| 1 | #include <iostream>
|
|---|
| 2 | #include <algorithm>
|
|---|
| 3 | using namespace std;
|
|---|
| 4 | #include <chrono>
|
|---|
| 5 | using namespace chrono;
|
|---|
| 6 | #include <uActor.h>
|
|---|
| 7 |
|
|---|
| 8 | unsigned int xr = 3'072, xc = 3'072, yc = 3'072, Processors = 1; // default values
|
|---|
| 9 |
|
|---|
| 10 | struct WorkMsg : public uActor::Message { // derived message
|
|---|
| 11 | int * Z;
|
|---|
| 12 | const int * const X, * const * Y;
|
|---|
| 13 | WorkMsg( int Z[], const int X[], const int * const Y[] ) :
|
|---|
| 14 | Message( uActor::Finished ), Z( Z ), X( X ), Y( Y ) {} // one-shot
|
|---|
| 15 | }; // WorkMsg
|
|---|
| 16 |
|
|---|
| 17 | time_point<steady_clock> starttime;
|
|---|
| 18 |
|
|---|
| 19 | _Actor MatrixMult {
|
|---|
| 20 | Allocation receive( Message & msg ) {
|
|---|
| 21 | Case ( WorkMsg, msg ) {
|
|---|
| 22 | int * z = msg_d->Z; // optimizations
|
|---|
| 23 | const int * const x = msg_d->X, * const * y = msg_d->Y;
|
|---|
| 24 | for ( unsigned int i = 0; i < yc; i += 1 ) { // multiply X_row by Y_col and sum products
|
|---|
| 25 | z[i] = 0;
|
|---|
| 26 | for ( unsigned int j = 0; j < xc; j += 1 ) {
|
|---|
| 27 | z[i] += x[j] * y[j][i];
|
|---|
| 28 | } // for
|
|---|
| 29 | } // for
|
|---|
| 30 | } // Case
|
|---|
| 31 |
|
|---|
| 32 | return Finished;
|
|---|
| 33 | } // MatrixMult:::receive
|
|---|
| 34 | }; // MatrixMult
|
|---|
| 35 |
|
|---|
| 36 | int main( int argc, char * argv[] ) {
|
|---|
| 37 | locale loc( getenv("LANG") );
|
|---|
| 38 | cout.imbue( loc );
|
|---|
| 39 |
|
|---|
| 40 | switch ( argc ) {
|
|---|
| 41 | case 5:
|
|---|
| 42 | if ( strcmp( argv[4], "d" ) != 0 ) { // default ?
|
|---|
| 43 | Processors = stoi( argv[4] );
|
|---|
| 44 | if ( Processors < 1 ) goto Usage;
|
|---|
| 45 | } // if
|
|---|
| 46 | case 4:
|
|---|
| 47 | if ( strcmp( argv[3], "d" ) != 0 ) { // default ?
|
|---|
| 48 | xr = stoi( argv[3] );
|
|---|
| 49 | if ( xr < 1 ) goto Usage;
|
|---|
| 50 | } // if
|
|---|
| 51 | case 3:
|
|---|
| 52 | if ( strcmp( argv[2], "d" ) != 0 ) { // default ?
|
|---|
| 53 | xc = stoi( argv[2] );
|
|---|
| 54 | if ( xc < 1 ) goto Usage;
|
|---|
| 55 | } // if
|
|---|
| 56 | case 2:
|
|---|
| 57 | if ( strcmp( argv[1], "d" ) != 0 ) { // default ?
