INT8 Fully-connected (#17641)

* Implement Int8 FC

* Integrate FC into INT8v2

test=develop

* int8 FC: transpose weights before computing scales

test=develop

* Add support for activation_type string in FC

test=develop

* Disable MKL-DNN's FC in VGG16 and 19

test=develop

* Disable FC quantization when mkldnn FC is disabled

test=develop

* Solve PADDLE_ENFORCES in FC int8

* Fix Paddle enforces and remove const cast

test=develop

* Fix style changes

test=develop

* Fix quantizer_tester test and add fc quantization

test=develop

* Fix FC test fail on CUDA

* Remove unnecessary log from quantize placement pass

test=develop

* Add Thread ID to FC hash key

test=develop

* Add comments to MKL-DNN FC Kernel

test=develop

* Refactor quantizer

test=develop

* Fix linter issues

test=develop

* Fix crash in slim googlenet

test=develop

* Fix PADDLE_ENFORCE messages

test=develop

cmake/operators.cmake

@@ -190,6 +190,10 @@ function(op_library TARGET)
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(transpose2, MKLDNN, FP32);\n")
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(transpose2, MKLDNN, S8);\n")
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(transpose2, MKLDNN, U8);\n")
+      elseif(${MKLDNN_FILE} STREQUAL "fc_mkldnn_op")
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(fc, MKLDNN, FP32);\n")
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(fc, MKLDNN, S8);\n")
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(fc, MKLDNN, U8);\n")
       else()
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n")
       endif()

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -905,15 +905,17 @@ PDNode *patterns::FCMKLDNN::operator()(paddle::framework::ir::PDNode *x,
 
   auto *fc_op = pattern->NewNode(fc_repr())->assert_is_op("fc");
   // Create variables
+  // Input
+  auto *input_var = pattern->NewNode(input_repr())
+                        ->AsInput()
+                        ->assert_is_op_input("fc", "Input");
   // Filter
   auto *fc_weight_var = pattern->NewNode(weights_repr())
                             ->AsInput()
-                            ->assert_is_persistable_var()
                             ->assert_is_op_input("fc", "W");
   // Bias
   auto *fc_bias_var = pattern->NewNode(bias_repr())
                           ->AsInput()
-                          ->assert_is_persistable_var()
                           ->assert_is_op_input("fc", "Bias");
   // Output
   auto *fc_out_var = pattern->NewNode(output_repr())
@@ -921,7 +923,8 @@ PDNode *patterns::FCMKLDNN::operator()(paddle::framework::ir::PDNode *x,
                          ->assert_is_op_output("fc", "Out")
                          ->assert_is_only_output_of_op("fc");
 
-  fc_op->LinksFrom({x, fc_weight_var, fc_bias_var}).LinksTo({fc_out_var});
+  fc_op->LinksFrom({input_var, fc_weight_var, fc_bias_var})
+      .LinksTo({fc_out_var});
   return fc_out_var;
 }
 

paddle/fluid/framework/ir/graph_pattern_detector.h

@@ -517,6 +517,7 @@ struct FCMKLDNN : public PatternBase {
   // declare operator node's name
   PATTERN_DECL_NODE(fc);
   // declare variable node's name
+  PATTERN_DECL_NODE(input);
   PATTERN_DECL_NODE(weights);
   PATTERN_DECL_NODE(bias);
   PATTERN_DECL_NODE(output);

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

@@ -17,6 +17,7 @@
 #include <utility>
 #include <vector>
 #include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/platform/errors.h"
 #include "paddle/fluid/string/pretty_log.h"
 
 namespace paddle {
@@ -43,6 +44,13 @@ void CPUQuantizePass::QuantizeInput(Graph* g, Node* op, Node* input,
                                     std::string input_name, double scale_to_one,
                                     bool is_unsigned,
                                     std::string scale_attr_name) const {
+  auto inputs = op->Op()->InputNames();
+  bool name_found =
+      std::find(inputs.begin(), inputs.end(), input_name) != inputs.end();
+  PADDLE_ENFORCE_EQ(
+      name_found, true,
+      platform::errors::InvalidArgument("%s isn't the input of the %s operator",
+                                        input_name, op->Op()->Type()));
   unsigned max = is_unsigned ? U8_MAX : S8_MAX;
   float scale = scale_to_one * max;
 
