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

ssize_t xr = 500, xc = 500, yc = 500, Processors = 1; // default values, must be signed

struct derived_actor { inline actor; };

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 ) {
	set_allocation( this, Nodelete );
	this.Z = Z;
	this.X = X;
	this.Y = Y;
}

allocation receive( derived_actor & receiver, derived_msg & msg ) {
	for ( i; yc ) {										// multiply X_row by Y_col and sum products
		msg.Z[i] = 0;
		for ( j; xc ) {
			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 = ato( argv[4] );
			if ( Processors < 1 ) fallthru default;
		} // if
	  case 4:
		if ( strcmp( argv[3], "d" ) != 0 ) {			// default ?
			xr = ato( argv[3] );
			if ( xr < 1 ) fallthru default;
		} // if
	  case 3:
		if ( strcmp( argv[2], "d" ) != 0 ) {			// default ?
			xc = ato( argv[2] );
			if ( xc < 1 ) fallthru default;
		} // if
	  case 2:
		if ( strcmp( argv[1], "d" ) != 0 ) {			// default ?
			yc = ato( argv[1] );
			if ( yc < 1 ) fallthru default;
		} // if
	  case 1:											// use defaults
		break;
	  default:
		exit | "Usage: " | argv[0]
			 | " [ yc (> 0) | 'd' (default " | yc
			 | ") ] [ xc (> 0) | 'd' (default " | xc
			 | ") ] [ xr (> 0) | 'd' (default " | xr
			 | ") ] [ processors (> 0) | 'd' (default " | Processors
			 | ") ]" ;
	} // switch

	int * Z[xr], * X[xr], * Y[xc];

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

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

	sout | "starting";
	actor_start( e );
	sout | "started";
	derived_msg messages[xr];
	derived_actor actors[xr];

	for ( r; xr ) {
		messages[r]{ Z[r], X[r], Y };
	} // for
	for ( r; xr ) {
		actors[r] | messages[r];
	} // for

	sout | "stopping";
	actor_stop();
	sout | "stopped";

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