switch to unique_tpr

Author: haojin2

src/operator/operator_tune-inl.h

@@ -116,10 +116,6 @@ class OperatorTune : public OperatorTuneByType<DType> {
     TuneAll();
   }
 
-  ~OperatorTune() {
-    delete[] data_set_;
-  }
-
   /*!
    * \brief Initialize the OperatorTune object
    * \return Whether the OperatorTune object was successfully initialized
@@ -128,7 +124,8 @@ class OperatorTune : public OperatorTuneByType<DType> {
     if (!initialized_) {
       initialized_ = true;
       // Generate some random data for calling the operator kernels
-      data_set_ = reinterpret_cast<DType*>(new char[0x100 * sizeof(DType)]);
+      data_set_ =
+        std::unique_ptr<DType[]>(reinterpret_cast<DType*>(new char[0x100 * sizeof(DType)]));
       std::random_device rd;
       std::mt19937 gen(rd());
       if (!std::is_integral<DType>::value) {
@@ -521,7 +518,7 @@ class OperatorTune : public OperatorTuneByType<DType> {
   /*! \brief Number of passes to obtain an average */
   static constexpr duration_t OUTSIDE_COUNT = (1 << OUTSIDE_COUNT_SHIFT);
   /*! \brief Random data for timing operator calls */
-  static DType* data_set_;
+  static std::unique_ptr<DType[]> data_set_;
   /*! \brief Operators tuned */
   static std::unordered_set<std::string> operator_names_;
   /*! \brief Arbitary object to modify in OMP loop */

src/operator/operator_tune.cc

@@ -39,7 +39,7 @@ double OperatorTuneBase::tuning_weight_scale_ = 0.0;
  */
 #define IMPLEMENT_OPERATOR_TUNE_STATICS_FOR_TYPE(__typ$) \
   template<> bool OperatorTune<__typ$>::initialized_ = false; \
-  template<> __typ$* OperatorTune<__typ$>::data_set_ = nullptr; \
+  template<> std::unique_ptr<__typ$[]> OperatorTune<__typ$>::data_set_ = nullptr; \
   template<> volatile tune::TuningMode OperatorTuneByType<__typ$>::tuning_mode_ = tune::kAuto; \
   template<> volatile int OperatorTune<__typ$>::volatile_int_ = 9;  /* arbitrary number */ \
   template<> std::unordered_set<std::string> OperatorTune<__typ$>::operator_names_({}); \

src/operator/tensor/elemwise_binary_scalar_op.h

@@ -236,7 +236,7 @@ class BinaryScalarOp : public UnaryOp {
     using namespace mshadow::expr;
     Stream<xpu> *s = ctx.get_stream<xpu>();
     const double alpha = nnvm::get<double>(attrs.parsed);
-    MSHADOW_TYPE_SWITCH(outputs[0].type_flag_, DType, {
+    MSHADOW_TYPE_SWITCH_WITH_BOOL(outputs[0].type_flag_, DType, {
       MXNET_ASSIGN_REQ_SWITCH(req[0], Req, {
         mxnet_op::Kernel<mxnet_op::op_with_req<OP, Req>, xpu>::Launch(
           s, inputs[0].Size(), outputs[0].dptr<DType>(), inputs[0].dptr<DType>(), DType(alpha));

tests/python/unittest/test_numpy_op.py

@@ -1669,7 +1669,7 @@ def hybrid_forward(self, F, a, b, *args, **kwargs):
         mx_test_x1 = mx.numpy.array(np_test_x1, dtype=ltype)
         mx_test_x2 = mx.numpy.array(np_test_x2, dtype=rtype)
         rtol = 1e-2 if ltype is np.float16 or rtype is np.float16 else 1e-3
-        atol = 1e-4 if ltype is np.float16 or rtype is np.float16 else 1e-5
+        atol = 1e-3 if ltype is np.float16 or rtype is np.float16 else 1e-5
         for hybridize in [True, False]:
             if hybridize:
                 mx_func.hybridize()