[cherry-pick] Squeeze2 and transpose2 fuse using oneDNN(#47712)

* suqeeze2 + transpose2 fuse onednn cherrypick 2.4

* format

* fix merge

Author: zh794390558

paddle/fluid/framework/ir/CMakeLists.txt

@@ -219,6 +219,7 @@ if(WITH_MKLDNN)
   pass_library(matmul_elementwise_add_mkldnn_fuse_pass inference DIR mkldnn)
   pass_library(matmul_activation_mkldnn_fuse_pass inference DIR mkldnn)
   pass_library(operator_scale_onednn_fuse_pass inference DIR mkldnn)
+  pass_library(squeeze2_transpose2_onednn_fuse_pass inference DIR mkldnn)
   pass_library(operator_unsqueeze2_onednn_fuse_pass inference DIR mkldnn)
   pass_library(operator_reshape2_onednn_fuse_pass inference DIR mkldnn)
   pass_library(cpu_quantize_placement_pass base DIR mkldnn)

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -1042,6 +1042,25 @@ PDNode *patterns::SeqConvEltAddRelu::operator()(
   return relu_out_var;
 }
 
+PDNode *patterns::Squeeze2Transpose2::operator()() {
+  auto *squeeze2_op_in = pattern->NewNode(squeeze2_op_in_repr())
+                             ->AsInput()
+                             ->assert_is_op_input("squeeze2", "X");
+  auto *squeeze2_op = pattern->NewNode(squeeze2_op_repr())
+                          ->assert_is_op("squeeze2")
+                          ->assert_has_n_outputs(2);
+  auto *squeeze2_op_out = pattern->NewNode(squeeze2_op_out_repr())
+                              ->AsIntermediate()
+                              ->assert_is_op_output("squeeze2", "Out")
+                              ->assert_is_op_input("transpose2", "X");
+  auto *transpose2_op =
+      pattern->NewNode(transpose2_op_repr())->assert_is_op("transpose2");
+
+  squeeze2_op->LinksFrom({squeeze2_op_in}).LinksTo({squeeze2_op_out});
+  transpose2_op->LinksFrom({squeeze2_op_out});
+  return transpose2_op;
+}
+
 PDNode *patterns::FC::operator()(paddle::framework::ir::PDNode *x,
                                  bool with_bias,
                                  bool with_relu) {

paddle/fluid/framework/ir/graph_pattern_detector.h

@@ -634,6 +634,20 @@ struct FCMKLDNN : public PatternBase {
   PATTERN_DECL_NODE(output);
 };
 
+// Squeeze2 + Transpose2
+// Forward pass
+struct Squeeze2Transpose2 : public PatternBase {
+  Squeeze2Transpose2(PDPattern* pattern, const std::string& name_scope)
+      : PatternBase(pattern, name_scope, "squeeze2_transpose2") {}
+
+  PDNode* operator()();
+
+  PATTERN_DECL_NODE(squeeze2_op_in);
+  PATTERN_DECL_NODE(squeeze2_op);
+  PATTERN_DECL_NODE(squeeze2_op_out);
+  PATTERN_DECL_NODE(transpose2_op);
+};
+
 // Embedding
 struct Embedding : public PatternBase {
   Embedding(PDPattern* pattern, const std::string& name_scope)
@@ -2002,6 +2016,12 @@ struct AddSupportInt8 : public PatternBase {
   out_var->inputs.clear();          \
   out_var->inputs.push_back(op);
 
+// Set the in_var as the input of the op
+#define IR_VAR_OP_LINK(in_var, op) \
+  in_var->outputs.clear();         \
+  in_var->outputs.push_back(op);   \
+  op->inputs.push_back(in_var);
+
 }  // namespace ir
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/ir/mkldnn/squeeze2_transpose2_onednn_fuse_pass.cc

