[DNNL] Added elementwise_add mkl-dnn inplace (#23477) (#24080)

test=release/2.0
Co-authored-by: NJacek Czaja <jacek.czaja@intel.com>

paddle/fluid/framework/ir/CMakeLists.txt

@@ -86,7 +86,7 @@ endif()
 
 if(WITH_MKLDNN)
     pass_library(mkldnn_placement_pass base DEPS placement_pass_base DIR mkldnn)
-    pass_library(mkldnn_inplace_pass inference DEPS mkldnn_placement_pass op_registry softmax_op softmax DIR mkldnn)
+    pass_library(mkldnn_inplace_pass inference DEPS mkldnn_placement_pass op_registry elementwise_add_op activation_op softmax_op softmax DIR mkldnn)
     pass_library(depthwise_conv_mkldnn_pass base DIR mkldnn)
     pass_library(conv_bias_mkldnn_fuse_pass inference DIR mkldnn)
     pass_library(conv_activation_mkldnn_fuse_pass inference DIR mkldnn)

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -1892,30 +1892,30 @@ PDNode *patterns::MultipleQuantize::operator()() {
 }
 
 PDNode *patterns::MKLDNNInPlace::operator()() {
-  // TODO(jczaja): Enable more mkl-dnn ops e.g. activation, elementwise_add,
-  // batch_norm....
   auto possible_inplace_op =
-      pattern->NewNode(inplace_to_be_op_repr())->assert_is_ops({"softmax"});
+      pattern->NewNode(inplace_to_be_op_repr())
+          ->assert_is_ops({"elementwise_add", "softmax"});
 
-  // TODO(jczaja): Enable more mkl-dnn ops e.g. activation, elementwise_add,
-  // batch_norm....
+  // TODO(jczaja): Enable more mkl-dnn ops e.g. activation, batch_norm....
   auto input = pattern->NewNode(inplace_to_be_op_in_repr())
-                   ->assert_is_ops_input({"softmax"})
+                   ->assert_is_ops_input({"elementwise_add", "softmax"})
                    ->AsInput();
-  // TODO(jczaja): Enable more mkl-dnn ops e.g. activation, elementwise_add,
-  // batch_norm....
+  // TODO(jczaja): Enable more mkl-dnn ops e.g. activation, batch_norm....
   auto output = pattern->NewNode(inplace_to_be_op_out_repr())
-                    ->assert_is_ops_output({"softmax"})
-                    ->AsIntermediate();
+                    ->assert_is_ops_output({"elementwise_add", "softmax"})
+                    ->AsOutput();
 
   auto next_op = pattern->NewNode(next_op_repr())->assert_is_op();
+  auto next_output = pattern->NewNode(next_op_out_repr())->AsOutput();
 
   // Check if op is MKL-DNN enabled
   possible_inplace_op->assert_op_attr("use_mkldnn", true);
 
+  // linked structure
   possible_inplace_op->LinksTo({output});
   possible_inplace_op->LinksFrom({input});
   next_op->LinksFrom({output});
+  next_op->LinksTo({next_output});
 
   return possible_inplace_op;
 }

paddle/fluid/framework/ir/graph_pattern_detector.h

@@ -1140,11 +1140,12 @@ struct MKLDNNInPlace : public PatternBase {
       : PatternBase(pattern, name_scope, "mkldnn_inplace") {}
   PDNode* operator()();
 
-  // MKL-DNN's in-place ops: BatchNorm, Softmax, Layer Norm
+  // MKL-DNN's in-place ops: BatchNorm, Softmax, Elementwise_add
   PATTERN_DECL_NODE(inplace_to_be_op);
   PATTERN_DECL_NODE(inplace_to_be_op_in);
   PATTERN_DECL_NODE(inplace_to_be_op_out);
   PATTERN_DECL_NODE(next_op);
+  PATTERN_DECL_NODE(next_op_out);
 };
 
