Mkldnn layout (#11040)

* Add MKLDNN layout support in Paddle

Add MKLDNN layout in Paddle so that MKLDNN friendly memory layout
can be used in MKLDNN enabled OP kernel. Before this commit, NCHW
is hardcode to be used in all MKLDNN op kernels. As a result,
non-optimized execution path is selected in MKLDNN primitive which
bring worse performance.
Besides framework change, three MKLDNN OP kernels were updated
for using new MKLDNN layout. They are conv/pool2d/batch_norm.
Other MKLDNN OP kernels need be also updated in similar way to
achieve best performance.

* Add MKLDNN layout support in activation OP

* Don't populate layout from input to output when kMKLDNN in

* Refine pool mkldnn op kernel

* MKLDNN layout

* Remove the inferitance from tensor file

* MKLDNN layout: refactoring

* Remove additional #define to register new operator

* Prepare mkldnn tests to work with layout

Author: mozga-intel

paddle/fluid/framework/data_layout.h

@@ -27,6 +27,7 @@ enum class DataLayout {
   kNHWC = 0,
   kNCHW = 1,
   kAnyLayout = 2,
+  kMKLDNN = 3,  // all layouts supported by MKLDNN internally
 };
 
 inline DataLayout StringToDataLayout(const std::string& str) {
@@ -41,6 +42,8 @@ inline DataLayout StringToDataLayout(const std::string& str) {
     return DataLayout::kNCHW;
   } else if (s == "ANYLAYOUT") {
     return DataLayout::kAnyLayout;
+  } else if (s == "MKLDNNLAYOUT") {
+    return DataLayout::kMKLDNN;
   } else {
     PADDLE_THROW("Unknown storage order string: %s", s);
   }
@@ -54,8 +57,10 @@ inline std::string DataLayoutToString(const DataLayout& data_layout) {
       return "NCHW";
     case DataLayout::kAnyLayout:
       return "ANY_LAYOUT";
+    case DataLayout::kMKLDNN:
+      return "MKLDNNLAYOUT";
     default:
-      PADDLE_THROW("unknown DataLayou %d", data_layout);
+      PADDLE_THROW("unknown DataLayout %d", data_layout);
   }
 }
 

paddle/fluid/framework/data_layout_transform.cc

@@ -16,6 +16,9 @@
 #include <vector>
 
 #include "paddle/fluid/operators/math/math_function.h"
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
 
 namespace paddle {
 namespace framework {
@@ -88,5 +91,85 @@ void TransDataLayout(const OpKernelType& kernel_type_for_var,
   out->set_layout(expected_kernel_type.data_layout_);
 }
 
+#ifdef PADDLE_WITH_MKLDNN
+using mkldnn::memory;
+using mkldnn::primitive;
+using mkldnn::reorder;
+
+void* GetDataFromTensor(const Tensor& tensor, mkldnn::memory::data_type type) {
+  switch (type) {
+    case mkldnn::memory::data_type::f32:
+      return platform::to_void_cast(tensor.data<float>());
+    case mkldnn::memory::data_type::s8:
+      return platform::to_void_cast(tensor.data<char>());
+    case mkldnn::memory::data_type::u8:
+      return platform::to_void_cast(tensor.data<unsigned char>());
+    case mkldnn::memory::data_type::s16:
+      return platform::to_void_cast(tensor.data<int16_t>());
+    case mkldnn::memory::data_type::s32:
+      return platform::to_void_cast(tensor.data<int32_t>());
+    default:
+      PADDLE_THROW("wrong mkldnn type provided");
+  }
+}
+#endif
+
+void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
+                               const OpKernelType& expected_kernel_type,
+                               const Tensor& in, Tensor* out) {
+  auto in_layout = kernel_type_for_var.data_layout_;
+  auto out_layout = expected_kernel_type.data_layout_;
+
+  PADDLE_ENFORCE(
+      in_layout == DataLayout::kMKLDNN && out_layout != DataLayout::kMKLDNN,
+      "TransDataLayoutFromMKLDNN only supports transform from MKLDNN to "
+      "non-MKLDNN");
+
+#ifdef PADDLE_WITH_MKLDNN
+  PADDLE_ENFORCE(in.format() != memory::format::format_undef &&
+                     in.format() != memory::format::any,
+                 "Input tensor should have specified memory format");
+
+  // Set default as NCHW in case not specified
+  out_layout =
+      out_layout == DataLayout::kAnyLayout ? DataLayout::kNCHW : out_layout;
+
+  auto& pool = platform::DeviceContextPool::Instance();
+  auto* dev_ctx = dynamic_cast<platform::MKLDNNDeviceContext*>(
+      pool.Get(expected_kernel_type.place_));
+  auto& cpu_engine = dev_ctx->GetEngine();
+
+  std::vector<int> in_tz = paddle::framework::vectorize2int(in.dims());
+  std::vector<int> out_tz = in_tz;
+
+  memory::data_type in_type = ToMKLDNNDataType(in.type());
+  PADDLE_ENFORCE(in_type != memory::data_type::data_undef,
+                 "Input tensor type is not supported: ", in.type().name());
+  memory::data_type out_type = in_type;
+
+  memory::format in_format =
+      in_tz.size() == 2 ? memory::format::nc : in.format();
+  memory::format out_format =
+      out_tz.size() == 2 ? memory::format::nc : ToMKLDNNFormat(out_layout);
+
+  void* in_data = GetDataFromTensor(in, in_type);
+
+  // output tensor has the same dims as input. Reorder don't change dims
+  out->Resize(in.dims());
+
+  auto out_data = out->mutable_data(expected_kernel_type.place_, in.type());
+
+  auto in_memory = memory({{{in_tz}, in_type, in_format}, cpu_engine}, in_data);
+  auto out_memory =
+      memory({{{out_tz}, out_type, out_format}, cpu_engine}, out_data);
+
+  platform::Reorder(in_memory, out_memory);
+
+  out->set_layout(out_layout);
+  // reset format since the out tensor will be feed to non-MKLDNN OPkernel
+  out->set_format(memory::format::format_undef);
+#endif
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/data_layout_transform.h