@@ -0,0 +1,86 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+#include "paddle/fluid/framework/ir/mkldnn/squeeze2_transpose2_onednn_fuse_pass.h"
+#include "paddle/fluid/framework/op_version_registry.h"
+#include "paddle/fluid/platform/mkldnn_reuse.h"
+#include "paddle/fluid/string/pretty_log.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+using string::PrettyLogDetail;
+
+void FuseSqueeze2Transpose2OneDNNPass::ApplyImpl(Graph *graph) const {
+  PADDLE_ENFORCE_NOT_NULL(graph,
+                          platform::errors::InvalidArgument(
+                              "Pointer to graph argument should not be NULL."));
+
+  FusePassBase::Init("squeeze2_transpose2_onednn_fuse_pass", graph);
+
+  GraphPatternDetector gpd;
+  patterns::Squeeze2Transpose2 squeeze2_transpose2_pattern(
+      gpd.mutable_pattern(), "squeeze2_transpose2_onednn_fuse_pass");
+  squeeze2_transpose2_pattern();
+
+  int found_count = 0;
+
+  auto handler = [&](const GraphPatternDetector::subgraph_t &subgraph,
+                     Graph *g) {
+    GET_IR_NODE_FROM_SUBGRAPH(
+        squeeze2_op_in, squeeze2_op_in, squeeze2_transpose2_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(
+        squeeze2_op, squeeze2_op, squeeze2_transpose2_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(
+        squeeze2_op_out, squeeze2_op_out, squeeze2_transpose2_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(
+        transpose2_op, transpose2_op, squeeze2_transpose2_pattern);
+
+    if (!transpose2_op->Op()->HasAttr("use_mkldnn") ||
+        (transpose2_op->Op()->HasAttr("use_mkldnn") &&
+         !(PADDLE_GET_CONST(bool,
+                            transpose2_op->Op()->GetAttr("use_mkldnn"))))) {
+      VLOG(4) << "Only oneDNN version of transpose2 can be fused after with "
+                 "squeeze2.";
+      return;
+    }
+
+    std::vector<int> squeeze2_axes =
+        PADDLE_GET_CONST(std::vector<int>, squeeze2_op->Op()->GetAttr("axes"));
+    transpose2_op->Op()->SetAttr("fused_squeeze2_axes", squeeze2_axes);
+    transpose2_op->Op()->SetInput("X", {squeeze2_op_in->Name()});
+
+    IR_VAR_OP_LINK(squeeze2_op_in, transpose2_op);
+    GraphSafeRemoveNodes(g, {squeeze2_op, squeeze2_op_out});
+    found_count++;
+  };
+
+  gpd(graph, handler);
+  AddStatis(found_count);
+  if ((!Has("disable_logs") || !Get<bool>("disable_logs"))) {
+    PrettyLogDetail("--- fused %d squeeze2 with transpose2", found_count);
+  }
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(squeeze2_transpose2_onednn_fuse_pass,
+              paddle::framework::ir::FuseSqueeze2Transpose2OneDNNPass);
+REGISTER_PASS_CAPABILITY(squeeze2_transpose2_onednn_fuse_pass)
+    .AddCombination(
+        paddle::framework::compatible::OpVersionComparatorCombination()
+            .GE("squeeze2", 0)
+            .GE("transpose2", 0));

paddle/fluid/framework/ir/mkldnn/squeeze2_transpose2_onednn_fuse_pass.h

@@ -0,0 +1,35 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+#include "paddle/fluid/framework/ir/graph.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+class FuseSqueeze2Transpose2OneDNNPass : public FusePassBase {
+ public:
+  virtual ~FuseSqueeze2Transpose2OneDNNPass() {}
+
+ protected:
+  void ApplyImpl(Graph *graph) const override;
+};
+
+}  // namespace ir
+}  // namespace framework
+
+}  // namespace paddle

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -307,6 +307,7 @@ void CpuPassStrategy::EnableMKLDNN() {
     passes_.insert(passes_.begin(), "mkldnn_placement_pass");
 
     for (auto &pass : std::vector<std::string>({
+             "squeeze2_transpose2_onednn_fuse_pass",
              "depthwise_conv_mkldnn_pass",    //
              "conv_bn_fuse_pass",             // Execute BN passes again to
              "conv_eltwiseadd_bn_fuse_pass",  // preserve correct pass order
@@ -386,6 +387,7 @@ void CpuPassStrategy::EnableMkldnnInt8() {
     passes_.push_back("mkldnn_placement_pass");
     passes_.push_back("simplify_with_basic_ops_pass");
     passes_.push_back("constant_folding_pass");
+    passes_.push_back("squeeze2_transpose2_onednn_fuse_pass");
     passes_.push_back("layer_norm_fuse_pass");
     passes_.push_back("attention_lstm_fuse_pass");
     passes_.push_back("seqconv_eltadd_relu_fuse_pass");

paddle/fluid/operators/mkldnn/transpose_mkldnn_op.cc

@@ -42,6 +42,9 @@ class TransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
 
     auto& astream = platform::MKLDNNDeviceContext::tls().get_stream();
 
+    platform::SetInMemDescWithLogicalLayoutFusesSupport(
+        ctx, const_cast<phi::DenseTensor*>(x), x->mem_desc());
+
     if (ndims == 1) {
       framework::TensorCopy(*x, x->place(), out);
       out->set_mem_desc(x->mem_desc());

paddle/fluid/operators/transpose_op.cc

@@ -39,11 +39,13 @@ class TransposeOp : public framework::OperatorWithKernel {
     size_t x_rank = x_dims.size();
     size_t axis_size = axis.size();
 
