merge

.flake8

@@ -37,9 +37,3 @@ per-file-ignores =
     .cmake-format.py: F821
     python/paddle/fluid/tests/unittests/dygraph_to_static/test_loop.py: F821
     python/paddle/fluid/tests/unittests/dygraph_to_static/test_closure_analysis.py: F821
-    # These files will be fixed in the future
-    python/paddle/fluid/tests/unittests/fft/test_fft_with_static_graph.py: F811
-    python/paddle/fluid/tests/unittests/test_activation_nn_grad.py: F811
-    python/paddle/fluid/tests/unittests/test_lstm_cudnn_op.py: F811
-    python/paddle/fluid/tests/unittests/test_matmul_v2_op.py: F811
-    python/paddle/fluid/tests/unittests/test_rrelu_op.py: F811

cmake/external/cutlass.cmake

@@ -17,7 +17,7 @@ include(ExternalProject)
 set(CUTLASS_PREFIX_DIR ${THIRD_PARTY_PATH}/cutlass)
 
 set(CUTLASS_REPOSITORY https://github.com/NVIDIA/cutlass.git)
-set(CUTLASS_TAG v2.9.1)
+set(CUTLASS_TAG v2.10.0)
 
 include_directories("${THIRD_PARTY_PATH}/cutlass/src/extern_cutlass/")
 include_directories("${THIRD_PARTY_PATH}/cutlass/src/extern_cutlass/include/")

cmake/external/dgc.cmake

@@ -23,14 +23,14 @@ set(DGC_INCLUDE_DIR
 set(DGC_LIBRARIES
     "${DGC_INSTALL_DIR}/lib/libdgc.a"
     CACHE FILEPATH "dgc library." FORCE)
-set(DGC_URL "https://fleet.bj.bcebos.com/dgc/collective_f66ef73.tgz")
+set(DGC_URL "https://fleet.bj.bcebos.com/dgc/collective_7369ff.tgz")
 include_directories(${DGC_INCLUDE_DIR})
 
 ExternalProject_Add(
   extern_dgc
   ${EXTERNAL_PROJECT_LOG_ARGS}
   URL ${DGC_URL}
-  URL_MD5 "94e6fa1bc97169d0e1aad44570fe3251"
+  URL_MD5 "ede459281a0f979da8d84f81287369ff"
   PREFIX "${DGC_PREFIX_DIR}"
   CONFIGURE_COMMAND ""
   BUILD_COMMAND make -j${NPROC}

paddle/fluid/framework/ir/graph_helper.cc

@@ -713,7 +713,7 @@ static void GetGraphOpDesc(const std::vector<Node *> &nodes,
         UpdateControlOpSkipEagerDeletionVars(*n, graph, graph_idx, n->Name());
       }
       ops->emplace_back(*n->Op());
-      VLOG(4) << n->ToString();
+      VLOG(5) << n->ToString();
     }
     // delete no OpDesc op
   }

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -2068,8 +2068,9 @@ PDNode *patterns::Flatten2Matmul::operator()() {
   return matmul_out;
 }
 
-PDNode *patterns::ConvResidual::operator()(bool with_residual_data) {
-  auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d");
+PDNode *patterns::ConvResidual::operator()(const std::string &conv_type,
+                                           bool with_residual_data) {
+  auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op(conv_type);
 
   if (!with_residual_data) {
     conv_op->assert_more([&](Node *x) {
@@ -2082,22 +2083,22 @@ PDNode *patterns::ConvResidual::operator()(bool with_residual_data) {
 
   auto input_var = pattern->NewNode(conv_input_repr())
                        ->AsInput()
-                       ->assert_is_op_input("conv2d", "Input");
+                       ->assert_is_op_input(conv_type, "Input");
 
   auto filter_var = pattern->NewNode(conv_filter_repr())
                         ->AsInput()
-                        ->assert_is_op_input("conv2d", "Filter");
+                        ->assert_is_op_input(conv_type, "Filter");
 
   auto output_var = pattern->NewNode(conv_output_repr())
                         ->AsOutput()
-                        ->assert_is_op_output("conv2d", "Output");
+                        ->assert_is_op_output(conv_type, "Output");
 
   std::vector<PDNode *> links_from{input_var, filter_var};
 
   if (with_residual_data) {
     auto res_conn_var = pattern->NewNode(conv_residual_data_repr())
                             ->AsInput()
-                            ->assert_is_op_input("conv2d", "ResidualData");
+                            ->assert_is_op_input(conv_type, "ResidualData");
     links_from.push_back(res_conn_var);
   }
 

paddle/fluid/framework/ir/graph_pattern_detector.h

@@ -1057,7 +1057,7 @@ struct ConvResidual : public PatternBase {
   ConvResidual(PDPattern* pattern, const std::string& name_scope)
       : PatternBase(pattern, name_scope, "conv_residual") {}
 
