Unification of BF16 enablement process (#31034)

* Unification of bfloat16 enablement process and refactor

* Remove unnecessary function

* Standardize the output name search

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -1829,9 +1829,8 @@ PDNode *patterns::OpDequant::operator()() {
   auto any_op = pattern->NewNode(any_op_repr())
                     ->assert_is_op()
                     ->assert_more([&](Node *node) {
-                      return (node->Op()->Type() == "matmul" ||
-                              node->Op()->Type() == "conv2d" ||
-                              node->Op()->Type() == "fc");
+                      return (node->Op()->HasAttr("force_fp32_output") ||
+                              node->Op()->HasProtoAttr("force_fp32_output"));
                     });
   auto dequant_in = pattern->NewNode(dequant_in_repr())
                         ->assert_is_op_input("dequantize", "Input");
@@ -1865,6 +1864,44 @@ PDNode *patterns::DequantScale::operator()() {
   return scale_out;
 }
 
+PDNode *patterns::ScaleQuant::operator()() {
+  auto scale_in = pattern->NewNode(scale_in_repr())
+                      ->AsInput()
+                      ->assert_is_op_input("scale", "X");
+  auto scale_op = pattern->NewNode(scale_op_repr())->assert_is_op("scale");
+
+  auto quant_in = pattern->NewNode(quant_in_repr())
+                      ->AsInput()
+                      ->assert_is_op_input("quantize", "Input");
+  auto quant_op = pattern->NewNode(quant_op_repr())->assert_is_op("quantize");
+
+  scale_op->LinksFrom({scale_in}).LinksTo({quant_in});
+  quant_op->LinksFrom({quant_in});
+
+  return quant_op;
+}
+
+PDNode *patterns::QuantConv::operator()() {
+  auto quant_in = pattern->NewNode(quant_in_repr())
+                      ->AsInput()
+                      ->assert_is_op_input("quantize", "Input");
+  auto quant_op = pattern->NewNode(quant_op_repr())->assert_is_op("quantize");
+
+  auto conv_in = pattern->NewNode(conv_in_repr())
+                     ->AsInput()
+                     ->assert_is_op_input("conv2d", "Input");
+  auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d");
+  conv_op->assert_more([&](Node *node) {
+    return node->Op()->GetAttrIfExists<std::string>("mkldnn_data_type") ==
+           "bfloat16";
+  });
+
+  quant_op->LinksFrom({quant_in}).LinksTo({conv_in});
+  conv_op->LinksFrom({conv_in});
+
+  return quant_op;
+}
+
 PDNode *patterns::ScaleMatmul::operator()() {
   auto scale_in = pattern->NewNode(scale_in_repr())
                       ->AsInput()
@@ -2191,10 +2228,11 @@ PDNode *patterns::QuantizePlacement::operator()(
 PDNode *patterns::Bfloat16Placement::operator()(
     const std::unordered_set<std::string> &bfloat16_enabled_op_types) {
   std::unordered_set<std::string> supported_op_types =
-      std::unordered_set<std::string>(
-          {"concat", "conv2d", "conv2d_transpose", "elementwise_add",
-           "elementwise_mul", "fc", "fusion_gru", "gelu", "layer_norm",
-           "matmul", "pool2d", "reshape2", "softmax", "sum", "transpose2"});
+      std::unordered_set<std::string>({"concat", "conv2d", "conv2d_transpose",
+                                       "elementwise_add", "elementwise_mul",
+                                       "fc", "fusion_gru", "gelu", "layer_norm",
+                                       "matmul", "pool2d", "relu", "reshape2",
+                                       "softmax", "sum", "transpose2"});
   if (!bfloat16_enabled_op_types.empty()) {
     supported_op_types = bfloat16_enabled_op_types;
   }
@@ -2240,25 +2278,12 @@ PDNode *patterns::LastBfloat16Ops::operator()() {
            "bfloat16";
   });
   auto *op_out = pattern->NewNode(op_out_repr())->AsOutput();
-
-  auto *next_op = pattern->NewNode(next_op_repr())->assert_is_op();
-  next_op->assert_more([&](Node *node) {
-    return node->Op()->GetAttrIfExists<std::string>("mkldnn_data_type") !=
-           "bfloat16";
-  });
-
   op->LinksTo({op_out});
-  next_op->LinksFrom({op_out});
-  return next_op;
+  return op_out;
 }
 
 PDNode *patterns::FirstBfloat16Ops::operator()() {
-  auto *prev_op = pattern->NewNode(prev_op_repr())->assert_is_op();
-  prev_op->assert_more([&](Node *node) {
-    return node->Op()->GetAttrIfExists<std::string>("mkldnn_data_type") !=
-           "bfloat16";
-  });
-  auto *op_in = pattern->NewNode(op_in_repr())->AsOutput();
+  auto *op_in = pattern->NewNode(op_in_repr())->AsInput();
 
   auto *op = pattern->NewNode(op_repr())->assert_is_op();
   op->assert_more([&](Node *node) {
@@ -2266,7 +2291,6 @@ PDNode *patterns::FirstBfloat16Ops::operator()() {
            "bfloat16";
   });
 
-  prev_op->LinksTo({op_in});
   op->LinksFrom({op_in});
   return op;
 }
@@ -2280,27 +2304,6 @@ PDNode *patterns::DuplicatedInputs::operator()() {
   return op;
 }
 
