Efficient GPU kernels for block-sparse matrix multiplication and convolution - View it on GitHub