@@ -122,6 +130,13 @@ 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 {
+  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(
+                        "%s isn't the output of the %s operator", output_name,
+                        op->Op()->Type()));
   unsigned max = is_unsigned ? U8_MAX : S8_MAX;
   float scale = scale_to_one * max;
 
@@ -228,6 +243,66 @@ void CPUQuantizePass::QuantizeConv(Graph* graph,
   PrettyLogDetail(msg_ss.str().c_str());
 }
 
+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);
+    auto* fc_op_desc = fc->Op();
+
+    // skip if should not be quantized
+    if (fc_op_desc->GetAttrIfExists<bool>("use_quantizer") != true ||
+        fc_op_desc->GetAttrIfExists<bool>("use_mkldnn") != true)
+      return;
+
+    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);
+
+    // get scales calculated after warmup, they scale variables to MAX=1.0
+    auto scales = Get<VarQuantScale>("quant_var_scales");
+
+    auto input_scale = scales[input->Name()].second.data<double>()[0];
+    bool is_input_unsigned = scales[input->Name()].first;
+    QuantizeInput(g, fc, input, "Input", input_scale, is_input_unsigned,
+                  "Scale_in");
+
+    auto weight_scale_tensor = scales[weights->Name()].second;
+    EigenVectorArrayMap eigen_tensor{weight_scale_tensor.data<double>(),
+                                     weight_scale_tensor.numel(), 1};
+    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);
+
+    auto output_scale = scales[output->Name()].second.data<double>()[0];
+    bool is_output_unsigned = scales[output->Name()].first;
+    DequantizeOutput(g, fc, output, "Out", output_scale, is_output_unsigned,
+                     "Scale_out");
+
+    ++quantize_fc_count;
+  };
+
+  gpd(graph, handler);
+  AddStatis(quantize_fc_count);
+
+  std::stringstream msg_ss;
+  msg_ss << "---    quantized " << quantize_fc_count << " fc ops";
+  PrettyLogDetail(msg_ss.str().c_str());
+}
+
 void CPUQuantizePass::QuantizePool(Graph* graph) const {
   GraphPatternDetector gpd;
   auto pattern = gpd.mutable_pattern();
@@ -418,6 +493,7 @@ void CPUQuantizePass::ApplyImpl(ir::Graph* graph) const {
   QuantizeConcat(graph);
   QuantizePriorBox(graph);
   QuantizeTranspose(graph);
+  QuantizeFc(graph);
 }
 
 }  // namespace ir

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

@@ -46,6 +46,8 @@ class CPUQuantizePass : public FusePassBase {
 
   void QuantizeConv(Graph* graph, bool with_residual_data = false) const;
 
+  void QuantizeFc(Graph* graph) const;
+
   void QuantizePool(Graph* graph) const;
 
   void QuantizeConcat(Graph* graph) const;

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

@@ -62,6 +62,10 @@ void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
     if (inputs.size() > 1) op->SetInput("W", {inputs[1]});
     if (inputs.size() > 2) op->SetInput("Bias", {inputs[2]});
     op->SetOutput("Out", {outputs[0]});
+    op->SetAttr("use_quantizer", use_quantizer);
+    op->SetAttr("Scale_in", 1.0f);
+    op->SetAttr("Scale_out", 1.0f);
+    op->SetAttr("Scale_weights", std::vector<float>{1.0f});
   } else if (type == "concat") {
     op->SetInput("X", inputs);
     op->SetOutput("Out", outputs);
@@ -71,13 +75,13 @@ void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
 
 namespace {
 static const std::initializer_list<std::string> variable_names{
-    "a", "w1", "c",  "d", "w2", "e",  "f", "g",
-    "h", "w3", "b1", "i", "j",  "w4", "b2"};
+    "a",  "w1", "c", "d", "w2", "e",  "f",  "g", "h",
+    "w3", "b1", "i", "j", "w4", "b2", "w5", "b3"};
 // (a,w1)->Conv1->c and c->Pool1->d
 //
 // (d,w2)->Conv2->e and e->Pool2->f
 //
-// d->Dropout1->g and g->Fc1->h and (h,w3,b1,i)->Conv3->j
+// d->Dropout1->g and (g, w5, b3)->Fc1->h and (h,w3,b1,i)->Conv3->j
 //
 // (d,w4, b2)->Conv4->i
 ProgramDesc BuildProgramDesc(bool use_mkldnn, bool use_quantizer) {
@@ -98,7 +102,8 @@ ProgramDesc BuildProgramDesc(bool use_mkldnn, bool use_quantizer) {
   SetOp(&prog, "pool2d", "Pool2", {"e"}, {"f"}, use_mkldnn, use_quantizer);
 