-PDNode *patterns::UnnecessaryReorders::operator()() {
-  auto prev_op = pattern->NewNode(prev_op_repr())->assert_is_op();
-  prev_op->assert_more([&](Node *node) {
-    return node->Op()->GetAttrIfExists<std::string>("mkldnn_data_type") ==
-           "bfloat16";
-  });
-
-  auto *quant_in = pattern->NewNode(quant_in_repr())
-                       ->assert_is_op_input("quantize", "Input");
-
-  auto *quant_op = pattern->NewNode(quant_op_repr())->assert_is_op("quantize");
-
-  auto *quant_out = pattern->NewNode(quant_out_repr())
-                        ->assert_is_op_output("quantize", "Output");
-
-  prev_op->LinksTo({quant_in});
-  quant_op->LinksFrom({quant_in}).LinksTo({quant_out});
-
-  return quant_out;
-}
-
 PDNode *patterns::MKLDNNInPlace::operator()() {
   const std::unordered_set<std::string> &supported_op_types = {
       "abs",

paddle/fluid/framework/ir/graph_pattern_detector.h

@@ -1135,11 +1135,36 @@ struct DequantScale : public PatternBase {
 
   PATTERN_DECL_NODE(dequant_op);
   PATTERN_DECL_NODE(dequant_out);
-
   PATTERN_DECL_NODE(scale_op);
   PATTERN_DECL_NODE(scale_out);
 };
 
+// Scale + Quantize
+struct ScaleQuant : public PatternBase {
+  ScaleQuant(PDPattern* pattern, const std::string& name_scope)
+      : PatternBase(pattern, name_scope, "scale_quant") {}
+
+  PDNode* operator()();
+
+  PATTERN_DECL_NODE(scale_in);
+  PATTERN_DECL_NODE(scale_op);
+  PATTERN_DECL_NODE(quant_in);
+  PATTERN_DECL_NODE(quant_op);
+};
+
+// Quantize + Conv2d
+struct QuantConv : public PatternBase {
+  QuantConv(PDPattern* pattern, const std::string& name_scope)
+      : PatternBase(pattern, name_scope, "quant_conv") {}
+
+  PDNode* operator()();
+
+  PATTERN_DECL_NODE(quant_in);
+  PATTERN_DECL_NODE(quant_op);
+  PATTERN_DECL_NODE(conv_in);
+  PATTERN_DECL_NODE(conv_op);
+};
+
 // Scale + Matmul
 struct ScaleMatmul : public PatternBase {
   ScaleMatmul(PDPattern* pattern, const std::string& name_scope)
@@ -1338,7 +1363,6 @@ struct LastBfloat16Ops : public PatternBase {
 
   PATTERN_DECL_NODE(op);
   PATTERN_DECL_NODE(op_out);
-  PATTERN_DECL_NODE(next_op);
 };
 
 struct FirstBfloat16Ops : public PatternBase {
@@ -1346,7 +1370,6 @@ struct FirstBfloat16Ops : public PatternBase {
       : PatternBase(pattern, name_scope, "first_bfloat16_ops") {}
   PDNode* operator()();
 
-  PATTERN_DECL_NODE(prev_op);
   PATTERN_DECL_NODE(op_in);
   PATTERN_DECL_NODE(op);
 };
@@ -1360,17 +1383,6 @@ struct DuplicatedInputs : public PatternBase {
   PATTERN_DECL_NODE(op);
 };
 
-struct UnnecessaryReorders : public PatternBase {
-  UnnecessaryReorders(PDPattern* pattern, const std::string& name_scope)
-      : PatternBase(pattern, name_scope, "unnecessary_reorders") {}
-  PDNode* operator()();
-
-  PATTERN_DECL_NODE(prev_op);
-  PATTERN_DECL_NODE(quant_in);
-  PATTERN_DECL_NODE(quant_op);
-  PATTERN_DECL_NODE(quant_out);
-};
-
 // Pattern used for enforcing inplace computation for in-place computation
 // supporting DNNL ops. softmax, batch_norm and layer_norm
 struct MKLDNNInPlace : public PatternBase {

paddle/fluid/framework/ir/mkldnn/cpu_bfloat16_pass.cc

@@ -12,12 +12,10 @@ limitations under the License. */
 #include "paddle/fluid/framework/ir/mkldnn/cpu_bfloat16_pass.h"
 
 #include <string>
-#include <unordered_set>
 #include <vector>
 
 #include "paddle/fluid/framework/ir/graph_pattern_detector.h"
 #include "paddle/fluid/framework/op_version_registry.h"
-#include "paddle/fluid/platform/mkldnn_helper.h"
 #include "paddle/fluid/string/pretty_log.h"
 
 namespace paddle {
@@ -33,8 +31,38 @@ void UnlinkNodes(ir::Node* a, ir::Node* b) {
                   b->inputs.end());
 }
 
+// Checking whether a reorder from FP32 to BF16 should be added before the input
+// to the operator
+bool IsPermittedInputName(const std::string& input_name) {
+  // Only the inputs listed in \"permitted_names\" requires quanitization before
+  // the bfloat16 operator. Other inputs, such as Filter and Bias are reordered
+  // in the kernel.
+  const std::vector<std::string> permitted_names = {"X", "Y", "Input",
+                                                    "ResidualData"};
+  return (std::find(permitted_names.begin(), permitted_names.end(),
+                    input_name) != permitted_names.end());
+}
+
+// Checking whether a reorder from BF16 to FP32 should be added after the output
+// to the operator
+bool IsPermittedOutputName(const std::string& output_name) {
+  // XShape is output in transpose2 and reshape2 operators used to store the
+  // shape and lod of X. So this output do not need dequantize before.
+  return (output_name != "XShape");
+}
+
 void AddQuantize(Graph* g, ir::Node* op, ir::Node* op_in,
                  int* quantize_counter) {
+  std::vector<std::string> input_names;
+
+  // Find the name of the input linking op to op_in
+  for (auto name : op->Op()->InputNames())
+    for (auto input_name : op->Op()->Input(name))
+      if (input_name == op_in->Name() && IsPermittedInputName(name))
+        input_names.push_back(name);
+
+  if (input_names.empty()) return;
+
   VarDesc quantize_out_desc(patterns::PDNodeName("quantize", "out"));
   auto* quantize_out_node = g->CreateVarNode(&quantize_out_desc);
 
