Merge pull request #10 from kunkunblueberry/tmp

fix the linking question of lacking function of hip

Author: dyzheng

source/module_base/module_device/rocm/memory_op.hip.cu

@@ -60,6 +60,19 @@ void set_memory_op<FPTYPE, base_device::DEVICE_GPU>::operator()(const base_devic
     hipErrcheck(hipMemset(arr, var, sizeof(FPTYPE) * size));
 }
 
+template <typename FPTYPE>
+void set_memory_2d_op<FPTYPE, base_device::DEVICE_GPU>::operator()(const base_device::DEVICE_GPU* dev,
+                                                                   FPTYPE* arr,
+                                                                   const size_t pitch,
+                                                                   const int var,
+                                                                   const size_t width,
+                                                                   const size_t height)
+{
+    hipErrcheck(hipMemset2D(arr, sizeof(FPTYPE) * pitch , var, sizeof(FPTYPE) * width, height));
+}
+
+
+
 template <typename FPTYPE>
 void synchronize_memory_op<FPTYPE, base_device::DEVICE_CPU, base_device::DEVICE_GPU>::operator()(
     const base_device::DEVICE_CPU* dev_out,
@@ -93,6 +106,49 @@ void synchronize_memory_op<FPTYPE, base_device::DEVICE_GPU, base_device::DEVICE_
     hipErrcheck(hipMemcpy(arr_out, arr_in, sizeof(FPTYPE) * size, hipMemcpyDeviceToDevice));
 }
 
+template <typename FPTYPE>
+void synchronize_memory_2d_op<FPTYPE, base_device::DEVICE_CPU, base_device::DEVICE_GPU>::operator()(
+    const base_device::DEVICE_CPU* dev_out,
+    const base_device::DEVICE_GPU* dev_in,
+    FPTYPE* arr_out,
+    const size_t dpitch,
+    const FPTYPE* arr_in,
+    const size_t spitch,
+    const size_t width,
+    const size_t height)
+{
+    hipErrcheck(hipMemcpy2D(arr_out, dpitch * sizeof(FPTYPE), arr_in, spitch * sizeof(FPTYPE), width * sizeof(FPTYPE), height, hipMemcpyDeviceToHost));
+}
+
+template <typename FPTYPE>
+void synchronize_memory_2d_op<FPTYPE, base_device::DEVICE_GPU, base_device::DEVICE_CPU>::operator()(
+    const base_device::DEVICE_GPU* dev_out,
+    const base_device::DEVICE_CPU* dev_in,
+    FPTYPE* arr_out,
+    const size_t dpitch,
+    const FPTYPE* arr_in,
+    const size_t spitch,
+    const size_t width,
+    const size_t height)
+{
+    hipErrcheck(hipMemcpy2D(arr_out, dpitch * sizeof(FPTYPE), arr_in, spitch * sizeof(FPTYPE), width * sizeof(FPTYPE), height, hipMemcpyHostToDevice));
+}
+
+template <typename FPTYPE>
+void synchronize_memory_2d_op<FPTYPE, base_device::DEVICE_GPU, base_device::DEVICE_GPU>::operator()(
+    const base_device::DEVICE_GPU* dev_out,
+    const base_device::DEVICE_GPU* dev_in,
+    FPTYPE* arr_out,
+    const size_t dpitch,
+    const FPTYPE* arr_in,
+    const size_t spitch,
+    const size_t width,
+    const size_t height)
+{
+    hipErrcheck(hipMemcpy2D(arr_out, dpitch * sizeof(FPTYPE), arr_in, spitch * sizeof(FPTYPE), width * sizeof(FPTYPE), height, hipMemcpyDeviceToDevice));
+}
+
+
 template <typename FPTYPE_out, typename FPTYPE_in>
 struct cast_memory_op<FPTYPE_out, FPTYPE_in, base_device::DEVICE_GPU, base_device::DEVICE_GPU> {
     void operator()(const base_device::DEVICE_GPU* dev_out,
@@ -183,6 +239,13 @@ template struct set_memory_op<double, base_device::DEVICE_GPU>;
 template struct set_memory_op<std::complex<float>, base_device::DEVICE_GPU>;
 template struct set_memory_op<std::complex<double>, base_device::DEVICE_GPU>;
 