   SetOp(&prog, "dropout", "Dropout1", {"d"}, {"g"}, use_mkldnn);
-  SetOp(&prog, "fc", "Fc1", {"g"}, {"h"}, use_mkldnn);
+  SetOp(&prog, "fc", "Fc1", {"g", "w5", "b3"}, {"h"}, use_mkldnn,
+        use_quantizer);
   SetOp(&prog, "conv2d", "Conv3", {"h", "w3", "b1", "i"}, {"j"}, use_mkldnn,
         use_quantizer);
 
@@ -194,13 +199,13 @@ TEST(CpuQuantizePass, quantize) {
   // (d->QUANT3->IN3,w2)->Conv2->OUT3->DEQUANT3->e and
   // e->QUANT4->IN4->Pool2->OUT4->DEQUANT4->f
   //
-  // d->Dropout1->g and g->Fc1->h and
+  // d->Dropout1->g and (g->QUANT8->IN8,w5,b3)->Fc1->OUT7->DEQUANT7->h and
   // (h->QUANT5->IN5,w3,b1,i->QUANT6->IN6)->Conv3->OUT5->DEQUANT5->j
   //
   // (d->QUANT7->IN7,w4, b2)->Conv4->DEQUANT6->OUT6->i
-  // Insert nodes: 7 Quant + 7 IN + 6 OUT + 6 DEQUANT
-  int added_nodes = 7 + 7 + 6 + 6;
-  MainTest(BuildProgramDesc(use_mkldnn, use_quantizer), 4, 2, 7, 6, added_nodes,
+  // Insert nodes: 8 Quant + 8 IN + 7 OUT + 7 DEQUANT
+  int added_nodes = 8 + 8 + 7 + 7;
+  MainTest(BuildProgramDesc(use_mkldnn, use_quantizer), 4, 2, 8, 7, added_nodes,
            2.0f * 127);
 }
 

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

@@ -26,12 +26,11 @@ namespace framework {
 namespace ir {
 
 void FCMKLDNNPass::ApplyImpl(ir::Graph* graph) const {
-  PADDLE_ENFORCE(graph);
+  PADDLE_ENFORCE_NOT_NULL(graph,
+                          platform::errors::InvalidArgument(
+                              "Pointer to graph argument should not be NULL."));
   Init("fc_mkldnn_pass", graph);
 
-  auto* scope = param_scope();
-  PADDLE_ENFORCE(scope);
-
   GraphPatternDetector gpd;
   auto* x = gpd.mutable_pattern()
                 ->NewNode("fc_mkldnn_pass/x")
@@ -49,18 +48,25 @@ void FCMKLDNNPass::ApplyImpl(ir::Graph* graph) const {
       return;
     }
     GET_IR_NODE_FROM_SUBGRAPH(fc, fc, fc_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(input, input, fc_pattern);
     GET_IR_NODE_FROM_SUBGRAPH(weights, weights, fc_pattern);
     GET_IR_NODE_FROM_SUBGRAPH(bias, bias, fc_pattern);
     GET_IR_NODE_FROM_SUBGRAPH(output, output, fc_pattern);
 
     OpDesc* desc = fc->Op();
-    auto in_size = fc->inputs[0]->Var()->GetShape().size();
-    if (in_size != 2 && in_size != 4) {
+    auto dims = fc->inputs[0]->Var()->GetShape();
+    auto dim_num = dims.size();
+    bool are_dims_supported = dim_num == 2 || dim_num == 4;
+    constexpr size_t height_axis = 2;
+    constexpr size_t width_axis = 3;
+    bool is_size_supported =
+        dim_num == 4 ? (dims[width_axis] == 1 && dims[height_axis] == 1) : true;
+    if (!are_dims_supported || !is_size_supported) {
       VLOG(3) << "Do not enable FC MKL-DNN for dimensions different than 2 & 4";
+      VLOG(3) << "Or when width and height are different than one";
       return;
     }
     desc->SetAttr("use_mkldnn", true);
-    PADDLE_ENFORCE(subgraph.count(x));
 
     found_fc_count++;
   };