-  PDNode* operator()(bool with_residual_data);
+  PDNode* operator()(const std::string& conv_type, bool with_residual_data);
 
   PATTERN_DECL_NODE(conv_op);
   PATTERN_DECL_NODE(conv_input);

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

@@ -319,7 +319,7 @@ void ComputePropagateScalesMkldnnPass::ComputeWeightScales(
     ir::Graph* graph, Scope* scope, StringPairMap* var_quant_scales) const {
   ComputeVarScales(graph,
                    scope,
-                   {"conv2d", "depthwise_conv2d"},
+                   {"conv2d", "depthwise_conv2d", "fused_conv2d"},
                    "Filter",
                    1,
                    var_quant_scales);
@@ -446,7 +446,7 @@ void ComputePropagateScalesMkldnnPass::UpdateReluOutputScales(
     if (op->Type() == "relu") {
       is_unsigned = true;
     } else {
-      if (op->Type() == "conv2d") {
+      if (op->Type() == "conv2d" || op->Type() == "fused_conv2d") {
         act_name = "fuse_activation";
         output_name = "Output";
       } else if (op->Type() == "fc") {

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

@@ -26,7 +26,7 @@ using string::PrettyLogDetail;
 
 void ConvActivationMkldnnFusePass::ApplyImpl(Graph* graph) const {
   auto act_types = phi::funcs::GetSupportedActivations();
-  std::vector<std::string> conv_types = {"conv2d"};
+  std::vector<std::string> conv_types = {"conv2d", "fused_conv2d"};
 
   for (auto& act_type : act_types) {
     FuseConvConcatAct(graph, act_type);
@@ -218,6 +218,45 @@ ConvActivationMkldnnFusePass::ConvActivationMkldnnFusePass() {
       .IsStringIn({"NCHW", "NHWC", "AnyLayout"})
       .End();
 
+  AddOpCompat(OpCompat("fused_conv2d"))
+      .AddInput("Input")
+      .IsTensor()
+      .End()
+      .AddInput("Filter")
+      .IsTensor()
+      .End()
+      .AddInput("Bias")
+      .IsOptional()
+      .IsTensor()
+      .End()
+      .AddInput("ResidualData")
+      .IsTensor()
+      .IsOptional()
+      .End()
+      .AddOutput("Output")
+      .IsTensor()
+      .End()
+      .AddAttr("strides")
+      .IsType<std::vector<int>>()
+      .End()
+      .AddAttr("paddings")
+      .IsType<std::vector<int>>()
+      .End()
+      .AddAttr("padding_algorithm")
+      .IsOptional()
+      .IsStringIn({"EXPLICIT", "SAME", "VALID"})
+      .End()
+      .AddAttr("groups")
+      .IsNumGE(1)
+      .End()
+      .AddAttr("dilations")
+      .IsType<std::vector<int>>()
+      .End()
+      .AddAttr("data_format")
+      .IsOptional()
+      .IsStringIn({"NCHW", "NHWC", "AnyLayout"})
+      .End();
+
   AddOpCompat(OpCompat("concat"))
       .AddInput("X")
       .End()

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

@@ -61,6 +61,40 @@ ConvBiasFusePass::ConvBiasFusePass() {
       .IsStringIn({"NCHW", "NHWC", "AnyLayout"})
       .End();
 
