add compare simple net of mkldnn and cpu

paddle/gserver/layers/MKLDNNConvLayer.cpp

@@ -251,22 +251,31 @@ void MKLDNNConvLayer::resetInValue(
   // create buffer and reorder if input value do not match
   cpuInVal_ = nullptr;
   cvtInVal_ = nullptr;
-  if (inputIsOnlyMKLDNN()) {
-    MKLDNNMatrixPtr dnnIn = std::dynamic_pointer_cast<MKLDNNMatrix>(inMat);
-    CHECK(dnnIn) << "Input should be MKLDNNMatrix";
-    if (dnnIn->getPrimitiveDesc() != in->getPrimitiveDesc()) {
-      CHECK_EQ(dnnIn->getFormat(), format::nc);
+
+  MKLDNNMatrixPtr dnnIn = std::dynamic_pointer_cast<MKLDNNMatrix>(inMat);
+  CHECK_EQ(inputIsOnlyMKLDNN(), dnnIn != nullptr);
+  if (dnnIn != nullptr && dnnIn->getPrimitiveDesc() == in->getPrimitiveDesc()) {
+    in = dnnIn;
+    return;
+  }
+  if (dnnIn) {
+    if (dnnIn->getFormat() == format::nc) {
       CHECK(ih_ == 1 && iw_ == 1) << "when input is nc format";
       // create a new one with nchw format and same data
       memory::dims inDims = memory::dims{bs_, ic_, 1, 1};
       dnnIn = MKLDNNMatrix::create(inMat, inDims, format::nchw, engine_);
-      CHECK(dnnIn->getPrimitiveDesc() == in->getPrimitiveDesc());
     }
-    in = dnnIn;
+    if (dnnIn->getPrimitiveDesc() == in->getPrimitiveDesc()) {
+      in = dnnIn;
+      return;
+    }
+    cpuInVal_ = dnnIn;
+    in = MKLDNNMatrix::create(nullptr, pd->src_primitive_desc());
+    cvtInVal_ = MKLDNNMatrix::createReorder(cpuInVal_, in);
+    CHECK(cvtInVal_) << "should not be emptry";
   } else {
-    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
     memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
-    cpuInVal_ = MKLDNNMatrix::create(cpuIn, inDims, format::nchw, engine_);
+    cpuInVal_ = MKLDNNMatrix::create(inMat, inDims, format::nchw, engine_);
     if (cpuInVal_->getPrimitiveDesc() != in->getPrimitiveDesc()) {
       // create new mkldnn matrix
       in = MKLDNNMatrix::create(nullptr, pd->src_primitive_desc());

paddle/trainer/tests/CMakeLists.txt

@@ -37,6 +37,19 @@ add_test(NAME test_CompareTwoNets
             --config_file_a=trainer/tests/sample_trainer_config_qb_rnn.conf --config_file_b=trainer/tests/sample_trainer_config_rnn.conf
     WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
 
+################ test_CompareMKLDNNandCPU ######################
+if(WITH_MKLDNN)
+  add_unittest_without_exec(test_CompareMKLDNNandCPU
+      test_CompareTwoNets.cpp)
+  add_test(NAME test_CompareMKLDNNandCPU
+    COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/
+          ${CMAKE_CURRENT_BINARY_DIR}/test_CompareMKLDNNandCPU
+              --config_file_a=trainer/tests/sample_trainer_config_simple_net.conf --use_mkldnn_a=True
+              --config_file_b=trainer/tests/sample_trainer_config_simple_net.conf --use_mkldnn_b=False
+              --use_gpu=False
+      WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
+endif()
+
 ############### test_CompareTwoOpts ###################
 add_unittest_without_exec(test_CompareTwoOpts
     test_CompareTwoOpts.cpp)

paddle/trainer/tests/sample_trainer_config_simple_net.conf

+# Copyright (c) 2017 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.
+
+from paddle.trainer_config_helpers import *
+
+################################### Data Configuration ###################################
+TrainData(ProtoData(files = "trainer/tests/mnist.list"))
+################################### Algorithm Configuration ###################################
+settings(batch_size = 1000,
+         learning_method = MomentumOptimizer(momentum=0.5, sparse=False))
+################################### Network Configuration ###################################
+data = data_layer(name ="input", size=784)
+
+tmp = img_conv_layer(input=data,
+            num_channels=1,
+            filter_size=3,
+            num_filters=32,
+            padding=1,
+            shared_biases=True,
+            act=ReluActivation())
+
+tmp = img_pool_layer(input=tmp,
+            pool_size=3,
+            stride=2,
+            padding=1,
+            pool_type=AvgPooling())
+            
+tmp = img_conv_layer(input=tmp,
+            filter_size=3,
+            num_filters=64,
+            padding=1,
+            shared_biases=True,
+            act=ReluActivation())
+
+tmp = img_pool_layer(input=tmp,
+            pool_size=3,
+            stride=2,
+            padding=1,
+            pool_type=MaxPooling())
+            
+tmp = fc_layer(input=tmp, size=64,
+               bias_attr=True,
+               act=ReluActivation())
+
+output = fc_layer(input=tmp, size=10,
+                  bias_attr=True,
+                  act=SoftmaxActivation())
+
+lbl = data_layer(name ="label", size=10)
+
+cost = classification_cost(input=output, label=lbl)
+outputs(cost)

paddle/trainer/tests/test_CompareTwoNets.cpp

@@ -26,12 +26,15 @@ DECLARE_int32(gpu_id);
 
 DECLARE_bool(local);
 DECLARE_bool(use_gpu);
+DECLARE_bool(use_mkldnn);
 
 DECLARE_string(config);
 DECLARE_string(nics);
 
 DEFINE_string(config_file_a, "", "config of one network to compare");
 DEFINE_string(config_file_b, "", "config of another network to compare");
+DEFINE_bool(use_mkldnn_a, false, "whether to use mkldnn to run network");
+DEFINE_bool(use_mkldnn_b, false, "whether to use mkldnn to run network");
 DEFINE_bool(need_high_accuracy,
             false,
             "whether need to run in double accuracy");
@@ -128,6 +131,12 @@ void compareGradient(ComData& comDataA, ComData& comDataB) {
                 matA.getWidth());
   }
 
+  if (FLAGS_use_mkldnn_a || FLAGS_use_mkldnn_b) {
+    // some format of mkldnn parameter is different with cpu
+    // test_MKLDNN will check the parameters
+    return;
+  }
+
   vector<ParameterPtr>& parametersA = comDataA.parameters;
   vector<ParameterPtr>& parametersB = comDataB.parameters;
 
@@ -167,10 +176,12 @@ void compareGradient(ComData& comDataA, ComData& comDataB) {
 
 TEST(Trainer, create) {
   ComData dataA;
+  FLAGS_use_mkldnn = FLAGS_use_mkldnn_a;
   calcGradient(dataA, FLAGS_config_file_a);
   LOG(INFO) << "\n\nforwardBackward of Network A is finished\n\n";
 
   ComData dataB;
+  FLAGS_use_mkldnn = FLAGS_use_mkldnn_b;
   calcGradient(dataB, FLAGS_config_file_b);
   LOG(INFO) << "\n\nforwardBackward of the Network B is finished\n\n";