C++ interface "NNPInter"

[WangXinyan940]

I wrote simple test cases for both Python and C++ interface.

Python version:

import deepmd.DeepPot as DP
import numpy as np

dp = DP('graph.pb')
pos = np.array([1., 1., 1., 2., 1., 1., 1., 2., 1., 4., 4., 4., 5., 4., 4., 4., 5., 4.]).reshape([1, -1])
box = np.diag(99.9 * np.ones(3)).reshape([1, -1])
atype = [0, 1, 1, 0, 1, 1]
e, f, v = dp.eval(pos, box, atype)
print(e)

C++ version:

#include "NNPInter.h"
#include <vector>
#include <iostream>

using namespace std;
typedef double VALUETYPE2;

int main(){
    vector<int> types{0, 1, 1, 0, 1, 1};
    vector<VALUETYPE2> positions{1, 1, 1, 2, 1, 1, 1, 2, 1, 4, 4, 4, 5, 4, 4, 4, 5, 4};
    vector<VALUETYPE2> box{99.9, 0, 0, 0, 99.9, 0, 0, 0, 99.9};
    vector<VALUETYPE2> forces(18, 0);
    vector<VALUETYPE2> virial(9, 0);
    double ener = 0.0;
    NNPInter model("graph.pb");
    model.compute(ener, forces, virial, positions, types, box);
    cout << ener << endl;
    return 0;
}

As shown, my input coordinates, box vectors and atom types are totally the same, but the C++ case failed while Python case succeeded. The error is:

2020-12-18 20:47:24.501720: E tensorflow/stream_executor/cuda/cuda_blas.cc:428] failed to run cuBLAS routine: CUBLAS_STATUS_EXECUTION_FAILED
Internal: 2 root error(s) found.
  (0) Internal: Blas GEMM launch failed : a.shape=(92, 1), b.shape=(1, 25), m=92, n=25, k=1
         [[{{node filter_type_0/MatMul}}]]
         [[o_energy/_25]]
  (1) Internal: Blas GEMM launch failed : a.shape=(92, 1), b.shape=(1, 25), m=92, n=25, k=1
         [[{{node filter_type_0/MatMul}}]]
0 successful operations.
0 derived errors ignored.

Did I used the C++ interface in a wrong way?

I compiled the whole things with tensorflow==1.14, gcc==5.3, cudatookit==10.1 and the latest deepmd-kit code from master branch. The deepmd-kit library is compiled with double precision.

The freezed model I used is attached here:
graph.zip