Cost function, and a small example?

[profmadden]

Hi -- can you give a little more detail on how you're calculating wire length, and what the overall metric will be? In the example, metal 2:

WL_M2 = (25 + 100 + 50) + (50 + 50) = 275

Where does the 25 come from (and why?). If there's a via, are we assuming it's in the middle of the GCell? What would the capacity for the GCell boundaries be in terms of overflow?

A small (routed) example, with the routing cost given, would be great; it would help sanity check the code.

[liangrj2014]

Hi! Thanks a lot for the question! We will update the contest introduction to give more details soon! Also, we will release the evaluation scripts, including sanity check by this week. Thanks for your suggestion! We will provide a routed example as well. Sorry for the late response!

[liangrj2014]

Hi profmadden! FYI. We've released the evlaution scripts and example global routing solutions.

[profmadden]

Hi -- the updated formats are great, and I very much appreciate the evaluator! Thanks!

To check that I'm doing this right -- the evaluator seems to take a long time, a few hours or more on some of the designs? And also, the sample routings have overflows (here's ariane133_68)

num of uncovered nets 0 num of unchecked nets: 0 Finished! numVia: 825323 totalWL: 7.18906e+09 Layer 0 worst oveflow 0 Layer 1 worst oveflow -3.14286 Layer 2 worst oveflow -5.68421 Layer 3 worst oveflow -1.71429 Layer 4 worst oveflow -5.63158 Layer 5 worst oveflow -1.71429 Layer 6 worst oveflow -5.78947 Layer 7 worst oveflow -3.14286 Layer 8 worst oveflow -7.21053 Layer 9 worst oveflow -22.1429 9.45929e+06 3.30129e+06 5.22501e+06 0.525937 0.183552 0.290511

All layers have global routing cells that overflow, with layer 9 having the worst overflow.

[liangrj2014]

Hi profmadden，

Thanks! I am you are doing right. Yes, the current version of evaluator is slow and we will see how to accelerate it. And the sample routings do have overflows, since we generate the sample routing by only conducting pattern routing without further addressing overflows. By releasing the evaluator and sample routing, we'd like to show how the performance metrics are calculated and what is the expected output format.

[profmadden]

Thanks again for the evaluator and reference routings -- I'm digging through the evaluator, and was wondering about commitVia. It looks like when there's a via, there's a fraction of capacity consumed in the global routing cell (which makes perfect sense), but also in the global routing cell for [layer][x-1][y] on the horizontal layer. Should the cost function also add demand to the next global routing cell at [layer][x+1][y]?

The evaluator currently does this -

    for (int layer = l; layer <= l + 1; ++layer) {
        int direction = layerDirections[layer];
        if (direction == 0) {
            if (x + 1 < X) {
                Demand[layer][x][y] = Demand[layer][x][y] + (layerMinLengths[layer] * demand) / hEdgeLengths[x];
            }
            if ( x > 0) {
                Demand[layer][x-1][y] = Demand[layer][x-1][y] + (layerMinLengths[layer] *  demand) / hEdgeLengths[x-1];
            }

        } else {
            if (y + 1 < Y) {
                Demand[layer][x][y] = Demand[layer][x][y] + (layerMinLengths[layer] *  demand) / vEdgeLengths[y];
            }
            if ( y > 0) {
                Demand[layer][x][y-1] = Demand[layer][x][y-1] + (layerMinLengths[layer] *  demand) / vEdgeLengths[y-1];
            }
        }
    }

Perhaps modify to something like this, to capture the routing impact in both the next and prior cells?

    for (int layer = l; layer <= l + 1; ++layer) {
        int direction = layerDirections[layer];
        float delta;
        if (direction == 0)
        {
          delta = (layerMinLengths[layer] * demand) / hEdgeLengths[x];
        }
        else
        {
           delta = (layerMinLengths[layer] * demand) / vEdgeLengths[x];
        } 

       Demand[layer][x][y] += delta;
        if (direction == 0) {
            if (x + 1 < X) {
                Demand[layer][x+1][y] += delta;
            }
            if ( x > 0) {
                Demand[layer][x-1][y] = delta;
            }

        } else {
            if (y + 1 < Y) {
                Demand[layer][x][y+1] = delta;
            }
            if ( y > 0) {
                Demand[layer][x][y-1] = delta;
            }
        }
    }

[liangrj2014]

Hi profmadden,

Thanks a lot for the comments!

The demand map, represented by the "Demand", records the demand on GCell edges in our context. Each entry, such as Demand[layer][x][y], indicates the demand for the GCell edge from GCell[layer][x][y] to GCell[layer][x+1][y]. Similarly, Demand[layer][x-1][y] signifies the demand for the GCell edge from GCell[layer][x-1][y] to GCell[layer][x][y]. Consequently, our implementation already capture the routing impact in both the next and prior cells. Hope it makes sense to you now. Please kindly let us know if you have further questions!