Using exprtk in a multithreaded program

Question

I need to write a program in which string expressions are evaluated quite frequently. An example of an expression is below:

"x0*a0*a0+x1*a1+x2*a2+x3*a3+x4*a4....."

The expressions can be long and a string can contain multiple such expressions.

I wrote some test code using the C++ library exprtk.

vector<std::string> observation_functions;
vector<std::string> string_indices;


template<typename T>
float* get_observation(float* sing_j, float* zrlist, int num_functions,int num_variables)
{
    //omp_set_nested(1);

    float* results = (float*)malloc(sizeof(float)*num_functions);
    exprtk::symbol_table<float> symbol_table;

    exprtk::expression<T> expression;
    exprtk::parser<T> parser;
    int i;
    for( i = 0; i < num_variables; i++)
    {
            symbol_table.add_variable("x"+string_indices[i], sing_j[i]);
            symbol_table.add_variable("a"+string_indices[i], zrlist[i]);
    }

    expression.register_symbol_table(symbol_table);
    for(i = 0; i < num_functions; i++)
    {
            parser.compile(observation_functions[i],expression);
            results[i] = expression.value();
    }
    return results;
}



int main()
{

    for( int i = 0; i < 52; i++)
    {

    ostringstream s2;
    s2<<i;
    string_indices.push_back(s2.str());
    }



    string hfun ="x0*a0*a0+x1*a1+x2*a2+x3*a3+x4*a4+x5*a5+x6*a6+x7*a7+x8*a8+x9*a9+x10*a10+x11*a11+x12*a12+x13*a13+x14*a14+x15*a15+x16*a16+x17*a17+x18*a18+x19*a19+x20*a20+x21*a21+x22*a22+x23*a23+x24*a24+x25*a25+x26*a26+x27*a27+x28*a28+x29*a29+x30*a30+x31*a31+x32*a32+x33*a33+x34*a34+x35*a35+x36*a36+x37*a37+x38*a38+x39*a39+x40*a40+x41*a41+x42*a42+x43*a43+x44*a44+x45*a45+x46*a46+x47*a47+x48*a48+x49*a49+x50*a50+x51*a51 ";


    boost::split(observation_functions, hfun, boost::is_any_of(" "));
    float *a=(float*)malloc(52*sizeof(float));
    float* c=(float*)malloc(52*sizeof(float));

    struct timeval t0,t1;
    gettimeofday(&t0, 0);
    for(int j=0; j < 210; j++)
        #pragma omp parallel for schedule(static,1) num_threads(8)
        for(int i=0;i<104;i++)
            float* b =get_observation<float>(a,c,1,52);
    gettimeofday(&t1, 0);
    long elapsed = (t1.tv_sec-t0.tv_sec)*1000000 + t1.tv_usec-t0.tv_usec;
    cout<<"elapsed:"<<elapsed<<endl;

}   

Note that this is test code. In the actual, each thread is going to evaluate the expression with a different set of values. This code works fine, but I need to make it go faster.

Based on some other experiments, I found that I cannot share a single symbol table with multiple threads to compute a single expression faster. Sharing a symbol table among multiple threads led to memory corruption errors.

Can someone please provide some suggestions on how I could improve the performance.

Answer 1

Assume, you have N threads. Than, create N sets of exprtk -related objects (including symbol_table , expression , parser ) in the main program ( outside the function, and outside for loops).

You could use vector<> to store them: eg, for expression objects it would be vector<expression> expressions;

Then, pass the references to those objects, when calling your function,

for(int i=0;i<104;i++)
    get_observation<float>(expressions[i], more params here..)

Template function definition: template <typename T> T* get_observation(expression & exp, more params here..)

You could also create one set of objects and make other by copying, as Aloalo suggested .

PS Prefer to use smart pointers, https://stackoverflow.com/a/19042634 not to forget delete the memory allocated somewhere locally.

Answer 2

You can try building exprtk objects only once and make a copy of them for each thread. This should be faster if copying of exprtk objects is faster than constructing them.