Fix meme leak bug, add result expect

mindspore/ccsrc/backend/kernel_compiler/gpu/math/eigh_c_gpu_kernel.h

@@ -114,10 +114,14 @@ class EighcGpuKernel : public GpuKernel {
     cudaMalloc(reinterpret_cast<void **>(&dev_input_shape), kShape2dDims * sizeof(size_t));
     size_t *dev_input_axis = nullptr;
     cudaMalloc(reinterpret_cast<void **>(&dev_input_axis), kShape2dDims * sizeof(size_t));
-    cudaMemcpyAsync(dev_input_shape, input_shape, kShape2dDims * sizeof(size_t), cudaMemcpyHostToDevice,
-                    reinterpret_cast<cudaStream_t>(stream_ptr));
-    cudaMemcpyAsync(dev_input_axis, input_axis, kShape2dDims * sizeof(size_t), cudaMemcpyHostToDevice,
-                    reinterpret_cast<cudaStream_t>(stream_ptr));
+    CHECK_CUDA_RET_WITH_EXCEPT(kernel_node_,
+                               cudaMemcpyAsync(dev_input_shape, input_shape, kShape2dDims * sizeof(size_t),
+                                               cudaMemcpyHostToDevice, reinterpret_cast<cudaStream_t>(stream_ptr)),
+                               "malloc input shape workspace failed");
+    CHECK_CUDA_RET_WITH_EXCEPT(kernel_node_,
+                               cudaMemcpyAsync(dev_input_axis, input_axis, kShape2dDims * sizeof(size_t),
+                                               cudaMemcpyHostToDevice, reinterpret_cast<cudaStream_t>(stream_ptr)),
+                               "malloc input shape workspace failed");
     CalTranspose(m_ * m_, output_v_addr, dev_input_shape, dev_input_axis, kShape2dDims, w_v_addr,
                  reinterpret_cast<cudaStream_t>(stream_ptr));
 
@@ -145,6 +149,12 @@ class EighcGpuKernel : public GpuKernel {
                   reinterpret_cast<cudaStream_t>(stream_ptr));
     CalTranspose(m_ * m_, w_v_addr, dev_input_shape, dev_input_axis, kShape2dDims, output_v_addr,
                  reinterpret_cast<cudaStream_t>(stream_ptr));
+    if (dev_input_shape) {
+      cudaFree(dev_input_shape);
+    }
+    if (dev_input_axis) {
+      cudaFree(dev_input_axis);
+    }
     // convert real scalar to complex
     if (d_work) {
       cudaFree(d_work);