Merge pull request #9385 from mozga-intel/mozga/mkldnn-fc

Implementation of MKLDNN FC

paddle/fluid/operators/fc_op.cc

@@ -0,0 +1,102 @@
+/* Copyright (c) 2018 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/operators/fc_op.h"
+#include <vector>
+
+namespace paddle {
+namespace operators {
+
+void FCOp::InferShape(framework::InferShapeContext* ctx) const {
+  PADDLE_ENFORCE(ctx->HasInput("Input"),
+                 "X(Input) of Fully Connected should not be null.");
+  PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                 "Out(Output) of Fully Connected should not be null.");
+  PADDLE_ENFORCE(ctx->HasInput("W"),
+                 "W(Input) of Fully Connected should not be null.");
+
+  auto in_dims = ctx->GetInputDim("Input");
+  auto w_dims = ctx->GetInputDim("W");
+  std::vector<int64_t> output_shape({in_dims[0], w_dims[1]});
+
+  PADDLE_ENFORCE(in_dims.size() == 2 || in_dims.size() == 4,
+                 "Fully Connected input should be 2-D or 4-D tensor.");
+
+  PADDLE_ENFORCE(w_dims.size() == 2 || w_dims.size() == 4,
+                 "Fully Connected input should be 2-D or 4-D tensor.");
+
+  ctx->SetOutputDim("Out", framework::make_ddim(output_shape));
+  ctx->ShareLoD("Input", "Out");
+}
+
+framework::OpKernelType FCOp::GetExpectedKernelType(
+    const framework::ExecutionContext& ctx) const {
+  framework::LibraryType library{framework::LibraryType::kMKLDNN};
+  framework::DataLayout layout{framework::DataLayout::kAnyLayout};
+
+  return framework::OpKernelType(
+      framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
+      layout, library);
+}
+
+void FCOpGrad::InferShape(framework::InferShapeContext* ctx) const {
+  auto in_dims = ctx->GetInputDim("Input");
+  auto w_dims = ctx->GetInputDim("W");
+
+  if (ctx->HasOutput(framework::GradVarName("Input"))) {
+    ctx->SetOutputDim(framework::GradVarName("Input"), in_dims);
+  }
+  if (ctx->HasOutput(framework::GradVarName("W"))) {
+    ctx->SetOutputDim(framework::GradVarName("W"), w_dims);
+  }
+}
+
+framework::OpKernelType FCOpGrad::GetExpectedKernelType(
+    const framework::ExecutionContext& ctx) const {
+  framework::LibraryType library{framework::LibraryType::kMKLDNN};
+  framework::DataLayout layout{framework::DataLayout::kAnyLayout};
+
+  return framework::OpKernelType(
+      framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
+      layout, library);
+}
+
+FCOpMaker::FCOpMaker(OpProto* proto, OpAttrChecker* op_checker)
+    : OpProtoAndCheckerMaker(proto, op_checker) {
+  AddInput("Input", "(Tensor) The input tensor of fully connected operator. ");
+  AddInput("W", "(Tensor), The second input tensor of fc op.");
+  AddOutput("Out", "(Tensor) The output tensor of fully connected operator. ");
+  AddAttr<bool>("use_mkldnn",
+                "(bool, default false) Only used in mkldnn kernel")
+      .SetDefault(false);
+  AddAttr<bool>("bias_attr", "(bool, default false) Only used in mkldnn kernel")
+      .SetDefault(false);
+  AddComment(R"DOC(
+  Fully Connected Operator.
+
+  The fully connected operation calculates the output based on the input, weights and bias attribute.
+  The size of each dimension of the parameters checked in the infer-shape.
+  The matrix of bias is generated by the mkldnn framework, when the bias_attr is True.
+  Additional parametrs are use_mkldnn and bias_attr.
+  The input(X) size and output(Out) size may be diffrent.
+
+  The fully connected layer only supports MKLDNN version
+)DOC");
+}
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP(fc, paddle::operators::FCOp, paddle::operators::FCOpMaker, fc_grad,
+            paddle::operators::FCOpGrad);

paddle/fluid/operators/fc_op.h

@@ -0,0 +1,52 @@
+/* Copyright (c) 2018 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/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+class FCOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override;
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override;
+};
+
+class FCOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override;
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override;
+};
+
+class FCOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  FCOpMaker(OpProto* proto, OpAttrChecker* op_checker);
+};
+
+}  // namespace operators
+}  // namespace paddle

python/paddle/fluid/layers/nn.py

@@ -133,6 +133,8 @@ def fc(input,
         bias_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for the bias
             of this layer. If it is set to None, no bias will be added to the output units.
         act (str, default None): Activation to be applied to the output of this layer.
+        use_mkldnn(bool): Use mkldnn kernel or not, it is valid only when the mkldnn
+            library is installed. Default: False
         name (str, default None): The name of this layer.
 
