#include <iostream>
#include <fstream>
#include <sstream>
#include <iomanip>
#include "q2matrixmultiply.h"
using namespace std;

int main( int argc, char * argv[] ) {
	intmax_t xrows, xycols, ycols, processors = 1;
	istream *xfile = nullptr, *yfile = nullptr;			// x, y matrix files
	bool IO = false;

	try {
		if ( argc < 4 || argc > 6 ) throw 1;			// minimum arguments
		switch ( argc ) {
		  case 6:
			try {
				yfile = new ifstream( argv[5] );
			} catch( uFile::Failure & ) {				// open failed ?
				cerr << "Error! Cannot open y-matrix input-file \"" << argv[5] << "\"" << endl;
				throw 1;
			} // try
			try {
				xfile = new ifstream( argv[4] );
			} catch( uFile::Failure & ) {				// open failed ?
				cerr << "Error! Cannot open x-matrix input-file \"" << argv[4] << "\"" << endl;
				throw 1;
			} // try
			IO = true;									// print
			goto getdim;
		  case 5:
			processors = convert( argv[4] ); if ( processors <= 0 ) throw 1; // invalid integer ?
			// FALL THROUGH
		  case 4:
		  getdim:
			ycols = convert( argv[3] ); if ( ycols <= 0 ) throw 1; // invalid integer ?
			// FALL THROUGH
		  case 3:
			xycols = convert( argv[2] ); if ( xycols <= 0 ) throw 1; // invalid integer ?
			// FALL THROUGH
		  case 2:
			xrows = convert( argv[1] ); if ( xrows <= 0 ) throw 1; // invalid integer ?
			// FALL THROUGH
		} // switch
	} catch( ... ) {
		cerr << "Usage: " << argv[0] << " xrows (> 0) xycols (> 0) ycols (> 0) [ processors (> 0) | x-matrix-file  y-matrix-file ]" << endl;
		exit( EXIT_FAILURE );							// TERMINATE
	} // try

	int r, c;
	int * Z[xrows], * X[xrows], * Y[xycols];

	for ( r = 0; r < xrows; r += 1 ) {					// create/initialize X matrix
		X[r] = new int[xycols];
		for ( c = 0; c < xycols; c += 1 ) {
			if ( IO ) *xfile >> X[r][c];
			else X[r][c] = 37;							// for timing
		} // for
	} // for
	for ( r = 0; r < xycols; r += 1 ) {					// create/initialize Y matrix
		Y[r] = new int[ycols];
		for ( c = 0; c < ycols; c += 1 ) {
			if ( IO ) *yfile >> Y[r][c];
			else Y[r][c] = 37;							// for timing
		} // for
	} // for
	for ( r = 0; r < xrows; r += 1 ) {					// create Z matrix
		Z[r] = new int[ycols];
	} // for

	uProcessor p[processors - 1];						// set processors used (have one CPU already)
	matrixmultiply( Z, X, xrows, xycols, Y, ycols );	// Z_xr,yc = X_xr,c * Y_c,yc

	if ( IO ) {
		unsigned int width = 8;
		for ( r = 0; r < xycols; r += 1 ) {
			cout << setw(8 * xycols + 3) << "" << "| ";	// print X separator
			for ( c = 0; c < ycols; c += 1 ) {			// print Y matrix
				cout << setw(width) << Y[r][c];
			} // for
			cout << endl;
		} // for

		cout << setfill('-');
		cout << setw(8 * xycols + 3) << "" << '*';		// print separator line
		cout << setw(8 * ycols + 1) << "";
		cout << setfill(' ') << endl;

		for ( r = 0; r < xrows; r += 1 ) {
			for ( c = 0; c < xycols; c += 1 ) {			// print X matrix
				cout << setw(width) << X[r][c];
			} // for
			cout << "   | ";
			for ( c = 0; c < ycols; c += 1 ) {			// print Z matrix
				cout << setw(width) << Z[r][c];
			} // for
			cout << endl;
		} // for
	} // if

	for ( r = 0; r < xrows; r += 1 ) {					// deallocate X matrix
		delete [] X[r];
		delete [] Z[r];
	} // for
	for ( r = 0; r < xycols; r += 1 ) {					// deallocate Y matrix
		delete [] Y[r];
	} // for

	if ( yfile != nullptr ) delete yfile;				// close files if specified
	if ( xfile != nullptr ) delete xfile;
} // main