 struct TransposeFlattenConcat : public PatternBase {

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

@@ -16,6 +16,7 @@
 #include <algorithm>
 #include <memory>
 #include <string>
+#include <unordered_map>
 #include <vector>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/lod_tensor.h"
@@ -30,6 +31,7 @@ void MKLDNNInPlacePass::ApplyImpl(ir::Graph* graph) const {
   PADDLE_ENFORCE_NOT_NULL(graph,
                           platform::errors::InvalidArgument(
                               "Pointer to graph argument should not be NULL."));
+  std::unordered_map<std::string, std::string> original_output_names;
   GraphPatternDetector gpd;
   patterns::MKLDNNInPlace mkldnn_inplace{gpd.mutable_pattern(),
                                          "mkldnn_inplace"};
@@ -40,72 +42,136 @@ void MKLDNNInPlacePass::ApplyImpl(ir::Graph* graph) const {
                      Graph* g) {
     VLOG(3) << "Start to handle MKL-DNN In-Place pass";
 
-    GET_IR_NODE_FROM_SUBGRAPH(inplace_to_be_op, inplace_to_be_op,
+    GET_IR_NODE_FROM_SUBGRAPH(current_op, inplace_to_be_op, mkldnn_inplace);
+    GET_IR_NODE_FROM_SUBGRAPH(current_op_in, inplace_to_be_op_in,
                               mkldnn_inplace);
-    GET_IR_NODE_FROM_SUBGRAPH(inplace_to_be_op_in, inplace_to_be_op_in,
-                              mkldnn_inplace);
-    GET_IR_NODE_FROM_SUBGRAPH(inplace_to_be_op_out, inplace_to_be_op_out,
+    GET_IR_NODE_FROM_SUBGRAPH(current_op_out, inplace_to_be_op_out,
                               mkldnn_inplace);
     GET_IR_NODE_FROM_SUBGRAPH(next_op, next_op, mkldnn_inplace);
+    GET_IR_NODE_FROM_SUBGRAPH(next_op_out, next_op_out, mkldnn_inplace);
 
-    if ((inplace_to_be_op->Op()->HasAttr("use_mkldnn") == false) ||
-        (boost::get<bool>(inplace_to_be_op->Op()->GetAttr("use_mkldnn")) ==
-         false)) {
+    if ((current_op->Op()->HasAttr("use_mkldnn") == false) ||
+        (boost::get<bool>(current_op->Op()->GetAttr("use_mkldnn")) == false)) {
       VLOG(3) << "do not perform mkl-dnn inplace: use_mkldnn missing or set to "
                  "false";
       return;
     }
 
-    auto& infer_inplace = OpInfoMap::Instance()
-                              .Get(inplace_to_be_op->Op()->Type())
-                              .infer_inplace_;
+    auto& infer_inplace =
+        OpInfoMap::Instance().Get(current_op->Op()->Type()).infer_inplace_;
     if (!infer_inplace) {
       VLOG(3) << "do not perform mkl-dnn inplace: missing InplaceInferer";
       return;
     }
 
-    // TODO(jczaja): Enable more ops
-    if (inplace_to_be_op->Op()->Type() != "softmax") {
-      VLOG(3)
-          << "Curently works for softmax only. TODO(jczaja): support other ops";
+    VLOG(3) << "DNNL Inplace op(" << current_op->id() << ") "
+            << "Curr Node In: " << current_op_in->Name()
+            << " Curr Node out: " << current_op_out->Name();
+
+    VLOG(3) << "DNNL Inplace next op(" << next_op->id() << ") "
+            << " next Node out: " << next_op_out->Name();
+
+    auto inputs = current_op->Op()->Inputs();
+    auto outputs = current_op->Op()->Outputs();
+    auto in_to_outs = infer_inplace(false);  // strictly no CUDA for MKL-DNN
+    VLOG(3) << "DNNL InplaceInferer op(" << current_op->id() << ") "
+            << in_to_outs.begin()->first << ": "
+            << inputs[in_to_outs.begin()->first][0] << " "
+            << in_to_outs.begin()->second << ": "
+            << outputs[in_to_outs.begin()->second][0];
+    // If InferInplace pattern does not contain input node then skip
+    auto inplace_input_vec = inputs[in_to_outs.begin()->first];
+    if (std::find(inplace_input_vec.begin(), inplace_input_vec.end(),
+                  current_op_in->Name()) == inplace_input_vec.end()) {
+      VLOG(3) << "DNNL in-place pass SKIP pattern ";
       return;
     }
 