|
|---|
| 58 | yc = stoi( argv[1] );
|
|---|
| 59 | if ( yc < 1 ) goto Usage;
|
|---|
| 60 | } // if
|
|---|
| 61 | case 1: // use defaults
|
|---|
| 62 | break;
|
|---|
| 63 | default:
|
|---|
| 64 | Usage:
|
|---|
| 65 | cerr << "Usage: " << argv[0]
|
|---|
| 66 | << " [ yc (> 0) | 'd' (default " << yc
|
|---|
| 67 | << ") ] [ xc (> 0) | 'd' (default " << xc
|
|---|
| 68 | << ") ] [ xr (> 0) | 'd' (default " << xr
|
|---|
| 69 | << ") ] [ processors (> 0) | 'd' (default " << Processors
|
|---|
| 70 | << ") ]" << endl;
|
|---|
| 71 | exit( EXIT_FAILURE );
|
|---|
| 72 | } // switch
|
|---|
| 73 |
|
|---|
| 74 | unsigned int r, c;
|
|---|
| 75 | int * Z[xr], * X[xr], * Y[xc];
|
|---|
| 76 |
|
|---|
| 77 | for ( r = 0; r < xr; r += 1 ) { // create/initialize X matrix
|
|---|
| 78 | X[r] = new int[xc];
|
|---|
| 79 | for ( c = 0; c < xc; c += 1 ) {
|
|---|
| 80 | X[r][c] = r * c % 37; // for timing
|
|---|
| 81 | } // for
|
|---|
| 82 | } // for
|
|---|
| 83 | for ( r = 0; r < xc; r += 1 ) { // create/initialize Y matrix
|
|---|
| 84 | Y[r] = new int[yc];
|
|---|
| 85 | for ( c = 0; c < yc; c += 1 ) {
|
|---|
| 86 | Y[r][c] = r * c % 37; // for timing
|
|---|
| 87 | } // for
|
|---|
| 88 | } // for
|
|---|
| 89 | for ( r = 0; r < xr; r += 1 ) { // create Z matrix
|
|---|
| 90 | Z[r] = new int[yc];
|
|---|
| 91 | } // for
|
|---|
| 92 |
|
|---|
| 93 | uExecutor * executor = new uExecutor( Processors, Processors, Processors == 1 ? 1 : Processors * 32, true, 0 );
|
|---|
| 94 | uActor::start( executor ); // start actor system
|
|---|
| 95 | uNoCtor<MatrixMult> * multiply = new uNoCtor<MatrixMult>[xr];
|
|---|
| 96 | uNoCtor<WorkMsg> * workMsg = new uNoCtor<WorkMsg>[xr];
|
|---|
| 97 |
|
|---|
| 98 | for ( unsigned int r = 0; r < xr; r += 1 ) {
|
|---|
| 99 | multiply[r].ctor();
|
|---|
| 100 | workMsg[r].ctor( Z[r], X[r], (const int * const *)Y );
|
|---|
| 101 | } // for
|
|---|
| 102 |
|
|---|
| 103 | starttime = steady_clock::now();
|
|---|
| 104 | for ( unsigned int r = 0; r < xr; r += 1 ) {
|
|---|
| 105 | *multiply[r] | *workMsg[r];
|
|---|
| 106 | } // for
|
|---|
| 107 |
|
|---|
| 108 | uActor::stop(); // wait for all actors to terminate
|
|---|
| 109 |
|
|---|
| 110 | cout << (steady_clock::now() - starttime).count() / 1'000'000'000.0 << endl;
|
|---|
| 111 |
|
|---|
| 112 | for ( r = 0; r < xr; r += 1 ) { // deallocate X and Z matrices
|
|---|
| 113 | delete [] X[r];
|
|---|
| 114 | delete [] Z[r];
|
|---|
| 115 | } // for
|
|---|
| 116 | for ( r = 0; r < xc; r += 1 ) { // deallocate Y matrix
|
|---|
| 117 | delete [] Y[r];
|
|---|
| 118 | } // for
|
|---|
| 119 |
|
|---|
| 120 | // malloc_stats();
|
|---|
| 121 | } // main
|
|---|
| 122 |
|
|---|
| 123 | // /usr/bin/time -f "%Uu %Ss %Er %Mkb" a.out
|
|---|
| 124 |
|
|---|
| 125 | // Local Variables: //
|
|---|
| 126 | // compile-command: "u++-work -g -Wall -Wextra -O3 -nodebug -DNDEBUG -multi uC++Matrix.cc" //
|
|---|
| 127 | // End: //
|
|---|
