fill rij with zero (deepmodeling#818)

fix deepmodeling#765.

Author: njzjz

source/lib/src/cuda/prod_env_mat.cu

@@ -502,6 +502,7 @@ void prod_env_mat_a_gpu_cuda(
   const int ndescrpt = nnei * 4;
   cudaErrcheck(cudaMemset(em, 0.0, sizeof(FPTYPE) * nloc * ndescrpt));
   cudaErrcheck(cudaMemset(em_deriv, 0.0, sizeof(FPTYPE) * nloc * ndescrpt * 3));
+  cudaErrcheck(cudaMemset(rij, 0., sizeof(FPTYPE) * nloc * nnei * 3));
 
   format_nbor_list_gpu_cuda(
       nlist, 
@@ -538,6 +539,7 @@ void prod_env_mat_r_gpu_cuda(
   const int ndescrpt = nnei * 1;
   cudaErrcheck(cudaMemset(em, 0.0, sizeof(FPTYPE) * nloc * ndescrpt));
   cudaErrcheck(cudaMemset(em_deriv, 0.0, sizeof(FPTYPE) * nloc * ndescrpt * 3));
+  cudaErrcheck(cudaMemset(rij, 0., sizeof(FPTYPE) * nloc * nnei * 3));
 
   format_nbor_list_gpu_cuda(
       nlist, 

source/lib/src/rocm/prod_env_mat.hip.cu

@@ -504,6 +504,7 @@ void prod_env_mat_a_gpu_rocm(
   const int ndescrpt = nnei * 4;
   hipErrcheck(hipMemset(em, 0.0, sizeof(FPTYPE) * nloc * ndescrpt));
   hipErrcheck(hipMemset(em_deriv, 0.0, sizeof(FPTYPE) * nloc * ndescrpt * 3));
+  hipErrcheck(hipMemset(rij, 0., sizeof(FPTYPE) * nloc * nnei * 3));
 
   format_nbor_list_gpu_rocm(
       nlist, 
@@ -540,6 +541,7 @@ void prod_env_mat_r_gpu_rocm(
   const int ndescrpt = nnei * 1;
   hipErrcheck(hipMemset(em, 0.0, sizeof(FPTYPE) * nloc * ndescrpt));
   hipErrcheck(hipMemset(em_deriv, 0.0, sizeof(FPTYPE) * nloc * ndescrpt * 3));
+  hipErrcheck(hipMemset(rij, 0., sizeof(FPTYPE) * nloc * nnei * 3));
 
   format_nbor_list_gpu_rocm(
       nlist,