+  AddOpCompat(OpCompat("fused_conv2d"))
+      .AddInput("Input")
+      .IsTensor()
+      .End()
+      .AddInput("Filter")
+      .IsTensor()
+      .End()
+      .AddInput("Bias")
+      .IsTensor()
+      .IsOptional()
+      .End()
+      .AddOutput("Output")
+      .IsTensor()
+      .End()
+      .AddAttr("strides")
+      .IsType<std::vector<int>>()
+      .End()
+      .AddAttr("paddings")
+      .IsType<std::vector<int>>()
+      .End()
+      .AddAttr("padding_algorithm")
+      .IsOptional()
+      .IsStringIn({"EXPLICIT", "SAME", "VALID"})
+      .End()
+      .AddAttr("groups")
+      .IsNumGE(1)
+      .End()
+      .AddAttr("dilations")
+      .IsType<std::vector<int>>()
+      .End()
+      .AddAttr("data_format")
+      .IsStringIn({"NCHW", "NHWC", "AnyLayout"})
+      .End();
+
   AddOpCompat(OpCompat("elementwise_add"))
       .AddInput("X")
       .IsTensor()
@@ -165,6 +199,40 @@ Conv3DBiasFusePass::Conv3DBiasFusePass() {
       .IsStringIn({"NDHWC", "NCDHW"})
       .End();
 
+  AddOpCompat(OpCompat("fused_conv3d"))
+      .AddInput("Input")
+      .IsTensor()
+      .End()
+      .AddInput("Filter")
+      .IsTensor()
+      .End()
+      .AddInput("Bias")
+      .IsTensor()
+      .IsOptional()
+      .End()
+      .AddOutput("Output")
+      .IsTensor()
+      .End()
+      .AddAttr("strides")
+      .IsType<std::vector<int>>()
+      .End()
+      .AddAttr("paddings")
+      .IsType<std::vector<int>>()
+      .End()
+      .AddAttr("padding_algorithm")
+      .IsOptional()
+      .IsStringIn({"EXPLICIT", "SAME", "VALID"})
+      .End()
+      .AddAttr("groups")
+      .IsNumGE(1)
+      .End()
+      .AddAttr("dilations")
+      .IsType<std::vector<int>>()
+      .End()
+      .AddAttr("data_format")
+      .IsStringIn({"NCHW", "NHWC", "AnyLayout"})
+      .End();
+
   AddOpCompat(OpCompat("elementwise_add"))
       .AddInput("X")
       .IsTensor()
@@ -203,6 +271,16 @@ phi::DenseTensor tensor_apply_eltwise(const phi::DenseTensor& vec_a,
 }
 
 void ConvBiasFusePass::ApplyImpl(ir::Graph* graph) const {
+  FuseConvBias(graph, type(), fused_type());
+  if (type() != fused_type()) {
+    // Is the second pass useful?
+    FuseConvBias(graph, fused_type(), fused_type());
+  }
+}
+
+void ConvBiasFusePass::FuseConvBias(ir::Graph* graph,
+                                    const std::string& conv_type,
+                                    const std::string& fused_conv) const {
   PADDLE_ENFORCE_NOT_NULL(
       graph, platform::errors::InvalidArgument("Graph cannot be nullptr."));
   FusePassBase::Init(name_scope_, graph);
@@ -216,9 +294,9 @@ void ConvBiasFusePass::ApplyImpl(ir::Graph* graph) const {
       gpd.mutable_pattern()
           ->NewNode(patterns::PDNodeName(name_scope_, "conv_input"))
           ->AsInput()
-          ->assert_is_op_input(type(), "Input");
+          ->assert_is_op_input(conv_type, "Input");
   patterns::ConvBias conv_bias_pattern(gpd.mutable_pattern(), name_scope_);
-  conv_bias_pattern(conv_input, type());
+  conv_bias_pattern(conv_input, conv_type);
   int found_conv_bias_count = 0;
   auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
                      Graph* g) {
@@ -249,7 +327,7 @@ void ConvBiasFusePass::ApplyImpl(ir::Graph* graph) const {
     // check if fuse can be done and if MKL-DNN should be used
     FuseOptions fuse_option = FindFuseOption(*conv, *eltwise);
     if (fuse_option == DO_NOT_FUSE || fuse_option == FUSE_NATIVE) {
-      VLOG(3) << "do not perform " + type() + "+bias fuse";
+      VLOG(3) << "do not perform " + conv_type + "+bias fuse";
       return;
     }
 
@@ -294,7 +372,7 @@ void ConvBiasFusePass::ApplyImpl(ir::Graph* graph) const {
       desc.SetInput("Filter", std::vector<std::string>({conv_weight->Name()}));
       desc.SetInput("Bias", std::vector<std::string>({eltwise_bias->Name()}));
       desc.SetOutput("Output", std::vector<std::string>({eltwise_out->Name()}));
-      desc.SetType(type());
+      desc.SetType(fused_conv);
 
       for (auto& attr : conv->Op()->GetAttrMap()) {
         desc.SetAttr(attr.first, attr.second);
@@ -323,6 +401,7 @@ void ConvBiasFusePass::ApplyImpl(ir::Graph* graph) const {
                             type());
   }
 }
+
 }  // namespace ir
 }  // namespace framework
 }  // namespace paddle

