| 1 | /* -*- mode: C++; tab-width: 8; indent-tabs-mode: t; c-basic-offset: 8 -*- */ |
|---|
| 2 | //-------------------------------------------------------------------------- |
|---|
| 3 | // Test for fgemv : 1 computation |
|---|
| 4 | // |
|---|
| 5 | //-------------------------------------------------------------------------- |
|---|
| 6 | // Clement Pernet |
|---|
| 7 | //------------------------------------------------------------------------- |
|---|
| 8 | |
|---|
| 9 | #define DEBUG 1 |
|---|
| 10 | #define TIME 1 |
|---|
| 11 | |
|---|
| 12 | #include <iomanip> |
|---|
| 13 | #include <iostream> |
|---|
| 14 | #include "fflas-ffpack/modular-balanced.h" |
|---|
| 15 | #include "timer.h" |
|---|
| 16 | #include "Matio.h" |
|---|
| 17 | #include "fflas-ffpack/fflas.h" |
|---|
| 18 | |
|---|
| 19 | |
|---|
| 20 | |
|---|
| 21 | using namespace std; |
|---|
| 22 | |
|---|
| 23 | typedef Modular<double> Field; |
|---|
| 24 | |
|---|
| 25 | int main(int argc, char** argv){ |
|---|
| 26 | |
|---|
| 27 | int m,n,k; |
|---|
| 28 | int nbit=atoi(argv[4]); // number of times the product is performed |
|---|
| 29 | cerr<<setprecision(10); |
|---|
| 30 | Field::Element alpha,beta; |
|---|
| 31 | |
|---|
| 32 | |
|---|
| 33 | if (argc != 8) { |
|---|
| 34 | cerr<<"Usage : test-fgemm <p> <A> <b> <i>" |
|---|
| 35 | <<" <alpha> <beta> <c>"<<endl |
|---|
| 36 | <<" to do i computations of c <- alpha AB + beta C" |
|---|
| 37 | <<endl; |
|---|
| 38 | exit(-1); |
|---|
| 39 | } |
|---|
| 40 | Field F(atoi(argv[1])); |
|---|
| 41 | |
|---|
| 42 | F.init( alpha, double(atoi(argv[5]))); |
|---|
| 43 | F.init( beta, double(atoi(argv[6]))); |
|---|
| 44 | |
|---|
| 45 | Field::Element * A; |
|---|
| 46 | Field::Element * b; |
|---|
| 47 | |
|---|
| 48 | b = read_field(F,argv[3],&n,&k); |
|---|
| 49 | A = read_field(F,argv[2],&m,&n); |
|---|
| 50 | |
|---|
| 51 | Field::Element * c; |
|---|
| 52 | c = new Field::Element[n]; |
|---|
| 53 | |
|---|
| 54 | |
|---|
| 55 | |
|---|
| 56 | Timer tim,t; t.clear();tim.clear(); |
|---|
| 57 | for(int i = 0;i<nbit;++i){ |
|---|
| 58 | c = read_field(F,argv[7],&m,&k); |
|---|
| 59 | t.clear(); |
|---|
| 60 | t.start(); |
|---|
| 61 | FFLAS::fgemv (F, FFLAS::FflasNoTrans,m,n,alpha, A,n, b,1, |
|---|
| 62 | beta, c, 1); |
|---|
| 63 | t.stop(); |
|---|
| 64 | tim+=t; |
|---|
| 65 | } |
|---|
| 66 | |
|---|
| 67 | #if DEBUG |
|---|
| 68 | Field::Element *d = new Field::Element[n]; |
|---|
| 69 | for (int i=0; i<m; ++i) |
|---|
| 70 | F.mul (d[i], beta, b[i]); |
|---|
| 71 | for (int i=0; i<m; ++i) |
|---|
| 72 | F.mulin (b[i], alpha); |
|---|
| 73 | for (int i=0; i<m; ++i) |
|---|
| 74 | for (int j=0; j<n; ++j) |
|---|
| 75 | F.axpyin (d[i], *(A+i*m+j), b[j]); |
|---|
| 76 | bool fail = false; |
|---|
| 77 | for (int i=0; i<m; ++i) |
|---|
| 78 | if (!F.areEqual(d[i], c[i])) |
|---|
| 79 | fail = true; |
|---|
| 80 | |
|---|
| 81 | if (fail) |
|---|
| 82 | cerr<<"FAIL"<<endl; |
|---|
| 83 | else |
|---|
| 84 | cerr<<"PASS"<<endl; |
|---|
| 85 | delete[] d; |
|---|
| 86 | #endif |
|---|
| 87 | delete[] A; |
|---|
| 88 | delete[] b; |
|---|
| 89 | delete[] c; |
|---|
| 90 | #if TIME |
|---|
| 91 | double mflops = (2.0*(m*n/1000000.0)*nbit/tim.usertime()); |
|---|
| 92 | cerr << m <<"x" <<n <<" : fgemv over Z/" |
|---|
| 93 | <<atoi(argv[1])<<"Z : [ " |
|---|
| 94 | <<mflops<<" MFops in "<<tim.usertime()/nbit<<"s]" |
|---|
| 95 | << endl; |
|---|
| 96 | |
|---|
| 97 | cerr<<"alpha, beta = "<<alpha <<", "<<beta <<endl; |
|---|
| 98 | |
|---|
| 99 | cout<<m<<" "<<n<<" "<<mflops<<" "<<tim.usertime()/nbit<<endl; |
|---|
| 100 | #endif |
|---|
| 101 | } |
|---|
| 102 | |
|---|