paddle/fluid/inference/api/analysis_predictor_tester.cc

@@ -276,7 +276,7 @@ class MkldnnQuantizerTest : public testing::Test {
 
   std::pair<bool, framework::LoDTensor> GetMaxChScalingFactor(
       const framework::LoDTensor& var_tensor, bool is_unsigned) const {
-    return mkldnn_quantizer->GetMaxChScalingFactor(var_tensor, is_unsigned);
+    return mkldnn_quantizer->GetMaxChScalingFactor(var_tensor, is_unsigned, 0);
   }
 
   std::pair<bool, framework::LoDTensor> GetKLScalingFactor(

paddle/fluid/inference/api/mkldnn_quantizer.cc

@@ -37,6 +37,11 @@ using framework::LoDTensor;
 using framework::ir::Graph;
 using ConstEigenVectorArrayMap =
     Eigen::Map<const Eigen::Array<float, Eigen::Dynamic, 1>>;
+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>;
 using string::PrettyLogH1;
 static LoDTensor CreateScaleTensor(int64_t channels_num = 1);
 
@@ -66,7 +71,7 @@ bool AnalysisPredictor::MkldnnQuantizer::CalculateScales() {
             bool is_unsigned = false;
             bool compute_scale = true;
             if (is_output) {
-              if (op->Type() == "conv2d") {
+              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");
@@ -138,7 +143,12 @@ void AnalysisPredictor::MkldnnQuantizer::CalculateSingleScale(
       scales_[var_name] = GetMaxScalingFactor(var_tensor, is_unsigned);
       break;
     case ScaleAlgo::MAX_CH:
-      scales_[var_name] = GetMaxChScalingFactor(var_tensor, is_unsigned);
+      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);
       break;
     case ScaleAlgo::KL:
       scales_[var_name] = GetKLScalingFactor(var_tensor, is_unsigned);
@@ -319,7 +329,7 @@ AnalysisPredictor::MkldnnQuantizer::GetMaxScalingFactor(
 
 std::pair<bool, LoDTensor>
 AnalysisPredictor::MkldnnQuantizer::GetMaxChScalingFactor(
-    const LoDTensor& var_tensor, bool is_unsigned) const {
+    const LoDTensor& var_tensor, bool is_unsigned, bool is_transposed) const {
   PADDLE_ENFORCE(var_tensor.dims().size() > 0, "Tensor dimension is empty.");
 
   ConstEigenVectorArrayMap eigen_tensor{var_tensor.data<float>(),
@@ -331,18 +341,23 @@ AnalysisPredictor::MkldnnQuantizer::GetMaxChScalingFactor(
         "Tensor is claimed to be unsigned, but its min value (%f) is < 0.0",
         min_val);
 
-  int channels = var_tensor.dims()[0];
-  LoDTensor scale_tensor = CreateScaleTensor(channels);
-  auto* scale_ptr = scale_tensor.mutable_data<double>(CPUPlace());
-
-  for (int i = 0; i < channels; ++i) {
-    const auto tensor = var_tensor.Slice(i, i + 1);
+  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]};
 
-    ConstEigenVectorArrayMap eigen_tensor{tensor.data<float>(), tensor.numel(),
-                                          1};
-    float max_abs = eigen_tensor.abs().maxCoeff();
-    scale_ptr[i] = 1.0 / max_abs;
+  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();
   }
+  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);
 
   return std::make_pair(is_unsigned, scale_tensor);
 }

paddle/fluid/inference/api/mkldnn_quantizer.h

@@ -79,7 +79,8 @@ class AnalysisPredictor::MkldnnQuantizer {
       const framework::LoDTensor& var_tensor, bool is_unsigned) const;
 
   std::pair<bool, framework::LoDTensor> GetMaxChScalingFactor(
-      const framework::LoDTensor& var_tensor, bool is_unsigned) const;
+      const framework::LoDTensor& var_tensor, bool is_unsigned,
+      bool is_transposed) const;
 
   std::pair<bool, framework::LoDTensor> GetMaxScalingFactor(
       const framework::LoDTensor& var_tensor, bool is_unsigned) const;

paddle/fluid/inference/api/mkldnn_quantizer_config.cc

@@ -37,6 +37,11 @@ MkldnnQuantizerConfig::MkldnnQuantizerConfig() {
 
   rules_["transpose2"]["X"] = ScaleAlgo::KL;
   rules_["transpose2"]["Out"] = ScaleAlgo::NONE;
+
+  rules_["fc"]["Input"] = ScaleAlgo::KL;
+  rules_["fc"]["W"] = ScaleAlgo::MAX_CH_T;
+  rules_["fc"]["Bias"] = ScaleAlgo::NONE;
+  rules_["fc"]["Out"] = ScaleAlgo::KL;
 }
 