-    // Iterate over all nodes  that are ops
-    // and check if in-place to be var is part of inputs
-    // if positive then do not perform inplace
-    for (const Node* n : graph->Nodes()) {
-      if (n->IsOp()) {
-        // Avoid searchin in op that is to be inplace
-        if ((n->id() != inplace_to_be_op->id())) {
-          auto* op = n->Op();
-          auto inputs = op->Inputs();
-          auto in_place_input = inplace_to_be_op_in->Name();
-          for (auto& it : inputs) {
-            for (auto& var_name : it.second) {
-              if (var_name == in_place_input) {
-                VLOG(3) << "MKL-DNN in-place pass: in-place var cannot be an "
-                           "input to more than one operator";
-                return;
-              }
-            }
+    // Checking if this particular node (to be inplaced, overwritten)
+    // is used anywhere else apart from inplaced op
+    auto input_consumers = current_op_in->outputs;
+    if (input_consumers.size() > 1) {
+      VLOG(3) << "DNNL in-place pass FAIL: in-place var cannot "
+                 "be an input to multiple operators";
+      return;
+    }
+
+    // If this op was alrady inplaced in previous pass placements
+    // then we need to update input of next op
+    // but original name to be changed is gone, so we need to remember it
+    // on first time given op is to be inplaced
+    if (current_op_in->Name() != current_op_out->Name()) {
+      original_output_names[current_op->Name() + current_op_in->Name()] =
+          current_op_out->Name();
+    } else {
+      VLOG(3) << "DNNL Inplace: Current op already inplaced! ";
+    }
+
+    // It may be that next op is reusing some of vars, we need to
+    // make sure that unwanted inplace is not created
+    // TODO(jczaja): Make UT for that one
+    for (auto& n : current_op_out->outputs) {
+      auto& n_op_infer_inplace =
+          OpInfoMap::Instance().Get(n->Op()->Type()).infer_inplace_;
+      if ((n_op_infer_inplace == nullptr)) {
+        for (auto& m : n->outputs) {
+          if (m->Name() == current_op_in->Name()) {
+            VLOG(3) << "DNNL in-place pass FAIL: in-place var cannot "
+                       "be an output to non-inplaced next op";
+            return;
           }
         }
       }
     }
 
-    auto original_name = inplace_to_be_op_out->Name();
-    inplace_to_be_op_out->RenameVar(inplace_to_be_op_in->Name());
+    auto original_name =
+        original_output_names[current_op->Name() + current_op_in->Name()];
+    current_op_out->RenameVar(current_op_in->Name());
 
     // Get mapping of input to output
-    auto in_to_outs = infer_inplace(false);  // strictly no CUDA for MKL-DNN
-    // TODO(jczaja): Support more complex situations
     auto out_name = in_to_outs.begin()->second;
-    inplace_to_be_op->Op()->SetOutput(
-        out_name, std::vector<std::string>({inplace_to_be_op_out->Name()}));
-    next_op->Op()->RenameInput(original_name, inplace_to_be_op_out->Name());
+    current_op->Op()->SetOutput(
+        out_name, std::vector<std::string>({current_op_out->Name()}));
+
+    // If next op in a line is doing inplace
+    // then we need to update its output as well
+
+    // Get inferer of next op
+    // If no inferer then we are done
+    auto& next_op_infer_inplace =
+        OpInfoMap::Instance().Get(next_op->Op()->Type()).infer_inplace_;
+    if (next_op_infer_inplace) {
+      auto in_to_outs = next_op_infer_inplace(false);
+      auto out_name = in_to_outs.begin()->second;
+      auto* op = next_op->Op();
+      auto inputs = op->Inputs();
+      auto outputs = op->Outputs();
+      // Check if in-place happened
+      // for variable we changed (original name)
+      // TODO(jczaja): make recursive propagation of inplace
+      auto next_op_inplace_inputs = inputs[in_to_outs.begin()->first];
+      if ((next_op_inplace_inputs == outputs[in_to_outs.begin()->second]) &&
+          (std::find(next_op_inplace_inputs.begin(),
+                     next_op_inplace_inputs.end(),
+                     original_name) != next_op_inplace_inputs.end())) {
+        VLOG(3) << "DNNL InPlace: Next Op is in-placed , updating its "
+                   "input "
+                   "and output var!";
+        next_op->Op()->SetOutput(
+            out_name, std::vector<std::string>({current_op_out->Name()}));
+        next_op_out->RenameVar(current_op_in->Name());
+        // Get ops that next_op_out is linked to and update their input
+        auto next_op_out_consumers = next_op_out->outputs;  // Has to be ops
+        for (auto& c : next_op_out_consumers) {
+          c->Op()->RenameInput(original_name, current_op_out->Name());
+        }
+      }
+    }
+
+    next_op->Op()->RenameInput(original_name, current_op_out->Name());
+
     found_inplace_count++;
-    VLOG(3) << "MKL-DNN InPlace applied!";
+    VLOG(3) << "DNNL InPlace applied!";
   };
 
