Fusing activation functions with preceding computation operators like convolution, matrix multiply, elementwise ops, etc., allows the use of higher precision for the activation computations without ...