@@ -14,6 +14,7 @@
 
 #pragma once
 
+#include <map>
 #include <vector>
 #include "paddle/fluid/framework/op_kernel_type.h"
 #include "paddle/fluid/framework/tensor.h"
@@ -22,6 +23,50 @@
 namespace paddle {
 namespace framework {
 
+#ifdef PADDLE_WITH_MKLDNN
+using MKLDNNFormat = mkldnn::memory::format;
+using MKLDNNDataType = mkldnn::memory::data_type;
+
+inline MKLDNNFormat ToMKLDNNFormat(const DataLayout& layout) {
+  switch (layout) {
+    case DataLayout::kNHWC:
+      return MKLDNNFormat::nhwc;
+    case DataLayout::kNCHW:
+      return MKLDNNFormat::nchw;
+    default:
+      PADDLE_THROW("Fail to convert layout %s to MKLDNN format",
+                   DataLayoutToString(layout));
+  }
+}
+
+inline DataLayout ToPaddleLayout(const MKLDNNFormat& format) {
+  switch (format) {
+    case MKLDNNFormat::nhwc:
+      return DataLayout::kNHWC;
+    case MKLDNNFormat::nchw:
+      return DataLayout::kNCHW;
+    default:
+      PADDLE_THROW("Fail to convert MKLDNN format to paddle layout");
+  }
+}
+
+inline MKLDNNDataType ToMKLDNNDataType(const std::type_index type) {
+  static const std::map<std::type_index, MKLDNNDataType> dict{
+      {std::type_index(typeid(float)), MKLDNNDataType::f32},  // NOLINT
+      {std::type_index(typeid(char)), MKLDNNDataType::s8},    // NOLINT
+      {std::type_index(typeid(unsigned char)), MKLDNNDataType::u8},
+      {std::type_index(typeid(int16_t)), MKLDNNDataType::s16},
+      {std::type_index(typeid(int32_t)), MKLDNNDataType::s32}};
+  auto iter = dict.find(type);
+  if (iter != dict.end()) return iter->second;
+  return MKLDNNDataType::data_undef;
+}
+#endif
+
+void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
+                               const OpKernelType& expected_kernel_type,
+                               const Tensor& in, Tensor* out);
+
 std::vector<int> GetAxis(const DataLayout& from, const DataLayout& to);
 
 void TransDataLayout(const OpKernelType& kernel_type_for_var,

paddle/fluid/framework/data_transform.cc

@@ -33,11 +33,34 @@ void DataTransform(const OpKernelType& expected_kernel_type,
   Tensor in;
   in.ShareDataWith(input_tensor);
   Tensor out;
+  DataLayout lin = kernel_type_for_var.data_layout_;
+  DataLayout lout = expected_kernel_type.data_layout_;
 