   gpd(graph, handler);

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

@@ -21,6 +21,9 @@
 
 USE_OP(softmax);
 USE_OP_DEVICE_KERNEL(softmax, MKLDNN);
+USE_OP(elementwise_add);
+USE_OP_DEVICE_KERNEL(elementwise_add, MKLDNN);
+USE_OP(relu);
 
 namespace paddle {
 namespace framework {
@@ -62,8 +65,9 @@ class MKLDNNInplacePassTest {
                                bool branched) {
     ProgramDesc prog;
 
-    for (auto& v : std::vector<std::string>(
-             {"a", "weights", "bias", "f", "g", "h", "i", "j", "k"})) {
+    for (auto& v :
+         std::vector<std::string>({"a", "weights", "bias", "f", "g", "h", "i",
+                                   "j", "k", "l", "m", "z"})) {
       auto* var = prog.MutableBlock(0)->Var(v);
       var->SetType(proto::VarType::SELECTED_ROWS);
       if (v == "weights" || v == "bias") {
@@ -83,9 +87,12 @@ class MKLDNNInplacePassTest {
     SetOp(&prog, "elementwise_add", "elementwise_add1",
           std::vector<std::string>({"h", "i"}), std::vector<std::string>({"j"}),
           mkldnn_enabled_op.compare("elementwise_add") == 0);
+    SetOp(&prog, "relu", "relu2", std::vector<std::string>({"j"}),
+          std::vector<std::string>({"k"}),
+          mkldnn_enabled_op.compare("softmax") == 0);
     if (branched == true) {
       SetOp(&prog, "softmax", "softmax2", std::vector<std::string>({"g"}),
-            std::vector<std::string>({"k"}),
+            std::vector<std::string>({"z"}),
             mkldnn_enabled_op.compare("softmax") == 0);
     }
 
@@ -105,12 +112,11 @@ class MKLDNNInplacePassTest {
     unsigned use_mkldnn_true_count = 0;
     std::unordered_map<std::string, std::string> input_names;
     std::unordered_map<std::string, std::string> output_names;
+
     input_names["softmax"] = "X";
     output_names["softmax"] = "Out";
-    input_names["batch_norm"] = "X";
-    output_names["batch_norm"] = "Y";
-    input_names["layer_norm"] = "X";
-    output_names["layer_norm"] = "Y";
+    input_names["elementwise_add"] = "X";
+    output_names["elementwise_add"] = "Out";
 
     VLOG(3) << DebugString(graph);
 
@@ -135,15 +141,18 @@ class MKLDNNInplacePassTest {
 
 TEST(MKLDNNInplacePass, inplace_softmax) {
   // softmax to be mkl-dnn enabled and made in-place
-
   MKLDNNInplacePassTest().MainTest("softmax", false, 1);
 }
 
 TEST(MKLDNNInplacePass, inplace_softmax_branched) {
-  // softmax to be mkl-dnn enabled and made in-place
+  // softmax's input is shared by two branches. so no in-place
   MKLDNNInplacePassTest().MainTest("softmax", true, 0);
 }
 