@@ -44,23 +72,12 @@ void AddQuantize(Graph* g, ir::Node* op, ir::Node* op_in,
   q_desc.SetOutput("Output",
                    std::vector<std::string>({quantize_out_node->Name()}));
   q_desc.SetAttr("Scale", 1.f);
+  q_desc.SetAttr("Shift", 0.0f);
   q_desc.SetAttr("bfloat16", true);
   q_desc.SetAttr("output_format", op->Op()->HasAttr("data_layout")
                                       ? op->Op()->GetAttr("data_layout")
                                       : std::string("NCHW"));
-  auto quantize_op = g->CreateOpNode(&q_desc);
-
-  std::vector<std::string> input_names;
-  for (auto name : op->Op()->InputNames()) {
-    for (auto input_name : op->Op()->Input(name)) {
-      if (input_name == op_in->Name()) input_names.push_back(name);
-    }
-  }
-
-  PADDLE_ENFORCE_NE(
-      input_names.empty(), true,
-      platform::errors::NotFound(
-          "Operator before operator should have input as op output"));
+  auto quantize_op = g->CreateOpNode(&q_desc);  // OpDesc will be copied.
 
   for (auto name = input_names.begin(); name < input_names.end(); name++)
     op->Op()->SetInput(*name,
@@ -99,11 +116,12 @@ void AddQuantizes(Graph* g, ir::Node* op, int* quantize_counter) {
     q_desc.SetOutput("Output",
                      std::vector<std::string>({quantize_out_node_names[i]}));
     q_desc.SetAttr("Scale", 1.f);
+    q_desc.SetAttr("Shift", 0.0f);
     q_desc.SetAttr("bfloat16", true);
     q_desc.SetAttr("output_format", op->Op()->HasAttr("data_layout")
                                         ? op->Op()->GetAttr("data_layout")
                                         : std::string("NCHW"));
-    auto quantize_op = g->CreateOpNode(&q_desc);
+    auto quantize_op = g->CreateOpNode(&q_desc);  // OpDesc will be copied.
 
     UnlinkNodes(inputs[i], op);
     IR_NODE_LINK_TO(inputs[i], quantize_op);
@@ -115,6 +133,9 @@ void AddQuantizes(Graph* g, ir::Node* op, int* quantize_counter) {
   op->Op()->SetInput("X", quantize_out_node_names);
 }
 
+// Operators like Concat and Sum have a single input name X, which actually
+// consists of multiple inputs. Such operators require a different way to find
+// pattern and add quantize ops.
 void AddReoderBeforeDuplicatedInputs(ir::Graph* graph, int* quantize_counter) {
   GraphPatternDetector gpd;
   patterns::DuplicatedInputs duplicated_inputs{gpd.mutable_pattern(),
@@ -128,38 +149,8 @@ void AddReoderBeforeDuplicatedInputs(ir::Graph* graph, int* quantize_counter) {
   gpd(graph, handler);
 }
 
-void RemoveUnnecessaryReorders(ir::Graph* graph, int* quantize_counter) {
-  GraphPatternDetector gpd;
-  patterns::UnnecessaryReorders unnecessary_reorders{gpd.mutable_pattern(),
-                                                     "unnecessary_reorders"};
-  unnecessary_reorders();
-  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
-                     Graph* g) {
-    GET_IR_NODE_FROM_SUBGRAPH(prev_op, prev_op, unnecessary_reorders);
-    GET_IR_NODE_FROM_SUBGRAPH(quant_in, quant_in, unnecessary_reorders);
-    GET_IR_NODE_FROM_SUBGRAPH(quant_op, quant_op, unnecessary_reorders);
-    GET_IR_NODE_FROM_SUBGRAPH(quant_out, quant_out, unnecessary_reorders);
-
-    std::string op_output_name;
-    for (auto name : prev_op->Op()->OutputNames())
-      for (auto output_name : prev_op->Op()->Output(name))
-        if (output_name == quant_in->Name()) op_output_name = name;
-
-    PADDLE_ENFORCE_NE(
-        op_output_name.empty(), true,
-        platform::errors::NotFound(
-            "Operator before operator should have input as op output"));
-
-    prev_op->Op()->SetOutput(op_output_name,
-                             std::vector<std::string>({quant_out->Name()}));
-
-    IR_NODE_LINK_TO(prev_op, quant_out);
-    GraphSafeRemoveNodes(graph, {quant_in, quant_op});
-    (*quantize_counter)--;
-  };
-  gpd(graph, handler);
-}
-
+// Adding quantize ops before all operators except Concat and Sum, which have
+// already been handled in AddReoderBeforeDuplicatedInputs
 void AddReoderBeforeSingleInputs(ir::Graph* graph, int* quantize_counter) {
   GraphPatternDetector gpd;
   patterns::FirstBfloat16Ops bfloat16_ops{gpd.mutable_pattern(),
@@ -167,12 +158,9 @@ void AddReoderBeforeSingleInputs(ir::Graph* graph, int* quantize_counter) {
   bfloat16_ops();
   auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                      Graph* g) {
-    GET_IR_NODE_FROM_SUBGRAPH(prev_op, prev_op, bfloat16_ops);
     GET_IR_NODE_FROM_SUBGRAPH(op_in, op_in, bfloat16_ops);
     GET_IR_NODE_FROM_SUBGRAPH(op, op, bfloat16_ops);
-    auto prev_op_type = prev_op->Op()->Type();
-    if (op->Op()->Type() != "conv2d" && prev_op_type != "quantize" &&
-        prev_op_type != "sum" && prev_op_type != "concat") {
+    if (op->Op()->Type() != "sum" && op->Op()->Type() != "concat") {
       AddQuantize(g, op, op_in, quantize_counter);
     }
   };
@@ -182,9 +170,8 @@ void AddReoderBeforeSingleInputs(ir::Graph* graph, int* quantize_counter) {
 void CPUBFloat16Pass::SetInputDataType(ir::Graph* graph) const {
   int quantize_counter = 0;
   AddReoderBeforeDuplicatedInputs(graph, &quantize_counter);
-  RemoveUnnecessaryReorders(graph, &quantize_counter);
   AddReoderBeforeSingleInputs(graph, &quantize_counter);
-  PrettyLogDetail("---    added %d quantize op before bfloat16 op",
+  PrettyLogDetail("---    added %d quantize ops before bfloat16 op",
                   quantize_counter);
 }
 
