What the Doubling Really Costs

A convolutional backbone follows the usual convention: every time pooling halves the grid, the filter count doubles. Two of its convolutional layers, both $3\times3$ with "same" padding, sit in consecutive stages:

layer	input block	filters	output block
P	$32\times32\times64$	$64$	$32\times32\times64$
Q	$16\times16\times128$	$128$	$16\times16\times128$

For each layer, count two different things for a single image:

its weights (ignore the biases),
the multiplications one forward pass through it performs — every output value costs one multiplication per weight in the filter that produced it.

Which statement about $Q$ compared with $P$ is correct?

What the Doubling Really Costs

Hints