   // do layout transform
-  if (NeedTransformLayout(expected_kernel_type.data_layout_,
-                          kernel_type_for_var.data_layout_)) {
-    TransDataLayout(kernel_type_for_var, expected_kernel_type, in, &out);
+  if (NeedTransformLayout(lout, lin)) {
+    if (lin == DataLayout::kMKLDNN || lout == DataLayout::kMKLDNN) {
+      PADDLE_ENFORCE(
+          !(lin == DataLayout::kMKLDNN && lout == DataLayout::kMKLDNN),
+          "No layout transform needed between two MKLDNN OPKernels");
+
+      if (lin != DataLayout::kMKLDNN && lout == DataLayout::kMKLDNN) {
+#ifdef PADDLE_WITH_MKLDNN
+        // Case1 - transform from Non-MKLDNN OPKernel to MKLDNN OPKernel
+        // Just set layout/format. No real transform occur
+        out.ShareDataWith(input_tensor);
+        out.set_layout(DataLayout::kMKLDNN);
+        out.set_format(ToMKLDNNFormat(lin));
+#endif
+      } else {
+        // Case2 - transfrom from MKLDNN OPKernel to Non-MKLDNN OPKernel
+        // Do transform via MKLDNN lib
+        TransDataLayoutFromMKLDNN(kernel_type_for_var, expected_kernel_type, in,
+                                  &out);
+      }
+    } else {
+      // Case3 - transfrom between Non-MKLDNN OPKernels
+      TransDataLayout(kernel_type_for_var, expected_kernel_type, in, &out);
+    }
     transformed = true;
     PassTensorData(&out, &in);
   }

paddle/fluid/framework/op_kernel_type.h

@@ -87,7 +87,14 @@ inline std::string KernelTypeToString(const OpKernelType& kernel_key) {
 }
 
 inline bool NeedTransformLayout(const DataLayout& l, const DataLayout& r) {
-  return l != DataLayout::kAnyLayout && r != DataLayout::kAnyLayout && l != r;
+  bool ret =
+      (l != DataLayout::kAnyLayout && r != DataLayout::kAnyLayout && l != r);
+#ifdef PADDLE_WITH_MKLDNN
+  // Layout transform needed for either non-MKLDNN to MKLDNN or vice versa
+  ret |= (l != DataLayout::kMKLDNN && r == DataLayout::kMKLDNN);
+  ret |= (l == DataLayout::kMKLDNN && r != DataLayout::kMKLDNN);
+#endif
+  return ret;
 }
 
 inline bool TransFromNeeded(const OpKernelType& l, const OpKernelType& r) {

paddle/fluid/framework/op_registry.h

@@ -83,8 +83,14 @@ struct OpKernelRegistrarFunctor<PlaceType, false, I, KernelTypes...> {
 
   void operator()(const char* op_type, const char* library_type) const {
     using T = typename KERNEL_TYPE::ELEMENT_TYPE;
+    std::string library(library_type);
+    std::string data_layout = "ANYLAYOUT";
+    if (library == "MKLDNN") {
+      data_layout = "MKLDNNLAYOUT";
+    }
     OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType(),
-                     DataLayout::kAnyLayout, StringToLibraryType(library_type));
+                     StringToDataLayout(data_layout),
+                     StringToLibraryType(library_type));
     OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KERNEL_TYPE);
 
     constexpr auto size = std::tuple_size<std::tuple<KernelTypes...>>::value;
@@ -99,7 +105,8 @@ struct OpKernelRegistrarFunctor<PlaceType, true, I, KernelType...> {
   void operator()(const char* op_type, const char* library_type) const {}
 };
 
-// User can register many kernel in one place. The data type could be different.
+// User can register many kernel in one place. The data type could be
+// different.
 template <typename PlaceType, typename... KernelType>
 class OpKernelRegistrar : public Registrar {
  public:

paddle/fluid/framework/operator.cc

@@ -444,10 +444,25 @@ class RuntimeInferShapeContext : public InferShapeContext {
     auto* out_tensor = out_var->GetMutable<LoDTensor>();
     out_tensor->set_lod(in_tensor.lod());
 
