuantizerTest, histogram_non_negative_to_3) {
+  // all non-negative values
+  const auto& values = non_negative_values;
+  auto min_val = *std::min_element(values.begin(), values.end());
+  auto max_val = *std::max_element(values.begin(), values.end());
+
+  framework::LoDTensor var_tensor;
+  var_tensor.Resize(framework::make_dim(values.size()));
+  std::copy(begin(values), end(values),
+            var_tensor.mutable_data<float>(platform::CPUPlace()));
+
+  std::vector<int> histogram;
+  float bin_width;
+
+  std::tie(histogram, bin_width) = Histogram(var_tensor, min_val, max_val, 3);
+
+  ASSERT_NEAR(bin_width, std::abs(max_val - min_val) / 3.f, abs_error)
+      << "Improperly calculated bin_width.";
+
+  ASSERT_EQ(histogram[0], 4);
+  ASSERT_EQ(histogram[1], 4);
+  ASSERT_EQ(histogram[2], 2);
+}
+
+TEST_F(MkldnnQuantizerTest, histogram_positive_and_negative_to_3) {
+  const auto& values = positive_and_negative_values;
+  auto min_val = *std::min_element(values.begin(), values.end());
+  auto max_val = *std::max_element(values.begin(), values.end());
+
+  framework::LoDTensor var_tensor;
+  var_tensor.Resize(framework::make_dim(values.size()));
+  std::copy(begin(values), end(values),
+            var_tensor.mutable_data<float>(platform::CPUPlace()));
+
+  std::vector<int> histogram;
+  float bin_width;
+
+  std::tie(histogram, bin_width) = Histogram(var_tensor, min_val, max_val, 3);
+
+  ASSERT_NEAR(bin_width, std::abs(max_val - min_val) / 3.0f, abs_error)
+      << "Improperly calculated bin_width.";
+
+  ASSERT_EQ(histogram[0], 3);
+  ASSERT_EQ(histogram[1], 5);
+  ASSERT_EQ(histogram[2], 2);
+}
+
+TEST_F(MkldnnQuantizerTest, histogram_zero_bins) {
+  const auto& values = non_negative_values;
+  auto min_val = *std::min_element(values.begin(), values.end());
+  auto max_val = *std::max_element(values.begin(), values.end());
+
+  framework::LoDTensor var_tensor;
+  var_tensor.Resize(framework::make_dim(values.size()));
+  std::copy(begin(values), end(values),
+            var_tensor.mutable_data<float>(platform::CPUPlace()));
+
+  ASSERT_THROW(Histogram(var_tensor, min_val, max_val, 0),
+               platform::EnforceNotMet);
+}
+
+TEST_F(MkldnnQuantizerTest, histogram_empty) {
+  // empty tensor
+  ASSERT_THROW(Histogram({}, -1, 1, 1), platform::EnforceNotMet);
+
+  // zero tensor
+  framework::LoDTensor var_tensor;
+  var_tensor.Resize({0});
+  var_tensor.mutable_data<double>(platform::CPUPlace());
+
+  ASSERT_THROW(Histogram(var_tensor, -1, 1, 1), platform::EnforceNotMet);
+}
+
+TEST_F(MkldnnQuantizerTest, kl_scaling_factor_signed) {
+  const auto& values = positive_and_negative_values;
+
+  framework::LoDTensor var_tensor;
+  var_tensor.Resize(framework::make_dim(values.size()));
+  std::copy(begin(values), end(values),
+            var_tensor.mutable_data<float>(platform::CPUPlace()));
+
+  bool is_unsigned;
+  framework::LoDTensor lod_tensor;
+
+  std::tie(is_unsigned, lod_tensor) = GetKLScalingFactor(var_tensor, false);
+
+  ASSERT_EQ(is_unsigned, false);
+  ASSERT_EQ(lod_tensor.numel(), 1);
+  ASSERT_NEAR(lod_tensor.data<double>()[0], 1.0 / 0.0899106152344, abs_error);
+}
+
+TEST_F(MkldnnQuantizerTest, max_scaling_factor_signed) {
+  const auto& values = positive_and_negative_values;
+  auto max_val = *std::max_element(values.begin(), values.end());
+
+  framework::LoDTensor var_tensor;
+  var_tensor.Resize(framework::make_dim(values.size()));
+  std::copy(begin(values), end(values),
+            var_tensor.mutable_data<float>(platform::CPUPlace()));
+
+  bool is_unsigned;
+  framework::LoDTensor lod_tensor;
+
+  std::tie(is_unsigned, lod_tensor) = GetMaxScalingFactor(var_tensor, false);
+
+  ASSERT_EQ(is_unsigned, false);
+  ASSERT_EQ(lod_tensor.numel(), 1);