@@ -193,55 +180,51 @@ void CPUBFloat16Pass::SetOutputDataType(ir::Graph* graph) const {
   patterns::LastBfloat16Ops bfloat16_ops{gpd.mutable_pattern(),
                                          "last_bfloat16_ops"};
   bfloat16_ops();
-  int force_fp32_counter = 0, dequantize_counter = 0;
+  int dequantize_counter = 0;
 
   auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                      Graph* g) {
     GET_IR_NODE_FROM_SUBGRAPH(op, op, bfloat16_ops);
     GET_IR_NODE_FROM_SUBGRAPH(op_out, op_out, bfloat16_ops);
-    GET_IR_NODE_FROM_SUBGRAPH(next_op, next_op, bfloat16_ops);
-    if ((op->Op()->HasAttr("force_fp32_output") ||
-         op->Op()->HasProtoAttr("force_fp32_output")) &&
-        !op->Op()->GetAttrIfExists<bool>("fuse_residual_connection")) {
-      op->Op()->SetAttr("force_fp32_output", true);
-      force_fp32_counter++;
-    } else if (op->Op()->Type() != "prior_box") {
-      VarDesc dequantize_out_desc(patterns::PDNodeName("dequantize", "out"));
-      auto* dequantize_out_node = g->CreateVarNode(&dequantize_out_desc);
+
+    if (op->Op()->Type() != "prior_box") {
+      // Find the name of the output linking op to op_out
+      std::vector<std::string> output_names;
+      for (auto name : op->Op()->OutputNames())
+        for (auto output_name : op->Op()->Output(name))
+          if (output_name == op_out->Name() && IsPermittedOutputName(name))
+            output_names.push_back(name);
+
+      if (output_names.empty()) return;
+
+      VarDesc dequantize_in_desc(patterns::PDNodeName("dequantize", "in"));
+      auto* dequantize_in_node = g->CreateVarNode(&dequantize_in_desc);
 
       OpDesc deq_desc;
       deq_desc.SetType("dequantize");
-      deq_desc.SetInput("Input", std::vector<std::string>({op_out->Name()}));
-      deq_desc.SetOutput(
-          "Output", std::vector<std::string>({dequantize_out_node->Name()}));
+      deq_desc.SetInput("Input",
+                        std::vector<std::string>({dequantize_in_node->Name()}));
+      deq_desc.SetOutput("Output", std::vector<std::string>({op_out->Name()}));
       deq_desc.SetAttr("Scale", 1.0f);
-      auto dequantize_op = g->CreateOpNode(&deq_desc);
+      deq_desc.SetAttr("Shift", 0.0f);
+      auto dequantize_op =
+          g->CreateOpNode(&deq_desc);  // OpDesc will be copied.
 
-      std::string next_op_input_name;
-      for (auto name : next_op->Op()->InputNames()) {
-        for (auto input_name : next_op->Op()->Input(name)) {
-          if (input_name == op_out->Name()) next_op_input_name = name;
-        }
-      }
+      for (auto name = output_names.begin(); name < output_names.end(); name++)
+        op->Op()->SetOutput(
+            *name, std::vector<std::string>({dequantize_in_node->Name()}));
+
+      UnlinkNodes(op, op_out);
+      IR_NODE_LINK_TO(op, dequantize_in_node);
+      IR_NODE_LINK_TO(dequantize_in_node, dequantize_op);
+      IR_NODE_LINK_TO(dequantize_op, op_out);
 
