Merge pull request #11666 from mozga-intel/mozga-intel/Batch_norm_support_other_type

The mkldnn batch norm supports other data format

paddle/fluid/framework/data_layout_transform.cc

@@ -147,9 +147,9 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
                  "Input tensor type is not supported: ", in.type().name());
   memory::data_type out_type = in_type;
 
-  auto in_format = MKLDNNFormatForSize(in_tz.size(), in.format());
+  auto in_format = platform::MKLDNNFormatForSize(in_tz.size(), in.format());
   auto out_format =
-      MKLDNNFormatForSize(in_tz.size(), ToMKLDNNFormat(out_layout));
+      platform::MKLDNNFormatForSize(in_tz.size(), ToMKLDNNFormat(out_layout));
 
   void* in_data = GetDataFromTensor(in, in_type);
 

paddle/fluid/framework/data_layout_transform.h

@@ -62,12 +62,6 @@ inline MKLDNNDataType ToMKLDNNDataType(const std::type_index type) {
   return MKLDNNDataType::data_undef;
 }
 
-inline MKLDNNFormat MKLDNNFormatForSize(size_t dims_size,
-                                        MKLDNNFormat default_format) {
-  return (dims_size == 1
-              ? mkldnn::memory::format::x
-              : dims_size == 2 ? mkldnn::memory::format::nc : default_format);
-}
 #endif
 
 void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,

paddle/fluid/framework/data_transform.cc

@@ -18,6 +18,10 @@ limitations under the License. */
 #include "paddle/fluid/framework/data_layout_transform.h"
 #include "paddle/fluid/framework/data_type_transform.h"
 
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
+
 namespace paddle {
 namespace framework {
 
@@ -48,8 +52,8 @@ void TransformData(const OpKernelType &expected_kernel_type,
         // Case1 - transform from Non-MKLDNN OPKernel to MKLDNN OPKernel
         // Just set layout/format. No real transform occur
 
-        auto out_format =
-            MKLDNNFormatForSize(in.dims().size(), ToMKLDNNFormat(lin));
+        auto out_format = platform::MKLDNNFormatForSize(in.dims().size(),
+                                                        ToMKLDNNFormat(lin));
 
         out.ShareDataWith(input_tensor);
         out.set_layout(DataLayout::kMKLDNN);

paddle/fluid/operators/batch_norm_mkldnn_op.cc

@@ -115,9 +115,12 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     if (fuse_with_relu) flags |= mkldnn::fuse_bn_relu;
 
     // create mkldnn memory from input x tensor
-    auto src_memory =
-        memory({{{src_tz}, memory::data_type::f32, x->format()}, mkldnn_engine},
-               to_void_cast(x_data));
+    mkldnn::memory::format input_format =
+        platform::MKLDNNFormatForSize(src_tz.size(), x->format());
+
+    auto src_memory = memory(
+        {{{src_tz}, memory::data_type::f32, input_format}, mkldnn_engine},
+        to_void_cast(x_data));
 
     // create primitive descriptor for batch norm forward
     using bn_fwd_types = bn_type_traits<mkldnn::batch_normalization_forward>;
@@ -251,15 +254,21 @@ class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
     using bn_bwd_types = bn_type_traits<mkldnn::batch_normalization_backward>;
 
     // create mkldnn memory from input diff_y tensor
-    auto user_diff_dst_memory =
-        memory({{{diff_dst_tz}, memory::data_type::f32, diff_y->format()},
-                mkldnn_engine},
-               to_void_cast(diff_y_data));
+
+    mkldnn::memory::format dst_format =
+        platform::MKLDNNFormatForSize(src_tz.size(), diff_y->format());
+
+    auto user_diff_dst_memory = memory(
+        {{{diff_dst_tz}, memory::data_type::f32, dst_format}, mkldnn_engine},
+        to_void_cast(diff_y_data));
 
     // create mkldnn memory from input x tensor
-    auto src_memory =
-        memory({{{src_tz}, memory::data_type::f32, x->format()}, mkldnn_engine},
-               to_void_cast(x_data));
+    mkldnn::memory::format input_format =
+        platform::MKLDNNFormatForSize(src_tz.size(), x->format());
+
+    auto src_memory = memory(
+        {{{src_tz}, memory::data_type::f32, input_format}, mkldnn_engine},
+        to_void_cast(x_data));
 
     // for diff_dst, try to use same format as dst in forward pass
     auto diff_dst_pd = batch_norm_fwd_pd.get()->dst_primitive_desc();

paddle/fluid/platform/mkldnn_helper.h

@@ -228,7 +228,7 @@ class MKLDNNHandler {
       return dstr;
     };
     return dims2str(operand_dims) + suffix;
-  };
+  }
 
  protected:
   const MKLDNNDeviceContext& dev_ctx_;
@@ -237,5 +237,15 @@ class MKLDNNHandler {
   bool is_reusing_;
 };
 
+inline mkldnn::memory::format MKLDNNFormatForSize(
+    size_t dims_size, mkldnn::memory::format data_format) {
+  if (dims_size == 1) {
+    return mkldnn::memory::format::x;
+  } else if (dims_size == 2) {
+    return mkldnn::memory::format::nc;
+  }
+  return data_format;
+}
+
 }  // namespace platform
 }  // namespace paddle