+template struct set_memory_2d_op<int, base_device::DEVICE_GPU>;
+template struct set_memory_2d_op<float, base_device::DEVICE_GPU>;
+template struct set_memory_2d_op<double, base_device::DEVICE_GPU>;
+template struct set_memory_2d_op<std::complex<float>, base_device::DEVICE_GPU>;
+template struct set_memory_2d_op<std::complex<double>, base_device::DEVICE_GPU>;
+
+
 template struct synchronize_memory_op<int, base_device::DEVICE_CPU, base_device::DEVICE_GPU>;
 template struct synchronize_memory_op<int, base_device::DEVICE_GPU, base_device::DEVICE_CPU>;
 template struct synchronize_memory_op<int, base_device::DEVICE_GPU, base_device::DEVICE_GPU>;
@@ -199,6 +262,23 @@ template struct synchronize_memory_op<std::complex<double>, base_device::DEVICE_
 template struct synchronize_memory_op<std::complex<double>, base_device::DEVICE_GPU, base_device::DEVICE_CPU>;
 template struct synchronize_memory_op<std::complex<double>, base_device::DEVICE_GPU, base_device::DEVICE_GPU>;
 
+template struct synchronize_memory_2d_op<int, base_device::DEVICE_CPU, base_device::DEVICE_GPU>;
+template struct synchronize_memory_2d_op<int, base_device::DEVICE_GPU, base_device::DEVICE_CPU>;
+template struct synchronize_memory_2d_op<int, base_device::DEVICE_GPU, base_device::DEVICE_GPU>;
+template struct synchronize_memory_2d_op<float, base_device::DEVICE_CPU, base_device::DEVICE_GPU>;
+template struct synchronize_memory_2d_op<float, base_device::DEVICE_GPU, base_device::DEVICE_CPU>;
+template struct synchronize_memory_2d_op<float, base_device::DEVICE_GPU, base_device::DEVICE_GPU>;
+template struct synchronize_memory_2d_op<double, base_device::DEVICE_CPU, base_device::DEVICE_GPU>;
+template struct synchronize_memory_2d_op<double, base_device::DEVICE_GPU, base_device::DEVICE_CPU>;
+template struct synchronize_memory_2d_op<double, base_device::DEVICE_GPU, base_device::DEVICE_GPU>;
+template struct synchronize_memory_2d_op<std::complex<float>, base_device::DEVICE_CPU, base_device::DEVICE_GPU>;
+template struct synchronize_memory_2d_op<std::complex<float>, base_device::DEVICE_GPU, base_device::DEVICE_CPU>;
+template struct synchronize_memory_2d_op<std::complex<float>, base_device::DEVICE_GPU, base_device::DEVICE_GPU>;
+template struct synchronize_memory_2d_op<std::complex<double>, base_device::DEVICE_CPU, base_device::DEVICE_GPU>;
+template struct synchronize_memory_2d_op<std::complex<double>, base_device::DEVICE_GPU, base_device::DEVICE_CPU>;
+template struct synchronize_memory_2d_op<std::complex<double>, base_device::DEVICE_GPU, base_device::DEVICE_GPU>;
+
+
 template struct cast_memory_op<float, float, base_device::DEVICE_GPU, base_device::DEVICE_GPU>;
 template struct cast_memory_op<double, double, base_device::DEVICE_GPU, base_device::DEVICE_GPU>;
 template struct cast_memory_op<float, double, base_device::DEVICE_GPU, base_device::DEVICE_GPU>;
@@ -271,4 +351,4 @@ template struct delete_memory_op<std::complex<float>*, base_device::DEVICE_GPU>;
 template struct delete_memory_op<std::complex<double>*, base_device::DEVICE_GPU>;
 
 } // namespace memory
-} // end of namespace base_device
+} // end of namespace base_device