-      PADDLE_ENFORCE_NE(
-          next_op_input_name.empty(), true,
-          platform::errors::NotFound(
-              "Operator before operator should have input as op output"));
-
-      next_op->Op()->SetInput(
-          next_op_input_name,
-          std::vector<std::string>({dequantize_out_node->Name()}));
-      UnlinkNodes(op_out, next_op);
-      IR_NODE_LINK_TO(op_out, dequantize_op);
-      IR_NODE_LINK_TO(dequantize_op, dequantize_out_node);
-      IR_NODE_LINK_TO(dequantize_out_node, next_op);
       dequantize_counter++;
     }
   };
   gpd(graph, handler);
-  PrettyLogDetail("---    added %d dequantize op and used %d force_fp32_output",
-                  dequantize_counter, force_fp32_counter);
+  PrettyLogDetail("---    added %d dequantize ops after bfloat16 op",
+                  dequantize_counter);
 }
 
 void CPUBFloat16Pass::ApplyImpl(ir::Graph* graph) const {

paddle/fluid/framework/ir/mkldnn/cpu_bfloat16_pass_tester.cc

@@ -26,8 +26,7 @@ namespace ir {
 void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
            const std::vector<std::string>& inputs,
            const std::vector<std::string>& outputs, bool use_mkldnn,
-           const std::string& mkldnn_data_type = "float32",
-           const bool force_fp32_output = false) {
+           const std::string& mkldnn_data_type = "float32") {
   auto* op = prog->MutableBlock(0)->AppendOp();
   op->SetType(type);
   op->SetAttr("use_mkldnn", use_mkldnn);
@@ -37,7 +36,6 @@ void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
     op->SetInput("Input", {inputs[0]});
     op->SetOutput("Output", {outputs[0]});
     op->SetAttr("mkldnn_data_type", mkldnn_data_type);
-    op->SetAttr("force_fp32_output", force_fp32_output);
   } else if (type == "pool2d" || type == "transpose2" || type == "reshape2" ||
              type == "dropout") {
     op->SetInput("X", {inputs[0]});
@@ -47,7 +45,6 @@ void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
     op->SetInput("Input", {inputs[0]});
     op->SetOutput("Out", {outputs[0]});
     op->SetAttr("mkldnn_data_type", mkldnn_data_type);
-    op->SetAttr("force_fp32_output", force_fp32_output);
   } else if (type == "concat" || type == "sum") {
     op->SetInput("X", inputs);
     op->SetOutput("Out", outputs);
@@ -58,7 +55,6 @@ void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
     if (inputs.size() > 1) op->SetInput("Y", {inputs[1]});
     op->SetOutput("Out", {outputs[0]});
     op->SetAttr("mkldnn_data_type", mkldnn_data_type);
-    if (type == "matmul") op->SetAttr("force_fp32_output", force_fp32_output);
   } else if (type == "layer_norm") {
     op->SetInput("X", {inputs[0]});
     op->SetOutput("Y", {outputs[0]});
@@ -79,8 +75,8 @@ void PreparePass(std::unique_ptr<ir::Graph>* graph, const ProgramDesc& prog,
   *current_nodes_num = (*graph)->Nodes().size();
 }
 
-void MainTest(const ProgramDesc& prog, int quant_count, int dequant_count,
-              int force_fp32_count, int added_nodes_count) {
+void MainTest(const ProgramDesc& prog, const int& quant_count,
+              const int& dequant_count, const int& added_nodes_count) {
   std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
   int original_nodes_num, current_nodes_num;
   PreparePass(&graph, prog, variable_names, &original_nodes_num,
@@ -88,7 +84,6 @@ void MainTest(const ProgramDesc& prog, int quant_count, int dequant_count,
 
   int quantize_nodes_count = 0;
   int dequantize_nodes_count = 0;
-  int force_fp32_nodes_count = 0;
   for (auto* node : graph->Nodes()) {
     if (node->IsOp()) {
       auto* op = node->Op();
@@ -96,16 +91,11 @@ void MainTest(const ProgramDesc& prog, int quant_count, int dequant_count,
         quantize_nodes_count++;
       } else if (op->Type() == "dequantize") {
         dequantize_nodes_count++;
-      } else if (op->Type() == "conv2d" || op->Type() == "matmul" ||
-                 op->Type() == "fc") {
-        if (op->GetAttrIfExists<bool>("force_fp32_output"))
-          force_fp32_nodes_count++;
       }
     }
   }
   EXPECT_EQ(quantize_nodes_count, quant_count);
   EXPECT_EQ(dequantize_nodes_count, dequant_count);
-  EXPECT_EQ(force_fp32_nodes_count, force_fp32_count);
   EXPECT_EQ(original_nodes_num + added_nodes_count, current_nodes_num);
 }
 
@@ -125,9 +115,10 @@ ProgramDesc BuildProgramDescConv(bool use_mkldnn) {
 
 TEST(CpuBfloat16Pass, convolution) {
   bool use_mkldnn = true;
-  // 0 added + 1 force_fp32_output
-  int added_nodes = 0;
-  MainTest(BuildProgramDescConv(use_mkldnn), 0, 0, 1, added_nodes);
+  int quant_op = 3;
+  int dequant_op = 3;
+  int added_nodes = quant_op * 2 + dequant_op * 2;
+  MainTest(BuildProgramDescConv(use_mkldnn), quant_op, dequant_op, added_nodes);
 }
 
 ProgramDesc BuildProgramDescDoubleInput(bool use_mkldnn) {
@@ -147,9 +138,11 @@ ProgramDesc BuildProgramDescDoubleInput(bool use_mkldnn) {
 
 TEST(CpuBfloat16Pass, double_input_ops) {
   bool use_mkldnn = true;
-  // 2 quant + 2 quant out
-  int added_nodes = 4;
-  MainTest(BuildProgramDescDoubleInput(use_mkldnn), 2, 0, 0, added_nodes);
+  int quant_op = 4;
+  int dequant_op = 3;
+  int added_nodes = quant_op * 2 + dequant_op * 2;
+  MainTest(BuildProgramDescDoubleInput(use_mkldnn), quant_op, dequant_op,
+           added_nodes);
 }
 