 ScaleAlgo MkldnnQuantizerConfig::scale_algo(

paddle/fluid/inference/api/paddle_mkldnn_quantizer_config.h

@@ -27,8 +27,10 @@ namespace paddle {
 // Algorithms for finding scale of quantized Tensors.
 enum class ScaleAlgo {
   NONE,      // Do not compute scale
-  MAX,     // Find scale based on the maximum absolute value
-  MAX_CH,  // Find scale based on the maximum absolute value per channel
+  MAX,       // Find scale based on the max absolute value
+  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
 };
 

paddle/fluid/operators/fc_op.cc

@@ -93,13 +93,21 @@ class FCOp : public framework::OperatorWithKernel {
       const framework::ExecutionContext& ctx) const override {
     framework::LibraryType library = framework::LibraryType::kPlain;
     framework::DataLayout layout = framework::DataLayout::kAnyLayout;
+    int customized_type_value =
+        framework::OpKernelType::kDefaultCustomizedTypeValue;
+    auto input_data_type =
+        OperatorWithKernel::IndicateVarDataType(ctx, "Input");
     if (ctx.Attr<bool>("use_mkldnn")) {
       library = framework::LibraryType::kMKLDNN;
       layout = framework::DataLayout::kMKLDNN;
+      using framework::proto::VarType;
+      customized_type_value = (input_data_type == VarType::INT8 ||
+                               input_data_type == VarType::UINT8)
+                                  ? kFCMKLDNNINT8
+                                  : kFCMKLDNNFP32;
     }
-    return framework::OpKernelType(
-        OperatorWithKernel::IndicateVarDataType(ctx, "Input"), ctx.GetPlace(),
-        layout, library);
+    return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout,
+                                   library, customized_type_value);
   }
 };
 
@@ -132,6 +140,27 @@ class FCOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<bool>(framework::kAllKernelsMustComputeRuntimeShape,
                   "Skip calling InferShape() function in the runtime.")
         .SetDefault(true);
+    /* int8 parameters */
+    AddAttr<bool>("use_quantizer",
+                  "(bool, default false) "
+                  "Set to true for operators that should be quantized and use "
+                  "int8 kernel. "
+                  "Only used on CPU.")
+        .SetDefault(false);
+    AddAttr<float>("Scale_in",
+                   "(float, default 1.0f), The quantize scale of input data")
+        .SetDefault(1.0f);
+    AddAttr<std::vector<float>>("Scale_weights",
+                                "(std::vector<float>, default {1.0f}), The "
+                                "quantize scale of weights data")
+        .SetDefault({1.0f});
+    AddAttr<float>("Scale_out",
+                   "(float, default 1.0f), The quantize scale of output data")
+        .SetDefault(1.0f);
+    AddAttr<bool>("force_fp32_output",
+                  "(bool, default false) Force INT8 kernel output FP32, only "
+                  "used in MKL-DNN INT8")
+        .SetDefault(false);
     AddComment(R"DOC(
 Fully Connected Operator.
 

paddle/fluid/operators/fc_op.h

@@ -21,6 +21,7 @@ limitations under the License. */
 
 namespace paddle {
 namespace operators {
+enum { kFCMKLDNNFP32 = 1, kFCMKLDNNINT8 = 2 };
 
 using Tensor = framework::Tensor;
 

paddle/fluid/operators/mkldnn/dequantize_mkldnn_op.cc

@@ -78,7 +78,6 @@ class DeQuantOpKernel : public framework::OpKernel<T> {
       auto dst_md = platform::MKLDNNMemDesc(
           {dst_tz}, memory::data_type::f32,
           platform::MKLDNNFormatForSize(dst_tz.size(), memory::format::nchw));
-
       auto dst_pd = mkldnn::memory::primitive_desc(dst_md, engine);
       dst_memory = std::make_shared<mkldnn::memory>(
           dst_pd, to_void_cast<float>(output_data));

python/paddle/fluid/tests/unittests/mkldnn/test_fc_mkldnn_op.py

@@ -42,6 +42,7 @@ class TestFCMKLDNNOp(OpTest):
 
     def setUp(self):
         self.op_type = "fc"
+        self._cpu_only = True
         self.use_mkldnn = True
         self.create_data()
         self.inputs = {'Input': self.matrix.input, 'W': self.matrix.weights}