+TEST(MKLDNNInplacePass, inplace_elementwise_add) {
+  // Two elementwise_add mkl-dnn enabled op instances to be made inplace
+  MKLDNNInplacePassTest().MainTest("elementwise_add", false, 1);
+}
 }  // namespace ir
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/operators/elementwise/mkldnn/elementwise_add_mkldnn_op.cc

@@ -56,39 +56,34 @@ class EltwiseAddMKLDNNKernel : public framework::OpKernel<T> {
         y->format(), MKLDNNMemoryFormat::undef,
         platform::errors::InvalidArgument("Wrong format set for Y tensor"));
 
-    const T* x_data = x->data<T>();
-    const T* y_data = y->data<T>();
-
     auto src_x_tz = framework::vectorize<int64_t>(x->dims());
     auto src_y_tz = framework::vectorize<int64_t>(y->dims());
     auto dst_tz = framework::vectorize<int64_t>(z->dims());
 
-    std::vector<float> scales = {1.0f, 1.0f};
+    // Currently MKL-DNN kernel supports only Z <- X + Y, shape(X) == shape(Y)
+    // TODO(jczaja): Binary primitive support broadcasting, so we can support
+    // this in kernel
+    platform::BinaryMKLDNNHandler<T> handler(
+        dnnl::algorithm::binary_add, src_x_tz, x->format(), y->format(),
+        dev_ctx, ctx.GetPlace(), ctx.OutputName("Out"));
 
-    const std::string key =
-        platform::CreateKey(src_x_tz, ctx.OutputName("Out"));
+    auto src_x_memory = handler.AcquireSrcMemory(x);
+    auto src_y_memory = handler.AcquireSecondSrcMemory(y);
 
-    platform::SumMKLDNNHandler handler(dev_ctx, mkldnn_engine, key);
+    // For Inplace src and and dst are the same memory object
+    auto dst_memory =
+        x->IsSharedBufferWith(*z) ? src_x_memory : handler.AcquireDstMemory(z);
 
-    auto src_x_memory = handler.AcquireSrcMemory(
-        {{src_x_tz}, platform::MKLDNNGetDataType<T>(), x->format()},
-        paddle::platform::to_void_cast(x_data));
-    auto src_y_memory = handler.AcquireSecondSrcMemory(
-        {{src_y_tz}, platform::MKLDNNGetDataType<T>(), y->format()},
-        paddle::platform::to_void_cast(y_data));
-    auto dst_md = memory::desc({dst_tz}, platform::MKLDNNGetDataType<T>(),
-                               MKLDNNMemoryFormat::any);
-    auto sum_pd = handler.AcquireSumPrimitiveDescriptor(
-        {src_x_memory, src_y_memory}, scales, dst_md);
-    T* z_data =
-        z->mutable_data<T>(ctx.GetPlace(), sum_pd->dst_desc().get_size());
-    auto dst_memory = handler.AcquireDstMemoryFromPrimitive(z_data);
-    auto sum_prim = handler.AcquireSum();
+    auto binary_prim = handler.AcquireForwardPrimitive();
 
     mkldnn::stream astream(mkldnn_engine);
-    sum_prim->execute(astream, {{MKLDNN_ARG_MULTIPLE_SRC, *src_x_memory},
-                                {MKLDNN_ARG_MULTIPLE_SRC + 1, *src_y_memory},
-                                {MKLDNN_ARG_DST, *dst_memory}});
+
+    std::unordered_map<int, dnnl::memory> args = {
+        {DNNL_ARG_SRC_0, *src_x_memory},
+        {DNNL_ARG_SRC_1, *src_y_memory},
+        {DNNL_ARG_DST, *dst_memory}};
+
+    binary_prim->execute(astream, args);
     astream.wait();
 
     z->set_layout(DataLayout::kMKLDNN);

paddle/fluid/operators/mkldnn/inplace_op_tests.cmake

-cc_test(test_mkldnn_op_inplace SRCS mkldnn/test_mkldnn_op_inplace.cc DEPS op_registry softmax_op softmax scope device_context enforce executor)
+cc_test(test_mkldnn_op_inplace SRCS mkldnn/test_mkldnn_op_inplace.cc DEPS op_registry elementwise_add_op softmax_op softmax scope device_context enforce executor)
 