 ProgramDesc BuildProgramDescDuplicatedInput(bool use_mkldnn) {
@@ -169,9 +162,11 @@ ProgramDesc BuildProgramDescDuplicatedInput(bool use_mkldnn) {
 
 TEST(CpuBfloat16Pass, duplicated_input_ops) {
   bool use_mkldnn = true;
-  // 3 quant + 3 quant out
-  int added_nodes = 6;
-  MainTest(BuildProgramDescDuplicatedInput(use_mkldnn), 3, 0, 0, added_nodes);
+  int quant_op = 5;
+  int dequant_op = 3;
+  int added_nodes = quant_op * 2 + dequant_op * 2;
+  MainTest(BuildProgramDescDuplicatedInput(use_mkldnn), quant_op, dequant_op,
+           added_nodes);
 }
 
 ProgramDesc BuildProgramDescDoubleOutputs(bool use_mkldnn) {
@@ -193,9 +188,11 @@ ProgramDesc BuildProgramDescDoubleOutputs(bool use_mkldnn) {
 
 TEST(CpuBfloat16Pass, double_outputs_ops) {
   bool use_mkldnn = true;
-  // 3 dequant + 3 dequant out
-  int added_nodes = 6;
-  MainTest(BuildProgramDescDoubleOutputs(use_mkldnn), 0, 3, 0, added_nodes);
+  int quant_op = 3;
+  int dequant_op = 3;
+  int added_nodes = quant_op * 2 + dequant_op * 2;
+  MainTest(BuildProgramDescDoubleOutputs(use_mkldnn), quant_op, dequant_op,
+           added_nodes);
 }
 
 }  // namespace ir

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

@@ -255,14 +255,21 @@ void CPUQuantizeSquashPass::OpDequantSquash(Graph* graph) const {
     GET_IR_NODE_FROM_SUBGRAPH(dequant_out, dequant_out, op_dequant_pattern);
 
     if (dequant_in->outputs.size() == 1) {
-      auto output_name = "Out";
-      if (any_op->Op()->Type() == "conv2d") {
+      if (any_op->Op()->Type() == "conv2d" ||
+          any_op->Op()->Type() == "conv2d_transpose") {
         // do not squash if fuse residual connection is true
         // because residual fusion does not support force output with fp32
         if (any_op->Op()->GetAttrIfExists<bool>("fuse_residual_connection"))
           return;
-        output_name = "Output";
       }
+      // Find the name of the output linking any_op to dequant_in
+      std::string output_name;
+      for (auto name : any_op->Op()->OutputNames())
+        for (auto out_name : any_op->Op()->Output(name))
+          if (out_name == dequant_in->Name()) output_name = name;
+
+      if (output_name.empty()) return;
+
       any_op->Op()->SetAttr("force_fp32_output", true);
       any_op->Op()->SetOutput(output_name,
                               std::vector<std::string>({dequant_out->Name()}));
@@ -363,10 +370,10 @@ void CPUQuantizeSquashPass::DequantScaleSquash(Graph* graph) const {
                         platform::errors::InvalidArgument(
                             "Dequantize scale(%f) should have positive value.",
                             dequant_scale));
-      PADDLE_ENFORCE_GT(scale_scale, 0.0f,
+      PADDLE_ENFORCE_NE(
+          scale_scale, 0.0f,
           platform::errors::InvalidArgument(
-                            "Scale(%f) of scale op should have positive value.",
-                            scale_scale));
+              "Scale(%f) should have a non-zero value", scale_scale));
 
       dequant_op->Op()->SetAttr("Scale", dequant_scale / scale_scale);
       dequant_op->Op()->SetOutput(
@@ -378,10 +385,86 @@ void CPUQuantizeSquashPass::DequantScaleSquash(Graph* graph) const {
   };
   gpd(graph, handler);
   AddStatis(found_dequant_scale_squash_count);
-  PrettyLogDetail("---    squashed %d scale with dequant",
+  PrettyLogDetail("---    squashed %d scale with dequantize op",
                   found_dequant_scale_squash_count);
 }
 