+  ASSERT_NEAR(lod_tensor.data<double>()[0], 1.0 / max_val, abs_error);
+}
+
+TEST_F(MkldnnQuantizerTest, max_scaling_factor_unsigned) {
+  const auto& values = non_negative_values;
+  auto max_val = *std::max_element(values.begin(), values.end());
+
+  framework::LoDTensor var_tensor;
+  var_tensor.Resize(framework::make_dim(values.size()));
+  std::copy(begin(values), end(values),
+            var_tensor.mutable_data<float>(platform::CPUPlace()));
+
+  bool is_unsigned;
+  framework::LoDTensor lod_tensor;
+
+  std::tie(is_unsigned, lod_tensor) = GetMaxScalingFactor(var_tensor, true);
+
+  ASSERT_EQ(is_unsigned, true);
+  ASSERT_EQ(lod_tensor.numel(), 1);
+  ASSERT_NEAR(lod_tensor.data<double>()[0], 1.0 / max_val, abs_error);
+}
+
+TEST_F(MkldnnQuantizerTest, max_scaling_factor_chwise_unsigned) {
+  const auto& values = non_negative_values;
+  auto max_val = *std::max_element(values.begin(), values.end());
+  int channels = 3;
+
+  framework::LoDTensor var_tensor;
+  var_tensor.Resize(framework::make_dim(channels, 1, 1, values.size()));
+  for (int i = 0; i < channels; i++)
+    std::copy(begin(values), end(values),
+              var_tensor.mutable_data<float>(platform::CPUPlace()) +
+                  i * values.size());
+
+  bool is_unsigned;
+  framework::LoDTensor lod_tensor;
+
+  std::tie(is_unsigned, lod_tensor) = GetMaxChScalingFactor(var_tensor, true);
+
+  ASSERT_EQ(is_unsigned, true);
+  ASSERT_EQ(lod_tensor.numel(), channels);
+  for (int i = 0; i < channels; i++) {
+    ASSERT_NEAR(lod_tensor.data<double>()[i], 1.0 / max_val, abs_error);
+  }
+}
+
+TEST_F(MkldnnQuantizerTest, kl_scaling_factor_unsigned) {
+  const auto& values = non_negative_values;
+
+  framework::LoDTensor var_tensor;
+  var_tensor.Resize(framework::make_dim(values.size()));
+  std::copy(begin(values), end(values),
+            var_tensor.mutable_data<float>(platform::CPUPlace()));
+
+  bool is_unsigned;
+  framework::LoDTensor lod_tensor;
+
+  std::tie(is_unsigned, lod_tensor) = GetKLScalingFactor(var_tensor, true);
+
+  ASSERT_EQ(is_unsigned, true);
+  ASSERT_EQ(lod_tensor.numel(), 1);
+  ASSERT_NEAR(lod_tensor.data<double>()[0], 1.0 / 0.0252845321362, abs_error);
+}
+
+const std::vector<std::vector<float>> wx = {
+    {0.04347931, -0.5643393, 0.7551297, 0.26713502, 0.8055306, 0.91144973},
+    {0.01707571, 0.12741385, 0.15419468, 0.66127586, 0.46821925, 0.9665961},
+    {0.40393898, 0.884427, -0.5853097, 0.5840954, 0.9170512, 0.98245513}};
+const std::vector<std::vector<float>> wh = {
+    {0.42484227, -0.9025513, 0.17087583, 0.8403284, 0.03325734, 0.92331886},
+    {0.32630175, 0.41691914, 0.99848574, 0.3504407, 0.06707559, 0.62239844}};
+
+TEST_F(MkldnnQuantizerTest, max_ch_gru_scaling_factor) {
+  framework::LoDTensor wx_tensor, wh_tensor, lod_tensor;
+
+  wx_tensor.Resize(framework::make_dim(wx.size(), wx[0].size()));
+  for (size_t i = 0; i < wx.size(); i++)
+    std::copy(
+        begin(wx[i]), end(wx[i]),
+        wx_tensor.mutable_data<float>(platform::CPUPlace()) + i * wx[0].size());
+
+  wh_tensor.Resize(framework::make_dim(wh.size(), wh[0].size()));
+  for (size_t i = 0; i < wh.size(); i++)
+    std::copy(
+        begin(wh[i]), end(wh[i]),
+        wh_tensor.mutable_data<float>(platform::CPUPlace()) + i * wh[0].size());
+
+  bool is_unsigned;
+  std::tie(is_unsigned, lod_tensor) =
+      GetMaxChGRUScalingFactor(wx_tensor, wh_tensor);
+
+  std::vector<double> scales = {2.35381475, 1.08304947, 1.32427582,
+                                1.19001095, 1.00151656, 1.01785819};
+  ASSERT_EQ(is_unsigned, false);
+  ASSERT_EQ(lod_tensor.numel(), static_cast<int64_t>(scales.size()));
+  for (int64_t i = 0; i < lod_tensor.numel(); i++) {
+    ASSERT_NEAR(lod_tensor.data<double>()[i], scales[i], abs_error);
+  }
+}
+}  // namespace paddle

