void->void pipeline MatrixMult {
    add FloatSource(10);
    add MatrixMultiply(10, 10, 10, 10);
    add FloatPrinter();
}

float->float pipeline MatrixMultiply(int x0, int y0, int x1, int y1) {
    // rearrange and duplicate the matrices as necessary:
    add RearrangeDuplicateBoth(x0, y0, x1, y1);
    add MultiplyAccumulateParallel(x0, x0);
}

float->float splitjoin RearrangeDuplicateBoth(int x0, int y0, 
    int x1, int y1) {
    split roundrobin(x0 * y0, x1 * y1);
    // the first matrix just needs to get duplicated 
    add DuplicateRows(x1, x0);
        
    // the second matrix needs to be transposed first
    // and then duplicated:
    add RearrangeDuplicate(x0, y0, x1, y1);
    join roundrobin;
}

float->float pipeline RearrangeDuplicate(int x0, int y0, int x1, int y1) {
    add Transpose(x1, y1);
    add DuplicateRows(y0, x1*y1);
}

float->float splitjoin Transpose(int x, int y) {
    split roundrobin;
    for (int i = 0; i < x; i++) add Identity<float>();
    join roundrobin(y);
}

float->float splitjoin MultiplyAccumulateParallel(int x, int n) {
    split roundrobin(x*2);
    for (int i = 0; i < n; i++) add MultiplyAccumulate(x);
    join roundrobin(1);
}

float->float filter MultiplyAccumulate(int rowLength) {
    work pop rowLength*2 push 1 {
        float result = 0;
        for (int x = 0; x < rowLength; x++) {
            result += (peek(0) * peek(1));
	    pop();
	    pop();
        }
	push(result);
    }
}

float->float pipeline DuplicateRows(int x, int y) {
    add DuplicateRowsInternal(x, y);
}

float->float splitjoin DuplicateRowsInternal(int times, int length) {
   split duplicate;
   for (int i = 0; i < times; i++) add Identity<float>();
   join roundrobin(length);
}

void->float filter FloatSource(float maxNum) {    
    float num;
    init {
	num = 0;
    }
    work push 1 {
       push(num);
       num++;
       if (num == maxNum) num = 0;
    }
}

float->void filter FloatPrinter {
    work pop 1 {
	println(pop());
    }
}

float->void filter Sink {
    int x;
    init {
	x = 0;
    }
    work pop 1 {
	pop();
	x++;
	if (x == 100) {
	    println("done..");
	    x = 0;
	}
    }
}