Merge branch 'fix-transpose_before_shape_on_gpu' into 'master'

remove transpose half and add transpose to buffer before shape() for gpu runtime.

See merge request !1123

mace/core/net_def_adapter.cc

@@ -50,15 +50,6 @@ std::string TransformedName(const std::string &input_name,
   return ss.str();
 }
 
-#ifdef MACE_ENABLE_OPENCL
-bool TransformRequiredOp(const std::string &op_type) {
-  static const std::unordered_set<std::string> kNoTransformOp = {
-      "Shape", "InferConv2dShape"
-  };
-  return kNoTransformOp.count(op_type) == 0;
-}
-#endif  // MACE_ENABLE_OPENCL
-
 void BuildTransposeOpDef(
     const std::string &input_name,
     const std::string &output_name,
@@ -514,8 +505,6 @@ MaceStatus NetDefAdapter::AdaptMemoryType(
   // (only support one kind of memory type for multiple outputs)
   op_registry_->GetInOutMemoryTypes(op_def->type(), context);
 #ifdef MACE_ENABLE_OPENCL
-  // if op is memory-unused op, no transformation
-  if (TransformRequiredOp(op_def->type())) {
   int input_size = op_def->input_size();
   for (int i = 0; i < input_size; ++i) {
     if (output_map->count(op_def->input(i)) == 0) {
@@ -585,7 +574,6 @@ MaceStatus NetDefAdapter::AdaptMemoryType(
       op_def->set_input(i, transformed_name);
     }
   }
-  }
 #else
   MACE_UNUSED(output_map);
   MACE_UNUSED(tensor_shape_map);

mace/ops/transpose.cc

@@ -27,7 +27,10 @@ namespace mace {
 namespace ops {
 
 template<DeviceType D, typename T>
-class TransposeOp : public Operation {
+class TransposeOp;
+
+template<DeviceType D>
+class TransposeOp<D, float> : public Operation {
  public:
   explicit TransposeOp(OpConstructContext *context)
       : Operation(context),
@@ -49,8 +52,8 @@ class TransposeOp : public Operation {
 
     Tensor::MappingGuard input_guard(input);
     Tensor::MappingGuard output_guard(output);
-    const T *input_data = input->data<T>();
-    T *output_data = output->mutable_data<T>();
+    const float *input_data = input->data<float>();
+    float *output_data = output->mutable_data<float>();
 
     return Transpose(&context->device()->cpu_runtime()->thread_pool(),
                      input_data, input->shape(), dims_, output_data);
@@ -63,8 +66,6 @@ class TransposeOp : public Operation {
 void RegisterTranspose(OpRegistryBase *op_registry) {
   MACE_REGISTER_OP(op_registry, "Transpose", TransposeOp,
                    DeviceType::CPU, float);
-  MACE_REGISTER_OP(op_registry, "Transpose", TransposeOp,
-                   DeviceType::CPU, half);
 }
 
 }  // namespace ops