Get offset-th output tensor. @param offset @return the tensor or null if failed to get it.

Add one kernel to the bound kernels.

@return shape of the tensor as long array.

Deallocate the memory.

Reset all state so we can re-use the buffer.

modifiers

eat space like string.

fill() is most frequently called with small byte counts (<= 4), which is why we're using loops rather than calling memset.

the exposed interface to use

Run some initialization before `Run`, it will invoke after `SetParam` and `SetContext`, that is both the param_ and context_ are valid.

Set data.