From b9dbb7c5cbad8e25cb16af07af6b58764c27ae6e Mon Sep 17 00:00:00 2001
From: tensor-tang <tangjian03@baidu.com>
Date: Tue, 14 Aug 2018 15:47:15 +0800
Subject: [PATCH] fix bias attri in mkldnn fc

---
 paddle/fluid/operators/fc_mkldnn_op.cc                 | 10 +++++++---
 .../paddle/fluid/tests/unittests/test_fc_mkldnn_op.py  |  9 ++-------
 2 files changed, 9 insertions(+), 10 deletions(-)

diff --git a/paddle/fluid/operators/fc_mkldnn_op.cc b/paddle/fluid/operators/fc_mkldnn_op.cc
index 68a47dd6ad..e595f1a627 100644
--- a/paddle/fluid/operators/fc_mkldnn_op.cc
+++ b/paddle/fluid/operators/fc_mkldnn_op.cc
@@ -125,14 +125,16 @@ class FCMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
 
     auto input = ctx.Input<Tensor>("Input");
     auto w = ctx.Input<Tensor>("W");
+    auto bias = ctx.Input<Tensor>("Bias");
 
     PADDLE_ENFORCE(input->dims().size() == 2 || input->dims().size() == 4,
                    "Input must be with 2 or 4 dimensions, i.e. NCHW");
-    // TODO(intel): the src weight is io and mkldnn weight need be transposed !
+    // TODO(intel friends): the native weight format is io,
+    // but the mkldnn weight format is oihw, which may need be transposed.
     PADDLE_ENFORCE(w->dims().size() == 2 || w->dims().size() == 4,
                    "Weights must be with 2 or 4 dimensions, i.e. OI or OIHW");
 
-    bool with_bias = ctx.Attr<bool>("bias_attr");
+    bool with_bias = bias != nullptr;
     MKLDNNMD<Tensor> md(input, w, with_bias);
 
     std::shared_ptr<mkldnn::inner_product_forward::primitive_desc> pd =
@@ -155,6 +157,7 @@ class FCMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     auto dst_memory = mem.dst(output_data);
     auto src_memory = mem.src(input_data);
     auto weights_memory = mem.weights(w_data);
+    // TODO(intel friends): bias memory should also be obtain from bias->data()
     auto bias_memory = mem.bias();
 
     auto forward = with_bias ? mkldnn::inner_product_forward(
@@ -217,7 +220,8 @@ class FCMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
     const Tensor* out_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
     const T* out_grad_data = out_grad->data<T>();
 
-    bool with_bias = ctx.Attr<bool>("bias_attr");
+    auto bias = ctx.Input<Tensor>("Bias");
+    bool with_bias = bias != nullptr;
 
     MKLDNNMD<Tensor> md(input, w, with_bias);
     MKLDNNMemory mem(&md, mkldnn_engine);
diff --git a/python/paddle/fluid/tests/unittests/test_fc_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_fc_mkldnn_op.py
index 3f547f3c48..099e6e6064 100644
--- a/python/paddle/fluid/tests/unittests/test_fc_mkldnn_op.py
+++ b/python/paddle/fluid/tests/unittests/test_fc_mkldnn_op.py
@@ -22,6 +22,7 @@ def fully_connected_naive(input, weights, bias_data=None):
     w_h, w_c = weights.shape
 
     x_data = np.reshape(input, [in_n, in_c * in_h * in_w])
+    # this transpose should be implemented at C code
     w_data = np.transpose(np.reshape(weights, (w_c, in_c * in_h * in_w)))
     result = None
 
@@ -43,15 +44,11 @@ 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.attrs = {'use_mkldnn': self.use_mkldnn, }
 
         self.outputs = {
             'Out': fully_connected_naive(self.matrix.input, self.matrix.weights)
@@ -85,13 +82,11 @@ class TestFCMKLDNNOp3(TestFCMKLDNNOp):
 
 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)
 
 
-- 
GitLab