LinearGo now supports Go 1.6

Last summer, I built LinearGo, a Go wrapper for the famous machine learning library LIBLINEAR. Basically, I used CGO to bridge LIBLINEAR’s C library and Go. At first, I tried to build a 1-to-1 mapping for all of the data structure defined in C liblinear to hide the implementation details of C behind Go code.

However, things become complicated after Go 1.6 as they change their pointer policy for CGO. We can no longer pass something that contains a Go pointer which points to a Go pointer (WTF I just said!?). For example, a Parameter contains a pointer to a FeatureNode (an array of FeatureNode in fact). Before Go 1.6, I simply pass the pointer of Parameter to C, contains a pointer to a FeatureNode (a Go struct). Things like this, which a Go pointer contains a Go pointer, is not allowed anymore in Go 1.6. Therefore, I’m forced to figure out some other way to get around this.

The solution I came up with is passing plain C data type only (things like int, double, or pointers to them), and then re-construct the struct in C. Therefore, I removed the corresponding Go structure and wrote some C helper function which I can call from Go.

For example, I have a call_train helper function, the signature of which is

1
2
3
4
struct model* call_train(
double* x, double* y, int n_rows, int n_cols,
double bias, int solver_type, double C, double p, double eps,
int nr_weight, int* weight_label, double* weight)
;

And the real train function looks like this

1
struct model* train(const struct problem *prob, const struct parameter *param);

In call_train, I manually construct prob and param and then call the real train with these arguments.

Luckily, after some monkey coding, LinearGo is now supporting Go 1.6. Moreover, I believe the current approach is more stable than previous one because we only work on simple data type when bridging while leave the complicated structure in pure C.