     Returns:
@@ -153,6 +155,31 @@ def fc(input,
     dtype = helper.input_dtype()
 
     mul_results = []
+    if use_mkldnn:
+        tmp = helper.create_tmp_variable(dtype)
+        input_shape = input.shape
+        param_shape = [
+            reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
+        ] + [size]
+
+        w = helper.create_parameter(
+            attr=helper.param_attr,
+            shape=param_shape,
+            dtype=dtype,
+            is_bias=False)
+        if bias_attr is None or bias_attr is False:
+            bias_attr = False
+        else:
+            bias_attr = True
+        helper.append_op(
+            type="fc",
+            inputs={"Input": input,
+                    "W": w},
+            outputs={"Out": tmp},
+            attrs={"use_mkldnn": use_mkldnn,
+                   "bias_attr": bias_attr})
+        return helper.append_activation(tmp)
+    else:
         for input_var, param_attr in helper.iter_inputs_and_params():
             input_shape = input_var.shape
             param_shape = [
@@ -170,19 +197,20 @@ def fc(input,
                 attrs={
                     "x_num_col_dims": num_flatten_dims,
                     "y_num_col_dims": 1,
-                'use_mkldnn': use_mkldnn
                 })
             mul_results.append(tmp)
 
-    # sum
         if len(mul_results) == 1:
             pre_bias = mul_results[0]
         else:
             pre_bias = helper.create_tmp_variable(dtype)
             helper.append_op(
-            type="sum", inputs={"X": mul_results}, outputs={"Out": pre_bias})
+                type="sum",
+                inputs={"X": mul_results},
+                outputs={"Out": pre_bias})
         # add bias
-    pre_activation = helper.append_bias_op(pre_bias, dim_start=num_flatten_dims)
+        pre_activation = helper.append_bias_op(
+            pre_bias, dim_start=num_flatten_dims)
         # add activation
         return helper.append_activation(pre_activation)
 

python/paddle/fluid/tests/unittests/test_fc_op.py

@@ -0,0 +1,99 @@
+# Copyright (c) 2018 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.
+
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+def fully_connected_naive(input, weights, bias_data=None):
+    in_n, in_c, in_h, in_w = input.shape
+    w_h, w_c = weights.shape
+
+    x_data = np.reshape(input, [in_n, in_c * in_h * in_w])
+    w_data = np.transpose(np.reshape(weights, (w_c, in_c * in_h * in_w)))
+    result = None
+
+    if not bias_data:
+        result = np.dot(x_data, w_data)
+    else:
+        result = np.dot(x_data, w_data) + bias_data
+
+    return result
+
+
+class MatrixGenerate:
+    def __init__(self, mb, ic, oc, h, w):
+        self.input = np.random.random((mb, ic, h, w)).astype("float32")
+        self.weights = np.random.random((ic * h * w, oc)).astype("float32")
+
+
+class TestFCMKLDNNOp(OpTest):
+    def setUp(self):
+        self.op_type = "fc"
+        self.use_mkldnn = True
+        self.with_bias = True
+        self.matrix = MatrixGenerate(1, 10, 15, 3, 3)
+
+        self.inputs = {'Input': self.matrix.input, 'W': self.matrix.weights}
+
+        self.attrs = {
+            'use_mkldnn': self.use_mkldnn,
+            'with_bias': self.with_bias
+        }
+
+        self.outputs = {
+            'Out': fully_connected_naive(self.matrix.input, self.matrix.weights)
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(set(['Input', 'W']), 'Out', max_relative_error=0.9)
+
+    def test_check_grad_no_weight(self):
+        self.check_grad(
+            ['Input'], 'Out', max_relative_error=0.5, no_grad_set=set('W'))
+
+
+class TestFCMKLDNNOp1(TestFCMKLDNNOp):
+    def init_op_type(self):
+        self.matrix = MatrixGenerate(2, 15, 48, 2, 2)
+
+
+class TestFCMKLDNNOp2(TestFCMKLDNNOp):
+    def init_op_type(self):
+        self.matrix = MatrixGenerate(2, 32, 40, 1, 1)
+
+
+class TestFCMKLDNNOp3(TestFCMKLDNNOp):
+    def init_op_type(self):
+        self.matrix = MatrixGenerate(2, 2, 4, 1, 1)
+
+
+class TestFCMKLDNNOp4(TestFCMKLDNNOp):
+    def init_op_type(self):
+        self.with_bias = False
+        self.matrix = MatrixGenerate(2, 32, 48, 2, 2)
+
+
+class TestFCMKLDNNOp4(TestFCMKLDNNOp):
+    def init_op_type(self):
+        self.with_bias = False
+        self.matrix = MatrixGenerate(2, 32, 1000, 6, 6)
+
+
+if __name__ == "__main__":
+    unittest.main()