Question d’entretien chez Neural Magic

Speeding up an already cuda kernel, proposing some optimizations.