paddle/fluid/operators/mkldnn/softmax_mkldnn_op.cc

@@ -45,7 +45,8 @@ class SoftmaxMKLDNNHandler
       : platform::MKLDNNHandlerT<T, mkldnn::softmax_forward,
                                  mkldnn::softmax_backward>(
             dev_ctx, dev_ctx.GetEngine(), cpu_place,
-            platform::CreateKey(dims, uniq_name)) {
+            // Softmax may be inplace then uniq_name is no longer unique
+            platform::CreateKey(dims, axis, uniq_name)) {
     auto md = mkldnn::memory::desc(dims, platform::MKLDNNGetDataType<T>(), fmt);
 
     this->AcquireForwardPrimitiveDescriptor(prop_kind::forward_scoring, md,
@@ -60,7 +61,7 @@ class SoftmaxMKLDNNHandler
       : platform::MKLDNNHandlerT<T, mkldnn::softmax_forward,
                                  mkldnn::softmax_backward>(
             dev_ctx, dev_ctx.GetEngine(), cpu_place,
-            platform::CreateKey(dims, uniq_name)) {
+            platform::CreateKey(dims, axis, uniq_name)) {
     auto data_softmax_md =
         mkldnn::memory::desc(dims, platform::MKLDNNGetDataType<T>(), fmt);
     auto diff_softmax_md =
@@ -95,13 +96,13 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel<T> {
     auto softmax_src_memory_p = handler.AcquireSrcMemory(input);
     auto softmax_p = handler.AcquireForwardPrimitive();
     // For Inplace src and and dst are the same memory object
-    auto softmax_dst_memory_p = input->Holder() == output->Holder()
+    auto softmax_dst_memory_p = input->IsSharedBufferWith(*output)
                                     ? softmax_src_memory_p
                                     : handler.AcquireDstMemory(output);
 
     mkldnn::stream astream(dev_ctx.GetEngine());
-    softmax_p->execute(astream, {{MKLDNN_ARG_SRC, *softmax_src_memory_p},
-                                 {MKLDNN_ARG_DST, *softmax_dst_memory_p}});
+    softmax_p->execute(astream, {{DNNL_ARG_SRC, *softmax_src_memory_p},
+                                 {DNNL_ARG_DST, *softmax_dst_memory_p}});
     astream.wait();
 
     const bool is_test = ctx.Attr<bool>("is_test");

paddle/fluid/operators/mkldnn/test_mkldnn_op_inplace.cc

@@ -27,38 +27,68 @@
 
 USE_OP(softmax);
 USE_OP_DEVICE_KERNEL(softmax, MKLDNN);
+USE_OP(elementwise_add);
+USE_OP_DEVICE_KERNEL(elementwise_add, MKLDNN);
 
 namespace paddle {
 namespace operators {
 
+struct InputVars {
+  std::string name;
+  framework::LoDTensor *tensor;
+};
+
 template <typename T>
-bool TestMain(const platform::Place &place, const framework::DDim &dims) {
+bool TestMain(const platform::Place &place, const std::string &op_type,
+              const framework::DDim &dims, const int num_inputs) {
   framework::Scope scope;
-  auto *x = scope.Var("x")->GetMutable<framework::LoDTensor>();
-  auto *y = scope.Var("y")->GetMutable<framework::LoDTensor>();
 
-  x->Resize(dims);
-  y->Resize(dims);
-
-  size_t numel = static_cast<size_t>(framework::product(dims));
-
-  auto x_ptr = x->mutable_data<T>(place);
-  auto y_ptr = y->mutable_data<T>(place);
+  std::vector<InputVars> input_names = {
+      {"x", scope.Var("x")->GetMutable<framework::LoDTensor>()},
+      {"x1", num_inputs > 1
+                 ? scope.Var("x1")->GetMutable<framework::LoDTensor>()
+                 : nullptr},
+      {"x2", num_inputs > 2
+                 ? scope.Var("x2")->GetMutable<framework::LoDTensor>()
+                 : nullptr},
+      {"x3", num_inputs > 3
+                 ? scope.Var("x3")->GetMutable<framework::LoDTensor>()
+                 : nullptr},
+      {"x4", num_inputs > 4
+                 ? scope.Var("x4")->GetMutable<framework::LoDTensor>()
+                 : nullptr}};
+  auto *y = scope.Var("y")->GetMutable<framework::LoDTensor>();
 