paddle/fluid/inference/api/paddle_mkldnn_quantizer_config.h

@@ -45,7 +45,9 @@ enum class ScaleAlgo {
   MAX_CH,    ///< Find scale based on the max absolute value per output channel
   MAX_CH_T,  ///< Find scale based on the max absolute value per output channel
              ///< of a transposed tensor
-  KL,        ///< Find scale based on KL Divergence
+  MAX_CH_GRU,  ///< Find scale based on the max absolute value per output
+               /// channel for fusion_gru/multi_gru operators
+  KL,          ///< Find scale based on KL Divergence
 };
 
 ///

paddle/fluid/inference/tests/api/CMakeLists.txt

@@ -195,6 +195,21 @@ function(inference_analysis_api_lexical_bfloat16_test_run TARGET_NAME test_binar
              --iterations=2)
 endfunction()
 
+function(inference_analysis_api_lexical_int8_test_run TARGET_NAME test_binary infer_model data_path fuse_multi_gru)
+    inference_analysis_test_run(${TARGET_NAME}
+    COMMAND ${test_binary}
+        ARGS --infer_model=${infer_model}
+             --infer_data=${data_path}
+             --batch_size=100
+             --cpu_num_threads=${CPU_NUM_THREADS_ON_CI}
+             --with_accuracy_layer=true
+             --use_analysis=true
+             --enable_int8=true
+             --quantized_accuracy=0.01
+             --fuse_multi_gru=${fuse_multi_gru}
+             --iterations=4)
+endfunction()
+
 function(preprocess_data2bin_test_run target py_script_source data_dir output_file)
 	py_test(${target} SRCS ${CMAKE_CURRENT_SOURCE_DIR}/${py_script_source}
 	        ARGS --data_dir=${data_dir}
@@ -475,6 +490,10 @@ if(WITH_MKLDNN)
   inference_analysis_api_lexical_test_run(test_analyzer_lexical_gru ${LEXICAL_TEST_APP} ${GRU_MODEL_PATH} ${GRU_DATA_PATH})
   # run bfloat16 lexical analysis test
   inference_analysis_api_lexical_bfloat16_test_run(test_analyzer_lexical_gru_bfloat16 ${LEXICAL_TEST_APP} ${GRU_MODEL_PATH} ${GRU_DATA_PATH})
+  # run post-training quantization lexical analysis test
+  inference_analysis_api_lexical_int8_test_run(test_analyzer_lexical_gru_int8 ${LEXICAL_TEST_APP} ${GRU_MODEL_PATH} ${GRU_DATA_PATH} false)
+  # run post-training quantization lexical analysis test with multi_gru fuse 
+  inference_analysis_api_lexical_int8_test_run(test_analyzer_lexical_gru_int8_multi_gru ${LEXICAL_TEST_APP} ${GRU_MODEL_PATH} ${GRU_DATA_PATH} true)
 
