diff --git a/CMakeLists.txt b/CMakeLists.txt
index cd31fc6..3dc6b57 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,5 +1,5 @@
 cmake_minimum_required(VERSION 3.11 FATAL_ERROR) # 3.11 to avoid issues with OpenMP + CUDA
-project(SpFFT LANGUAGES CXX VERSION 1.0.0)
+project(SpFFT LANGUAGES CXX VERSION 1.0.1)
 set(SPFFT_SO_VERSION 1)
 set(SPFFT_VERSION ${PROJECT_VERSION})
 
diff --git a/src/execution/execution_gpu.cpp b/src/execution/execution_gpu.cpp
index 0336916..05c7298 100644
--- a/src/execution/execution_gpu.cpp
+++ b/src/execution/execution_gpu.cpp
@@ -272,7 +272,7 @@ auto ExecutionGPU<T>::forward_xy(const T* input) -> void {
       gpu::check_status(gpu::memcpy_async(static_cast<void*>(spaceDomainDataExternalGPU_.data()),
                                           static_cast<const void*>(inputPtrHost),
                                           spaceDomainDataExternalGPU_.size() * sizeof(T),
-                                          gpu::flag::MemcpyDeviceToHost, stream_.get()));
+                                          gpu::flag::MemcpyHostToDevice, stream_.get()));
     }
     transformXY_->forward(inputPtrGPU, freqDomainXYGPU_.data());
   }
diff --git a/tests/mpi_tests/test_transpose_gpu.cpp b/tests/mpi_tests/test_transpose_gpu.cpp
index 1fb98d8..c67ba1d 100644
--- a/tests/mpi_tests/test_transpose_gpu.cpp
+++ b/tests/mpi_tests/test_transpose_gpu.cpp
@@ -109,8 +109,8 @@ static void check_space_domain(const HostArrayView3D<std::complex<double>>& real
   for (SizeType z = 0; z < numLocalXYPlanes; ++z) {
     for (SizeType x = 0; x < fullView.dim_outer(); ++x) {
       for (SizeType y = 0; y < fullView.dim_mid(); ++y) {
-        EXPECT_EQ(realView(z, x, y).real(), fullView(x, y, z + planeOffset).real());
-        EXPECT_EQ(realView(z, x, y).imag(), fullView(x, y, z + planeOffset).imag());
+        EXPECT_EQ(realView(z, y, x).real(), fullView(x, y, z + planeOffset).real());
+        EXPECT_EQ(realView(z, y, x).imag(), fullView(x, y, z + planeOffset).imag());
       }
     }
   }