Inconsistency for contribution calculation between forward and backward pass?

[Developer-KS]

While going through the forward and backward cuda code in the diff-gaussian-rasterization module, I seem to find a slight inconsistency between the two passes (or maybe I have been mistaken). See the following two screenshot:

The first one is the forward pass and the second one is the backward pass. It seems the last contributor gaussian's (1- alpha) is not recorded in the final_T value in the forward pass while still being accounted for in the backward pass. Is this a mistake, or am I missing something here?

Also, why do we need this "test_T" in the forward pass anyway? Isn't it more clear to use T's value for the testing directly?

[Developer-KS]

By the way, this will have a larger impact when someone changes the T threshold to a larger value or uses an alpha clipping threshold that's closer to 1.0f, i.e. when the last contributor's contribution is significant.

[flm8620]

I think there is no bug in the code. When the forward pass skips, it actually skips the 1 over the last contributor, and the assignment last_contributor = contributor is not executed when the loop is continued.

The code is indeed confusing. But the effect is correct. if the check test_T < 0.0001 is true, and we know alpha <= 0.99, then we have T * (1-0.99) < 0.0001, so T < 0.01, which is negligible.