+// squash scale with quantize
+void CPUQuantizeSquashPass::ScaleQuantSquash(Graph* graph) const {
+  GraphPatternDetector gpd;
+  patterns::ScaleQuant scale_quant_pattern{gpd.mutable_pattern(),
+                                           "scale_quant"};
+  scale_quant_pattern();
+
+  int found_scale_quant_squash_count = 0;
+  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
+                     Graph* g) {
+    VLOG(4) << "squash scale-quant ops pair";
+
+    GET_IR_NODE_FROM_SUBGRAPH(scale_in, scale_in, scale_quant_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(scale_op, scale_op, scale_quant_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(quant_in, quant_in, scale_quant_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(quant_op, quant_op, scale_quant_pattern);
+
+    if (quant_in->outputs.size() == 1 &&
+        scale_op->Op()->GetAttrIfExists<float>("bias") == 0.0) {
+      auto quant_scale = quant_op->Op()->GetAttrIfExists<float>("Scale");
+      auto scale_scale = scale_op->Op()->GetAttrIfExists<float>("scale");
+
+      PADDLE_ENFORCE_GT(
+          quant_scale, 0.0f,
+          platform::errors::InvalidArgument(
+              "Quantize scale(%f) should have positive value.", quant_scale));
+      PADDLE_ENFORCE_NE(
+          scale_scale, 0.0f,
+          platform::errors::InvalidArgument(
+              "Scale(%f) should have a non-zero value", scale_scale));
+
+      quant_op->Op()->SetAttr("Scale", quant_scale * scale_scale);
+      quant_op->Op()->SetInput("Input",
+                               std::vector<std::string>({scale_in->Name()}));
+      IR_NODE_LINK_TO(scale_in, quant_op);
+      GraphSafeRemoveNodes(graph, {scale_op, quant_in});
+      found_scale_quant_squash_count++;
+    }
+  };
+  gpd(graph, handler);
+  AddStatis(found_scale_quant_squash_count);
+  PrettyLogDetail("---    squashed %d scale with quantize op",
+                  found_scale_quant_squash_count);
+}
+
+// squash quantize if is before bfloat16 conv2d
+void CPUQuantizeSquashPass::QuantizeBf16Conv(Graph* graph) const {
+  GraphPatternDetector gpd;
+  patterns::QuantConv pattern{gpd.mutable_pattern(), "quant_conv"};
+  pattern();
+
+  int found_quant_conv_squash_count = 0;
+  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
+                     Graph* g) {
+    VLOG(4) << "squash quant-conv2d ops pair";
+
+    GET_IR_NODE_FROM_SUBGRAPH(quant_in, quant_in, pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(quant_op, quant_op, pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(conv_in, conv_in, pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(conv_op, conv_op, pattern);
+
+    if (conv_in->outputs.size() == 1 &&
+        quant_op->Op()->GetAttrIfExists<float>("Scale") == 1.0) {
+      conv_op->Op()->SetInput("Input",
+                              std::vector<std::string>({quant_in->Name()}));
+      IR_NODE_LINK_TO(quant_in, conv_op);
+      GraphSafeRemoveNodes(graph, {quant_op, conv_in});
+      found_quant_conv_squash_count++;
+    }
+  };
+  gpd(graph, handler);
+  AddStatis(found_quant_conv_squash_count);
+  PrettyLogDetail("---    squashed %d quantize with bfloat16 conv2d op",
+                  found_quant_conv_squash_count);
+}
+
 void CPUQuantizeSquashPass::ApplyImpl(ir::Graph* graph) const {
   PADDLE_ENFORCE_NOT_NULL(
       graph,
@@ -389,6 +472,8 @@ void CPUQuantizeSquashPass::ApplyImpl(ir::Graph* graph) const {
           "The graph in function CPUQuantizeSquashPass::ApplyImpl is null."));
   FusePassBase::Init("cpu_quantize_squash_pass", graph);
 
+  DequantScaleSquash(graph);
+  ScaleQuantSquash(graph);
   std::unordered_map<const Node*, int> nodes_keep_counter;
   FindNodesToKeep(graph, &nodes_keep_counter);
   DequantQuantSquash(graph, &nodes_keep_counter);
@@ -396,7 +481,7 @@ void CPUQuantizeSquashPass::ApplyImpl(ir::Graph* graph) const {
   RequantOpSquash(graph);
   OpDequantSquash(graph);
   MultipleQuantizeSquash(graph);
-  DequantScaleSquash(graph);
+  QuantizeBf16Conv(graph);
 }
 
 }  // namespace ir

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

@@ -78,6 +78,16 @@ class CPUQuantizeSquashPass : public FusePassBase {
    */
   void DequantScaleSquash(Graph* graph) const;
 
+  /*
+   * Squash scale if scale is before quantize
+   */
+  void ScaleQuantSquash(Graph* graph) const;
+
+  /*
+   * Squash quantize if is before bfloat16 conv2d
+   */
+  void QuantizeBf16Conv(Graph* graph) const;
+
   const std::string name_scope_{"squash"};
 };
 

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

@@ -24,7 +24,8 @@ namespace ir {
 void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
            const std::vector<std::string>& inputs,
            const std::vector<std::string>& outputs, bool use_mkldnn,
-           const std::vector<float> scale = {}, float bias = 0.0) {
+           const std::vector<float> scale = {}, float bias = 0.0,
+           const std::string& mkldnn_data_type = "float32") {
   auto* op = prog->MutableBlock(0)->AppendOp();
   op->SetType(type);
   op->SetAttr("use_mkldnn", use_mkldnn);
@@ -36,6 +37,8 @@ void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
     if (inputs.size() > 1) op->SetInput("Filter", {inputs[1]});
     if (inputs.size() > 2) op->SetInput("Bias", {inputs[2]});
     op->SetOutput("Output", {outputs[0]});
+    op->SetAttr("force_fp32_output", false);
+    op->SetAttr("mkldnn_data_type", mkldnn_data_type);
   } else if (type == "quantize") {
     op->SetInput("Input", {inputs[0]});
     op->SetOutput("Output", {outputs[0]});
@@ -52,6 +55,7 @@ void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
   } else if (type == "concat") {
     op->SetInput("X", inputs);
     op->SetOutput("Out", outputs);
+    op->SetAttr("mkldnn_data_type", mkldnn_data_type);
   } else if (type == "fc") {
     op->SetInput("Input", {inputs[0]});
     PADDLE_ENFORCE_EQ(inputs.size(), 2UL,
@@ -63,6 +67,8 @@ void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
     op->SetOutput("Out", outputs);
     if (scale.size() > 0) op->SetAttr("Scale_in", scale[0]);
     if (scale.size() > 1) op->SetAttr("Scale_out", scale[1]);
+    op->SetAttr("force_fp32_output", false);
+    op->SetAttr("mkldnn_data_type", mkldnn_data_type);
   } else if (type == "scale") {
     op->SetInput("X", {inputs[0]});
     op->SetOutput("Out", {outputs[0]});
@@ -74,6 +80,8 @@ void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
     op->SetOutput("Out", {outputs[0]});
     if (scale.size() > 0) op->SetAttr("Scale_x", scale[0]);
     if (scale.size() > 1) op->SetAttr("Scale_out", scale[1]);
+    op->SetAttr("force_fp32_output", false);
+    op->SetAttr("mkldnn_data_type", mkldnn_data_type);
   }
 }
 
