Phil Ashby

OK, the dependency chain we have currently: `backscrub -> tensorflow/v2.8.0 -> flatbuffers/v1.12.0 (see /tools/cmake/modules/flatbuffers.cmake)`. Shifting our dependency to `tensorflow/v2.9.0` might fix, as that depends on `flatbuffers/v2.0.5`: @oshikore can you please...

As I suspected, the earlier tag isn't useful, so we'll need to force a specific checkout beyond the last tagged release. I have a PR (#147) that contains the CMake...

Hmmn. Tensorflow v2.8.0 as we are currently using does not compile against flatbuffers v2.0.6, so we _must_ upgrade Tensorflow before we can upgrade flatbuffers... unfortunately the latest Tensorflow (v2.9.1) only...

Do we want to consider parallel processing the multiple passes? IMO this is probably not worth the extra pain unless we run very low on time budget..

I'm aware that those were written because OpenCV doesn't do them.. but happy to be proven wrong if we up the minimum OCV version :smile: Found this Q&A: https://answers.opencv.org/question/211941/how-to-use-simd-feature-of-opencv-as-third-party-library/ I'm...

OK! Thanks for the pointer towards `cv::blendLinear`, I've just tried it and the performance gain seems marginal (~0.8 of previous run time). Digging in, I find that ~1/3 of the...

Yep - when we get down to shaving a few % that's worth it - looking for 10x improvement through SIMD first :smile:

So: I tried MMX, loading a pixel at a time (3 channels) into an `_m64` executing the multiplies, adds & a shift to divide down, then extracted the pixel back...

Sounds like you know waay more than I ever will @BenBE :smile: I've never looked into SIMD stuff before now, so my feeble attempt was through cut/paste/fixup from stack overflow...

Hmmn, that seems to be the `read_thread` trying to join itself because it held the last reference to the `shared_ptr`.. might want to review the use of `shared_ptr` (again!)