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

@@ -32,24 +32,32 @@ class ConvBiasFusePass : public FusePassBase {
   ConvBiasFusePass();
   virtual ~ConvBiasFusePass() {}
   virtual std::string type() const { return "conv2d"; }
+  virtual std::string fused_type() const { return "fused_conv2d"; }
 
  protected:
   void ApplyImpl(ir::Graph* graph) const override;
+  void FuseConvBias(ir::Graph* graph,
+                    const std::string& conv_type,
+                    const std::string& fused_conv) const;
+
   const std::string name_scope_{"conv_bias_mkldnn_fuse"};
 };
+
 /*
  * Fuse the Conv3D and Elementwise_add to a Conv3DBiasOp.
  */
 class Conv2DTransposeBiasFusePass : public ConvBiasFusePass {
  public:
   Conv2DTransposeBiasFusePass();
   std::string type() const override { return "conv2d_transpose"; }
+  std::string fused_type() const override { return "conv2d_transpose"; }
 };
 
 class Conv3DBiasFusePass : public ConvBiasFusePass {
  public:
   Conv3DBiasFusePass();
   std::string type() const override { return "conv3d"; }
+  std::string fused_type() const override { return "fused_conv3d"; }
 };
 }  // namespace ir
 }  // namespace framework

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

@@ -139,7 +139,8 @@ void MainTest(bool convWithExistingBias) {
   int conv_bias_count = 0;
 
   for (auto* node : graph->Nodes()) {
-    if (node->IsOp() && node->Op()->Type() == "conv2d") {
+    if (node->IsOp() && (node->Op()->Type() == "conv2d" ||
+                         node->Op()->Type() == "fused_conv2d")) {
       auto* op = node->Op();
       ASSERT_TRUE(op->HasAttr("use_mkldnn"));
       EXPECT_TRUE(PADDLE_GET_CONST(bool, op->GetAttr("use_mkldnn")));

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

@@ -388,11 +388,12 @@ void CPUQuantizePass::GetQuantInfo(Graph* graph) const {
 }
 
 void CPUQuantizePass::QuantizeConv(Graph* graph,
+                                   const std::string& conv_type,
                                    bool with_residual_data) const {
   GraphPatternDetector gpd;
   auto pattern = gpd.mutable_pattern();
   patterns::ConvResidual conv_pattern{pattern, name_scope_};
-  conv_pattern(with_residual_data);
+  conv_pattern(conv_type, with_residual_data);
 
   int quantize_conv_count = 0;
   auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
@@ -510,7 +511,7 @@ void CPUQuantizePass::QuantizeConv(Graph* graph,
   AddStatis(quantize_conv_count);
 
   LogQuantizedOpsCounter(
-      "conv2d",
+      conv_type,
       quantize_conv_count,
       ((with_residual_data) ? "with residual connection" : ""));
 }
@@ -1247,8 +1248,10 @@ void CPUQuantizePass::ApplyImpl(ir::Graph* graph) const {
       platform::errors::InvalidArgument("Scope cannot be nullptr."));
 
   GetQuantInfo(graph);
-  QuantizeConv(graph, false /* with_residual_data */);
-  QuantizeConv(graph, true /* with_residual_data */);
+  QuantizeConv(graph, "conv2d", false /* with_residual_data */);
+  QuantizeConv(graph, "conv2d", true /* with_residual_data */);
+  QuantizeConv(graph, "fused_conv2d", false /* with_residual_data */);
+  QuantizeConv(graph, "fused_conv2d", true /* with_residual_data */);
   QuantizePool(graph);
   QuantizeConcat(graph);
   QuantizePriorBox(graph);

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

@@ -49,7 +49,9 @@ class CPUQuantizePass : public FusePassBase {
  protected:
   void ApplyImpl(ir::Graph* graph) const override;
 
-  void QuantizeConv(Graph* graph, bool with_residual_data) const;
+  void QuantizeConv(Graph* graph,
+                    const std::string& conv_type,
+                    bool with_residual_data) const;
   void QuantizeFc(Graph* graph, bool with_residual_data) const;
   void QuantizePool(Graph* graph) const;
   void QuantizeConcat(Graph* graph) const;