-    PADDLE_ENFORCE_EQ(x_rank,
+    // Note: x_rank > axis_size when fuse squeeze2 + transpose2, else x_rank ==
+    // axis_size
+    PADDLE_ENFORCE_GE(x_rank,
                       axis_size,
                       platform::errors::InvalidArgument(
                           "The input tensor's dimension "
-                          "should be equal to the axis's size. "
+                          "should be equal to or greater than the axis's size. "
                           "But received input tensor's dimension is %d, "
                           "axis's size is %d",
                           x_rank,

paddle/fluid/platform/mkldnn_reuse.h

@@ -120,12 +120,10 @@ static void SetOutMemDescWithUnsqueeze2FuseSupport(
   const std::vector<int64_t>& op_tz = out_md.dims();
   std::vector<int64_t> unsqueezed_op_tz(
       op_tz.size() + fused_unsqueeze2_axes.size(), 0);
-
   for (const auto& axis : fused_unsqueeze2_axes) {
     int positive_axis = axis < 0 ? unsqueezed_op_tz.size() + axis : axis;
     unsqueezed_op_tz[positive_axis] = 1;
   }
-
   int j = 0;
   for (size_t i = 0; i < unsqueezed_op_tz.size(); ++i) {
     if (unsqueezed_op_tz[i] == 0) {
@@ -143,20 +141,17 @@ static void SetOutMemDescWithReshape2FuseSupport(
   std::vector<int64_t> fused_reshape2_shape(
       ctx.Attr<std::vector<int>>("fused_reshape2_shape").begin(),
       ctx.Attr<std::vector<int>>("fused_reshape2_shape").end());
-
   const int out_shape_numel = out->numel();
   const int new_shape_numel = std::accumulate(fused_reshape2_shape.begin(),
                                               fused_reshape2_shape.end(),
                                               1,
                                               std::multiplies<int64_t>());
-
   for (size_t i = 0; i < fused_reshape2_shape.size(); ++i) {
     if (fused_reshape2_shape[i] == -1) {
       fused_reshape2_shape[i] = -out_shape_numel / new_shape_numel;
       break;
     }
   }
-
   out->set_mem_desc(out_md.reshape(fused_reshape2_shape));
   out->Resize(phi::make_ddim(fused_reshape2_shape));
 }
@@ -169,11 +164,58 @@ static void SetOutMemDescWithLogicalLayoutFusesSupport(
     SetOutMemDescWithUnsqueeze2FuseSupport(ctx, out, out_md);
   } else if (ctx.HasAttr("fused_reshape2_shape")) {
     SetOutMemDescWithReshape2FuseSupport(ctx, out, out_md);
+  } else if (ctx.HasAttr("fused_squeeze2_axes")) {
+    out->set_mem_desc(out_md);
+    out->Resize(phi::make_ddim(out_md.dims()));
   } else {
     out->set_mem_desc(out_md);
   }
 }
 
+static void SetInMemDescWithSqueeze2FuseSupport(
+    const framework::ExecutionContext& ctx,
+    phi::DenseTensor* in,
+    const dnnl::memory::desc& in_md) {
+  const std::vector<int> fused_squeeze2_axes =
+      ctx.Attr<std::vector<int>>("fused_squeeze2_axes");
+  const std::set<int64_t> squeeze2_axes_set(fused_squeeze2_axes.begin(),
+                                            fused_squeeze2_axes.end());
+  const std::vector<int64_t>& x_vec_dims = in_md.dims();
+  std::vector<int64_t> squeezed_op_tz(
+      x_vec_dims.size() - fused_squeeze2_axes.size(), 0);
+
+  int j = 0;
+  for (size_t i = 0; i < x_vec_dims.size(); ++i) {
+    if (squeeze2_axes_set.count(i) ||
+        squeeze2_axes_set.count(i - x_vec_dims.size())) {
+      PADDLE_ENFORCE_EQ(
+          x_vec_dims[i],
+          1,
+          platform::errors::InvalidArgument(
+              "Squeeze2 input '%d' dim should be equal to one, but get '%d'.",
+              i,
+              x_vec_dims[i]));
+      continue;
+    }
+    squeezed_op_tz[j++] = x_vec_dims[i];
+  }
+
+  in->set_mem_desc(in_md.reshape(squeezed_op_tz));
+  in->Resize(phi::make_ddim(squeezed_op_tz));
+}
+
+static void SetInMemDescWithLogicalLayoutFusesSupport(
+    const framework::ExecutionContext& ctx,
+    phi::DenseTensor* in,
+    const dnnl::memory::desc& in_md) {
+  if (ctx.HasAttr("fused_squeeze2_axes")) {
+    SetInMemDescWithSqueeze2FuseSupport(ctx, in, in_md);
+  } else {
+    in->set_mem_desc(in_md);
+    in->Resize(phi::make_ddim(in_md.dims()));
+  }
+}
+
 template <typename T>
 constexpr bool IsInt8() {
   return std::is_same<T, int8_t>::value || std::is_same<T, uint8_t>::value;