+  // Initialize input data
   std::uniform_real_distribution<T> dist(static_cast<T>(10.0),
                                          static_cast<T>(20.0));
   std::mt19937 engine;
+  size_t numel = static_cast<size_t>(framework::product(dims));
+  for (int i = 0; i < num_inputs; ++i) {
+    input_names[i].tensor->Resize(dims);
+    auto data_ptr = input_names[i].tensor->mutable_data<T>(place);
+    for (size_t i = 0; i < numel; ++i) {
+      data_ptr[i] = dist(engine);
+    }
+  }
 
+  // Initialize output
+  y->Resize(dims);
+  auto y_ptr = y->mutable_data<T>(place);
   for (size_t i = 0; i < numel; ++i) {
-    x_ptr[i] = dist(engine);
     y_ptr[i] = static_cast<T>(0);
   }
 
   auto &pool = platform::DeviceContextPool::Instance();
 
   // Out of place (reference) computation
-  auto op_ref = framework::OpRegistry::CreateOp(
-      "softmax", {{"X", {"x"}}}, {{"Out", {"y"}}}, {{"use_mkldnn", {true}}});
+  auto op_ref = num_inputs > 1 ? framework::OpRegistry::CreateOp(
+                                     op_type, {{"X", {"x"}}, {"Y", {"x1"}}},
+                                     {{"Out", {"y"}}}, {{"use_mkldnn", {true}}})
+                               : framework::OpRegistry::CreateOp(
+                                     op_type, {{"X", {"x"}}}, {{"Out", {"y"}}},
+                                     {{"use_mkldnn", {true}}});
+
   op_ref->Run(scope, place);
   pool.Get(place)->Wait();
 
@@ -66,15 +96,20 @@ bool TestMain(const platform::Place &place, const framework::DDim &dims) {
   auto &ref_tensor = scope.FindVar("y")->Get<framework::LoDTensor>();
 
   // In-place (to be tested) computation
-  auto op = framework::OpRegistry::CreateOp(
-      "softmax", {{"X", {"x"}}}, {{"Out", {"x"}}}, {{"use_mkldnn", {true}}});
+  auto op = num_inputs > 1 ? framework::OpRegistry::CreateOp(
+                                 op_type, {{"X", {"x"}}, {"Y", {"x1"}}},
+                                 {{"Out", {"x"}}}, {{"use_mkldnn", {true}}})
+                           : framework::OpRegistry::CreateOp(
+                                 op_type, {{"X", {"x"}}}, {{"Out", {"x"}}},
+                                 {{"use_mkldnn", {true}}});
+
   op->Run(scope, place);
   platform::DeviceContextPool::Instance().Get(place)->Wait();
 
   // Get in-place result
   auto &out_tensor = scope.FindVar("x")->Get<framework::LoDTensor>();
   PADDLE_ENFORCE_EQ(
-      &out_tensor, x,
+      &out_tensor, input_names[0].tensor,
       platform::errors::InvalidArgument(
           "Input and output vars should share tensor for In-place test"));
 
@@ -88,7 +123,13 @@ bool TestMain(const platform::Place &place, const framework::DDim &dims) {
 TEST(test_softmax_inplace, cpu_place) {
   framework::DDim dims({32, 64});
   platform::CPUPlace p;
-  ASSERT_TRUE(TestMain<float>(p, dims));
+  ASSERT_TRUE(TestMain<float>(p, "softmax", dims, 1));
+}
+
+TEST(test_elementwise_add_inplace, cpu_place) {
+  framework::DDim dims({1, 12, 20, 20});
+  platform::CPUPlace p;
+  ASSERT_TRUE(TestMain<float>(p, "elementwise_add", dims, 2));
 }
 
