In CNN is tipically put on in cascade differents types of convolution layers, for example a 2D Convolution along with 2D Average Pooling. The convolution has the associative property:
$$(A*B)*C=A*(B*C)$$
In my case, $B$ and $C$ are the kernels of 2D Convolution and 2D Average pooling. Both are known and $B*C$ can be precomputed.
It's possible perform the convolution of kernels $K= B * C$ and later the convolution of $A*K$ instead of $(A*B)*C$?