Remove CopyElem (#6591)

Co-authored-by: Noneflow-ci-bot <69100618+oneflow-ci-bot@users.noreply.github.com>

oneflow/core/kernel/kernel_util.h

@@ -85,18 +85,6 @@ template<typename T>
 struct KernelUtil<DeviceType::kGPU, T, typename std::enable_if<IsIntegral<T>::value>::type>
     : public GpuKernelUtilIf<T, KernelUtil<DeviceType::kGPU, T>> {};
 
-template<typename T, typename U>
-typename std::enable_if<std::is_same<T, U>::value>::type CopyElem(const T* in_dptr, U* out_dptr,
-                                                                  int64_t elem_num) {
-  Memcpy<DeviceType::kCPU>(nullptr, out_dptr, in_dptr, elem_num * sizeof(T));
-}
-
-template<typename T, typename U>
-typename std::enable_if<!std::is_same<T, U>::value>::type CopyElem(const T* in_dptr, U* out_dptr,
-                                                                   int64_t elem_num) {
-  FOR_RANGE(int64_t, i, 0, elem_num) { *(out_dptr++) = static_cast<U>(*(in_dptr++)); }
-}
-
 }  // namespace oneflow
 
 #endif  // ONEFLOW_CORE_KERNEL_KERNEL_UTIL_H_

oneflow/user/kernels/image_batch_align_kernel.cpp

@@ -35,7 +35,7 @@ void CopyFromTensorBuffer(T* image_ptr, const TensorBuffer& image_buffer, const
   FOR_RANGE(int, i, 0, h) {
     const F* from = image_buffer.data<F>() + i * w * c;
     T* to = image_ptr + i * batch_width * channels;
-    CopyElem(from, to, w * c);
+    std::transform(from, from + w * c, to, [](F v) { return static_cast<T>(v); });
   }
 }
 

oneflow/user/kernels/ofrecord_decoder_kernels.cpp

@@ -40,7 +40,8 @@ void DecodeOneRawOFRecord(const Feature& feature, T* dptr, int64_t sample_elem_c
     const auto& value0 = feature.bytes_list().value(0);
     auto in_dptr = reinterpret_cast<const int8_t*>(value0.c_str());
     sample_elem_cnt = std::min<int64_t>(sample_elem_cnt, value0.size());
-    CopyElem<int8_t, T>(in_dptr, dptr, sample_elem_cnt);
+    std::transform(in_dptr, in_dptr + sample_elem_cnt, dptr,
+                   [](int8_t v) { return static_cast<T>(v); });
   }
 #define DEFINE_ONE_ELIF(PbT, CppT)                                                       \
   else if (feature.has_##PbT##_list()) {                                                 \
@@ -56,7 +57,8 @@ void DecodeOneRawOFRecord(const Feature& feature, T* dptr, int64_t sample_elem_c
         CHECK_EQ(sample_elem_cnt, list.value_size());                                    \
       }                                                                                  \
     }                                                                                    \
-    CopyElem<CppT, T>(in_dptr, dptr, sample_elem_cnt);                                   \
+    std::transform(in_dptr, in_dptr + sample_elem_cnt, dptr,                             \
+                   [](CppT v) { return static_cast<T>(v); });                            \
     if (padding_elem_num > 0) {                                                          \
       std::memset(dptr + sample_elem_cnt, 0, padding_elem_num * sizeof(T));              \
     }                                                                                    \