Dmitry Kurtaev

I believe that there is a way to make Runner.Listener protect self-hoster runners using Secrets. From the description: > Secrets are not passed to workflows that are triggered by a...

@Octopus136, @dbrenn74, I think we can try reproduce problem with any model unless memory allocation problem.

Experiment is not a valid for `float64` data type. Two architectures may load them differently. Even if `input_img_float.bin` is the same on disk - in the memory they might be...

Ok, got it. Bit-exactness is related to integer data types only, not float32/float64. So in case of float64 regular resize is used: https://github.com/opencv/opencv/blob/32db3f943d73f30d8f23b7dc57bb681b5de6e764/modules/imgproc/src/resize.cpp#L4167-L4168

First, I was confused why there is a need to have `imread` with pre-allocated memory while `VideoCapture` can do something like that (for example, read a sequence of the same...

@thewoz, you may refer to current imread by `@ref imread(const String&, int)` but add a note that `img` param can be pre-allocated and memory is reused if sizes match.

@thewoz, https://github.com/opencv/opencv/pull/24415/files#r1386723679

As discussed on team meeting, let's check Mat is contiguous.

Docker and WSL are not perfect candidates for performance benchmarking. Both do not have full access to the system resources. So this is not OpenCV issue.

@ukoehler, please add a link to model. If possible, provide per-layer timings using [getPerfProfile](https://docs.opencv.org/4.x/db/d30/classcv_1_1dnn_1_1Net.html#a06ce946f675f75d1c020c5ddbc78aedc) on both 4.5.2 and 4.8.0. This might help determine which layer has regression. ```cpp std::vector timings;...