#include <actor.hfa>
#include <fstream.hfa>
#include <stdlib.hfa>
#include <string.h>
#include <stdio.h>

unsigned int xr = 100, xc = 100, yc = 100, Processors = 1; // default values

struct derived_actor {
    inline actor;
};
void ?{}( derived_actor & this ) { ((actor &)this){}; }

struct derived_msg {
    inline message;
    int * Z;
	int * X;
    int ** Y;
};

void ?{}( derived_msg & this ) {}
void ?{}( derived_msg & this, int * Z, int * X, int ** Y ) {
    ((message &) this){ Finished };
    this.Z = Z;
    this.X = X;
    this.Y = Y;
}

Allocation receive( derived_actor & receiver, derived_msg & msg ) {
    for ( unsigned int i = 0; i < yc; i += 1 ) { // multiply X_row by Y_col and sum products
        msg.Z[i] = 0;
        for ( unsigned int j = 0; j < xc; j += 1 ) {
            msg.Z[i] += msg.X[j] * msg.Y[j][i];
        } // for
    } // for
    return Finished;
}

int main( int argc, char * argv[] ) {
    switch ( argc ) {
	  case 5:
		if ( strcmp( argv[4], "d" ) != 0 ) {			// default ?
			Processors = atoi( argv[4] );
			if ( Processors < 1 ) goto Usage;
		} // if
	  case 4:
		if ( strcmp( argv[3], "d" ) != 0 ) {			// default ?
			xr = atoi( argv[3] );
			if ( xr < 1 ) goto Usage;
		} // if
	  case 3:
		if ( strcmp( argv[2], "d" ) != 0 ) {			// default ?
			xc = atoi( argv[2] );
			if ( xc < 1 ) goto Usage;
		} // if
	  case 2:
		if ( strcmp( argv[1], "d" ) != 0 ) {			// default ?
			yc = atoi( argv[1] );
			if ( yc < 1 ) goto Usage;
		} // if
	  case 1:											// use defaults
		break;
	  default:
	  Usage:
		sout | "Usage: " | argv[0]
			 | " [ yc (> 0) | 'd' (default " | yc
			 | ") ] [ xc (> 0) | 'd' (default " | xc
			 | ") ] [ xr (> 0) | 'd' (default " | xr
			 | ") ] [ processors (> 0) | 'd' (default " | Processors
			 | ") ]" ;
		exit( EXIT_FAILURE );
	} // switch

    unsigned int r, c;
	int * Z[xr], * X[xr], * Y[xc];

	for ( r = 0; r < xr; r += 1 ) {						// create/initialize X matrix
		X[r] = aalloc( xc );
		for ( c = 0; c < xc; c += 1 ) {
			X[r][c] = r * c % 37;						// for timing
		} // for
	} // for
	for ( r = 0; r < xc; r += 1 ) {						// create/initialize Y matrix
		Y[r] = aalloc( yc );
		for ( c = 0; c < yc; c += 1 ) {
			Y[r][c] = r * c % 37;						// for timing
		} // for
	} // for
	for ( r = 0; r < xr; r += 1 ) {						// create Z matrix
		Z[r] = aalloc( yc );
	} // for

    executor e{ Processors, Processors, Processors == 1 ? 1 : Processors * 16, true };

    printf("starting\n");

    start_actor_system( e );

    printf("started\n");

    derived_msg messages[xr];

    derived_actor actors[xr];

	for ( unsigned int r = 0; r < xr; r += 1 ) {
		messages[r]{ Z[r], X[r], Y };
	} // for

	for ( unsigned int r = 0; r < xr; r += 1 ) {
		actors[r] | messages[r];
	} // for

    printf("stopping\n");

    stop_actor_system();

    printf("stopped\n");

    for ( r = 0; r < xr; r += 1 ) {						// deallocate X and Z matrices
		free( X[r] );
        free( Z[r] );
	} // for
	for ( r = 0; r < xc; r += 1 ) {						// deallocate Y matrix
        free( Y[r] );
	} // for

    return 0;
}