 }  // namespace operators

paddle/fluid/platform/mkldnn_helper.h

@@ -101,6 +101,11 @@ inline void MatchShapeToLayout(framework::Tensor* tensor_in,
   }
 }
 
+struct mkldnn_dummy_primitive {
+  struct primitive_desc {};
+  struct desc {};
+};
+
 inline mkldnn::memory::desc MKLDNNMemDesc(const std::vector<int64_t>& dims,
                                           mkldnn::memory::data_type data_type,
                                           MKLDNNMemoryFormat format) {

paddle/fluid/platform/mkldnn_reuse.h

@@ -30,7 +30,8 @@ namespace platform {
 using user_function = std::function<std::shared_ptr<float>(const float*)>;
 using memory = mkldnn::memory;
 
-template <typename T, typename TForward, typename TBackward>
+template <typename T, typename TForward,
+          typename TBackward = mkldnn_dummy_primitive>
 class MKLDNNHandlerT {
  public:
   MKLDNNHandlerT(const MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine,
@@ -351,6 +352,35 @@ class MKLDNNHandler {
   std::string key_common_;
 };
 
+template <typename T>
+class BinaryMKLDNNHandler : public platform::MKLDNNHandlerT<T, dnnl::binary> {
+ public:
+  BinaryMKLDNNHandler(const dnnl::algorithm algo,
+                      const std::vector<int64_t>& dims,
+                      const MKLDNNMemoryFormat src0_fmt,
+                      const MKLDNNMemoryFormat src1_fmt,
+                      const platform::MKLDNNDeviceContext& dev_ctx,
+                      platform::Place cpu_place, const std::string& uniq_name)
+      : platform::MKLDNNHandlerT<T, dnnl::binary>(
+            dev_ctx, dev_ctx.GetEngine(), cpu_place,
+            platform::CreateKey(dims, uniq_name)) {
+    // TODO(jczaja): Add function checking if data already exists
+    auto src0_md = dnnl::memory::desc(dims, MKLDNNGetDataType<T>(), src0_fmt);
+    auto src1_md = dnnl::memory::desc(dims, MKLDNNGetDataType<T>(), src1_fmt);
+    auto dst_md =
+        memory::desc(dims, MKLDNNGetDataType<T>(), MKLDNNMemoryFormat::any);
+
+    this->AcquireForwardPrimitiveDescriptor(algo, src0_md, src1_md, dst_md);
+  }
+
+  std::shared_ptr<mkldnn::memory> AcquireSecondSrcMemory(
+      const framework::Tensor* input) {
+    const T* input_data = input->data<T>();
+    return this->AcquireMemoryFromPrimitive(
+        this->fwd_pd_->src_desc(), to_void_cast<T>(input_data), "@src1_mem_p");
+  }
+};
+
 class SumMKLDNNHandler : public MKLDNNHandler {
  public:
   SumMKLDNNHandler(const platform::MKLDNNDeviceContext& dev_ctx,
@@ -419,7 +449,7 @@ class ActivationMKLDNNHandler
       : platform::MKLDNNHandlerT<T, mkldnn::eltwise_forward,
                                  mkldnn::eltwise_backward>(
             dev_ctx, dev_ctx.GetEngine(), cpu_place,
-            platform::CreateKey(dims, unique_name)) {
+            platform::CreateKey(dims, "a", algorithm, unique_name)) {
     auto md = mkldnn::memory::desc(dims, platform::MKLDNNGetDataType<T>(), fmt);
 
     this->AcquireForwardPrimitiveDescriptor(mkldnn::prop_kind::forward_training,
@@ -437,7 +467,7 @@ class ActivationMKLDNNHandler
       : platform::MKLDNNHandlerT<T, mkldnn::eltwise_forward,
                                  mkldnn::eltwise_backward>(
             dev_ctx, dev_ctx.GetEngine(), cpu_place,
-            platform::CreateKey(dims, unique_name)) {
+            platform::CreateKey(dims, "a", algorithm, unique_name)) {
     auto diff_dst_md = platform::MKLDNNMemDesc(
         dims, platform::MKLDNNGetDataType<T>(), diff_fmt);
     auto src_md =