fix the issue when call contiguous() if one tensor's format is channels_last, but it is not contiguous (#33)

* fix the issue when call contiguous() if one tensor's format is channels_last, but it is not contiguous

Author: XiaobingSuper

torch_ipex/csrc/cpu/Conv.cpp

@@ -34,7 +34,7 @@ at::Tensor convolution_impl(
     const ideep::attr_t& attr) {
 // TODO: the input will be actively converted to channels last format
 // after the 5-D tensor supports channels last format.
-  auto input_ = IS_CONTIGUOUS_ANY(input) ? input : input.contiguous();
+  auto input_ = IS_CONTIGUOUS_ANY(input) ? input : input.contiguous(input.suggest_memory_format());
   const ideep::tensor mkldnn_input = at::native::itensor_view_from_dense(input_);
   ideep::tensor mkldnn_weight = get_conv_prepacked_weight(mkldnn_input, weight, stride, padding, dilation, groups, attr);
   auto kernel_size = mkldnn_weight.get_dims();
@@ -105,7 +105,7 @@ void convolution_inplace_impl(
     const ideep::attr_t& attr) {
 // TODO: the input will be actively converted to channels last format
 // after the 5-D tensor supports channels last format.
-  auto input_ = IS_CONTIGUOUS_ANY(input) ? input : input.contiguous();
+  auto input_ = IS_CONTIGUOUS_ANY(input) ? input : input.contiguous(input.suggest_memory_format());
   const ideep::tensor mkldnn_input = at::native::itensor_view_from_dense(input_);
   ideep::tensor mkldnn_weight = get_conv_prepacked_weight(mkldnn_input, weight, stride, padding, dilation, groups, attr);
   auto kernel_size = mkldnn_weight.get_dims();
@@ -114,7 +114,7 @@ void convolution_inplace_impl(
       calc_conv_output_size(input_size, kernel_size, padding, stride, dilation);
 
   bool is_channels_last = input_.suggest_memory_format() == at::MemoryFormat::ChannelsLast;
-  output = IS_CONTIGUOUS_ANY(output) ? output : output.contiguous();
+  output = IS_CONTIGUOUS_ANY(output) ? output : output.contiguous(output.suggest_memory_format());
   output = output.to(input_.suggest_memory_format());
   ideep::tensor mkldnn_output = at::native::itensor_view_from_dense(output);
 

torch_ipex/csrc/cpu/Pooling.cpp

@@ -73,7 +73,7 @@ at::Tensor pooling_impl(
 
 // TODO: the input will be actively converted to channels last format
 // after the 5-D tensor supports channels last format.
-  auto input_ = IS_CONTIGUOUS_ANY(input) ? input : input.contiguous();
+  auto input_ = IS_CONTIGUOUS_ANY(input) ? input : input.contiguous(input.suggest_memory_format());
   const ideep::tensor mkldnn_input = at::native::itensor_view_from_dense(input_);
   std::vector<int64_t> output_sizes;
 

torch_ipex/csrc/cpu/WeightPrepack.cpp

@@ -25,7 +25,7 @@ ideep::tensor get_conv_prepacked_weight(
   if (it != cached_weights.end()) {
     return std::get<1>(it->second);
   } else { 
-    auto weight_ = IS_CONTIGUOUS_ANY(weight) ? weight : weight.contiguous();
+    auto weight_ = IS_CONTIGUOUS_ANY(weight) ? weight : weight.contiguous(weight.suggest_memory_format());
     ideep::tensor w = at::native::itensor_view_from_dense(weight_);
     // TODO: 3d check
     bool is_channels_last = input.get_desc().is_nhwc();

torch_ipex/csrc/cpu/WeightPrepack.h

@@ -10,9 +10,9 @@ namespace torch_ipex {
 namespace cpu {
 
 // Get the convolution's expected ideep weight tensor, may be a block weight.
-// if the expected weight doesn't not exist, it will create an expected weight according
+// if the expected weight doesn't exist, it will create an expected weight according
 // to the queried desc of OneDNN conv, and the expected weight will be cached.
-// TODO: if weight is a block weight, direct return the weight(ideep tensor).
+// TODO: if weight is already in the expected format, return the weight(ideep tensor) directly.
 
 // input: an ideep tensor, getting from the convolution's input,
 // weight: convolution's weight
@@ -28,9 +28,9 @@ ideep::tensor get_conv_prepacked_weight(
     const ideep::attr_t& attr);
 
 // Get the linear's expected ideep weight tensor, may be a block weight.
-// if the expected weight doesn't not exist, it will create an expected weight according
+// if the expected weight doesn't exist, it will create an expected weight according
 // to the queried desc of OneDNN linear, and the expected weight will be cached.
-// TODO: if weight is a block weight, direct return the weight(ideep tensor).
+// TODO: if weight is already in the expected format, return the weight(ideep tensor) directly.
 
 // input: an ideep tensor, getting from the linear's input,
 // weight: linear's weight