   ### optimized FP32 vs. Quant INT8 tests
   

paddle/fluid/inference/tests/api/analyzer_lexical_analysis_gru_tester.cc

@@ -54,6 +54,51 @@ std::vector<size_t> ReadSentenceLod(std::ifstream &file, size_t offset,
   return sentence_lod;
 }
 
+std::shared_ptr<std::vector<PaddleTensor>> WarmupData(
+    const std::vector<std::vector<PaddleTensor>> &test_data,
+    int num_images = 1) {
+  int data_size = test_data.size();
+
+  PADDLE_ENFORCE_LE(static_cast<size_t>(num_images), data_size,
+                    platform::errors::InvalidArgument(
+                        "The requested quantization warmup data size must be "
+                        "lower or equal to the test data size. But received"
+                        "warmup size is %d and test data size is %d",
+                        num_images, data_size));
+  int words_shape = test_data[0][0].shape[0];
+  PaddleTensor words;
+  words.name = "words";
+  words.shape = {words_shape, 1};
+  words.dtype = PaddleDType::INT64;
+  words.data.Resize(sizeof(int64_t) * words_shape);
+
+  int target_shape = test_data[0][1].shape[0];
+  PaddleTensor targets;
+  targets.name = "targets";
+  targets.shape = {target_shape, 1};
+  targets.dtype = PaddleDType::INT64;
+  targets.data.Resize(sizeof(int64_t) * target_shape);
+
+  for (int i = 0; i < num_images; i++) {
+    std::copy_n(
+        static_cast<int64_t *>(test_data[i][0].data.data()) + i * words_shape,
+        words_shape,
+        static_cast<int64_t *>(words.data.data()) + i * words_shape);
+    words.lod = test_data[i][0].lod;
+
+    std::copy_n(
+        static_cast<int64_t *>(test_data[i][1].data.data()) + i * target_shape,
+        target_shape,
+        static_cast<int64_t *>(targets.data.data()) + i * target_shape);
+    targets.lod = test_data[i][1].lod;
+  }
+
+  auto warmup_data = std::make_shared<std::vector<PaddleTensor>>(2);
+  (*warmup_data)[0] = std::move(words);
+  (*warmup_data)[1] = std::move(targets);
+  return warmup_data;
+}
+
 template <typename T>
 class TensorReader {
  public:
@@ -210,7 +255,19 @@ TEST(Analyzer_lexical_test, Analyzer_lexical_analysis) {
   if (FLAGS_use_analysis) {
     AnalysisConfig analysis_cfg;
     SetAnalysisConfig(&analysis_cfg, FLAGS_cpu_num_threads);
-    if (FLAGS_enable_bf16) analysis_cfg.EnableMkldnnBfloat16();
+    if (FLAGS_enable_bf16) {
+      analysis_cfg.EnableMkldnnBfloat16();
+    } else if (FLAGS_enable_int8) {
+      if (FLAGS_fuse_multi_gru)
+        analysis_cfg.pass_builder()->AppendPass("multi_gru_fuse_pass");
+
+      std::shared_ptr<std::vector<PaddleTensor>> warmup_data =
+          WarmupData(input_slots_all);
+      analysis_cfg.EnableMkldnnQuantizer();
+      analysis_cfg.mkldnn_quantizer_config()->SetWarmupData(warmup_data);
+      analysis_cfg.mkldnn_quantizer_config()->SetWarmupBatchSize(
+          FLAGS_batch_size);
+    }
     std::vector<double> acc_analysis(3);
     acc_analysis = Lexical_Test(input_slots_all, &outputs, &analysis_cfg, true);
     for (size_t i = 0; i < acc_analysis.size(); i++) {

paddle/fluid/inference/tests/api/tester_helper.h

@@ -73,6 +73,8 @@ DEFINE_int32(warmup_iters, 1, "Number of batches to process during warmup.");
 
 DEFINE_bool(enable_profile, false, "Turn on profiler for fluid");
 DEFINE_int32(cpu_num_threads, 1, "Number of threads for each paddle instance.");
+DEFINE_bool(fuse_multi_gru, false,
+            "Running the inference program with multi_gru_fuse_pass");
 
 namespace paddle {
 namespace inference {