-    // TODO(dzhwinter) : reuse ShareLoD in most operators.
-    // Need to call ShareLayout explicitly in sequence related ops.
-    // Shall we have a better method to shared info between in/out Tensor?
-    out_tensor->set_layout(in_tensor.layout());
+// TODO(dzhwinter) : reuse ShareLoD in most operators.
+// Need to call ShareLayout explicitly in sequence related ops.
+// Shall we have a better method to shared info between in/out Tensor?
+#ifdef PADDLE_WITH_MKLDNN
+    // Fix me: ugly workaround below
+    // Correct solution:
+    //    set_layout() should NOT be called here (i.e. ShareLoD). Instead,
+    //    layout of output tensor should be set "manually" in Compute()
+    //    of each OPKernel. The reason layout should NOT be shared between
+    //    input and output "automatically" (now by InferShape()->ShareLoD())
+    //    is that layout transform may occur after InferShape().
+    // Workaround:
+    //    Skip set_layout() when input layout is kMKLDNN
+    //    This is to avoid kMKLDNN is populated wrongly into a non-MKLDNN
+    //    OPKernel. In all MKLDNN OPkernel, set_layout(kMKLDNN) should be called
+    //    in Compute()
+    if (in_tensor.layout() != DataLayout::kMKLDNN)
+#endif
+      out_tensor->set_layout(in_tensor.layout());
   }
 
   void ShareLayout(const std::string& in, const std::string& out, size_t i = 0,
@@ -665,7 +680,8 @@ OpKernelType OperatorWithKernel::GetExpectedKernelType(
 OpKernelType OperatorWithKernel::GetKernelTypeForVar(
     const std::string& var_name, const Tensor& tensor,
     const OpKernelType& expected_kernel_type) const {
-  return OpKernelType(expected_kernel_type.data_type_, tensor.place());
+  return OpKernelType(expected_kernel_type.data_type_, tensor.place(),
+                      tensor.layout());
 }
 
 }  // namespace framework

paddle/fluid/framework/tensor.h

@@ -34,6 +34,28 @@ namespace framework {
 class LoDTensor;
 
 class Tensor {
+#ifdef PADDLE_WITH_MKLDNN
+
+ public:
+  inline mkldnn::memory::format format() const { return format_; }
+
+  inline void set_format(const mkldnn::memory::format format) {
+    format_ = format;
+  }
+
+ protected:
+  /**
+   * @brief the detail format of memory block which have layout as kMKLDNN
+   *
+   * @note MKLDNN lib support various memory format like nchw, nhwc, nChw8C,
+   *       nChw16c, etc. For a MKLDNN memory block, layout will be set as
+   *       DataLayout::kMKLDNN meanwhile detail memory format will be kept in
+   *       this field.
+   */
+
+  mkldnn::memory::format format_ = mkldnn::memory::format::format_undef;
+#endif
+
  public:
   template <typename T, size_t D, int MajorType, typename IndexType>
   friend struct EigenTensor;
@@ -195,8 +217,10 @@ class Tensor {
    *       N,C,H,W for respectively the batch size, the number of
    *       feature maps, the height.
    */
-
-  DataLayout layout_ = DataLayout::kNHWC;
+  // Fix me: here just change the default layout to kNCHW
+  // it doesn't fix the real issue, i.e. feeder should set up tensor layout
+  // according to actual input data
+  DataLayout layout_ = DataLayout::kNCHW;
 
   /**
    * @brief   A PlaceHolder may be shared by more than one tensor.

paddle/fluid/framework/tensor_test.cc

@@ -209,7 +209,7 @@ TEST(Tensor, ReshapeToMatrix) {
 
 TEST(Tensor, Layout) {
   framework::Tensor src;
-  ASSERT_EQ(src.layout(), framework::DataLayout::kNHWC);
+  ASSERT_EQ(src.layout(), framework::DataLayout::kNCHW);
   src.set_layout(framework::DataLayout::kAnyLayout);
   ASSERT_EQ(src.layout(), framework::DataLayout::kAnyLayout);
 }

paddle/fluid/operators/activation_op.cc

@@ -58,14 +58,16 @@ framework::OpKernelType GetKernelType(const framework::ExecutionContext& ctx,
                                       const framework::OperatorWithKernel& oper,
                                       const std::string& name) {
   framework::LibraryType library{framework::LibraryType::kPlain};
+
+  framework::DataLayout layout = framework::DataLayout::kAnyLayout;
 #ifdef PADDLE_WITH_MKLDNN
   auto it = oper.Attrs().find("use_mkldnn");
   if (library == framework::LibraryType::kPlain && it != oper.Attrs().end() &&
       platform::CanMKLDNNBeUsed(ctx)) {
     library = framework::LibraryType::kMKLDNN;
+    layout = framework::DataLayout::kMKLDNN;
   }
 #endif
-  framework::DataLayout layout = framework::DataLayout::kAnyLayout;
   return framework::OpKernelType(
       framework::ToDataType(ctx.Input<framework::Tensor>(name)->type()),
       ctx.GetPlace(), layout, library);