@@ -299,6 +307,20 @@ ProgramDesc BuildDequantScaleProgramDesc(bool use_mkldnn, float dequant_scale,
   return prog;
 }
 
+// a->Scale->b
+// b->Quant->c
+ProgramDesc BuildScaleQuantProgramDesc(bool use_mkldnn, float scale_scale,
+                                       float quant_scale, float bias) {
+  ProgramDesc prog;
+  for (auto& v : variable_names) {
+    prog.MutableBlock(0)->Var(v);
+  }
+  SetOp(&prog, "scale", "Scale", {"a"}, {"b"}, use_mkldnn, {scale_scale}, bias);
+  SetOp(&prog, "quantize", "Quant", {"b"}, {"c"}, use_mkldnn, {quant_scale});
+
+  return prog;
+}
+
 // {x,y}->Matmul->b
 // b->Dequant->c
 ProgramDesc BuildMatmulDequantProgramDesc(bool use_mkldnn,
@@ -341,6 +363,22 @@ ProgramDesc BuildRequantOpProgramDesc(bool use_mkldnn, float requant_scale_in,
   return prog;
 }
 
+// a->Quant->b
+// b->Conv2d->c
+ProgramDesc BuildQuantConv2dProgramDesc(const bool& use_mkldnn,
+                                        const float& quant_scale,
+                                        const std::string& mkldnn_data_type) {
+  ProgramDesc prog;
+  for (auto& v : variable_names) {
+    prog.MutableBlock(0)->Var(v);
+  }
+  SetOp(&prog, "quantize", "Quant", {"a"}, {"b"}, use_mkldnn, {quant_scale});
+  SetOp(&prog, "conv2d", "Conv2d", {"b"}, {"c"}, use_mkldnn, {}, 0.0f,
+        mkldnn_data_type);
+
+  return prog;
+}
+
 void InitTensorHolder(Scope* scope, const paddle::platform::Place& place,
                       const char* var_name) {
   auto x = scope->Var(var_name);
@@ -664,6 +702,22 @@ TEST(CpuQuantizeSquashPass, dequantize_scale_with_bias) {
                  "Dequant", "Scale", dequant_scale);
 }
 
+// if scale has no bias
+TEST(CpuQuantizeSquashPass, scale_with_no_bias_quantize) {
+  constexpr auto scale_scale = 1.5432f;
+  constexpr auto quant_scale = 1.2345f;
+  constexpr auto bias = 0.0f;
+  auto use_mkldnn = true;
+  // remove: dequant out, scale op
+  auto remove_nodes = 2;
+  CountNodeTest(
+      BuildScaleQuantProgramDesc(use_mkldnn, scale_scale, quant_scale, bias),
+      remove_nodes);
+  EqualScaleTest(
+      BuildScaleQuantProgramDesc(use_mkldnn, scale_scale, quant_scale, bias),
+      "Scale", "Quant", quant_scale * scale_scale);
+}
+
 TEST(CpuQuantizeSquashPass, matmul_with_dequant) {
   auto dequant_scale = 1.2345f;
   auto use_mkldnn = true;
@@ -688,6 +742,17 @@ TEST(CpuQuantizeSquashPass, requantize_with_matmul_fc_conv) {
   EqualScaleTest(program_desc, "Conv", "Scale_in", requant_scale_in);
 }
 
+TEST(CpuQuantizeSquashPass, quant_bf16_conv2d) {
+  auto quant_scale = 1.0f;
+  auto use_mkldnn = true;
+  auto mkldnn_data_type = "bfloat16";
+  // remove: quant_op, conv_in
+  auto remove_nodes = 2;
+  CountNodeTest(
+      BuildQuantConv2dProgramDesc(use_mkldnn, quant_scale, mkldnn_data_type),
+      remove_nodes);
+}
+
 }  // namespace ir
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -268,6 +268,7 @@ void CpuPassStrategy::EnableMkldnnBfloat16() {
   if (!use_mkldnn_bfloat16_) {
     passes_.push_back("cpu_bfloat16_placement_pass");
     passes_.push_back("cpu_bfloat16_pass");
+    passes_.push_back("cpu_quantize_squash_pass");
   }
   use_mkldnn_bfloat16_ = true;
 #else

paddle/fluid/operators/mkldnn/requantize_mkldnn_op.cc

@@ -156,4 +156,5 @@ class ReQuantOpKernel : public framework::OpKernel<T> {
 namespace ops = paddle::operators;
 
 REGISTER_OP_KERNEL(requantize, MKLDNN, ::paddle::platform::CPUPlace,
-                   ops::ReQuantOpKernel<int8_t>, ops::ReQuantOpKernel<uint8_t>);
+                   ops::ReQuantOpKernel<int8_t>, ops::ReQuantOpKernel<uint8_t>,
+                   ops::ReQuantOpKernel<paddle::platform::bfloat16>);