Faster Neighborhood Attention: Reducing the O(n^2) Cost of Self Attention at the Threadblock Level