w3sip

Same problem -- any chance it can be looked into?

The problem actually is the Numpy exports get optimized out during linking. Adding `-Wl,-force_load,libnumpy.a` to the binary linking with Kivy solved it for me.

I'm seeing extremely poor performance as well, on Xavier NX. The attraction of using nvmpi for us was to offload some of the decoder load. So, the expectation was when...

The problem seems to be not with this wrapper, but the actual nvidia implementation. I'm seeing the same behavior if running with nvidia-supplied ffmpeg on Jetson. Tracking it here -...

@grantthomas - see https://forums.developer.nvidia.com/t/the-nvidia-ffmpeg-package-supports-hardware-accelerated-decode-on-jetson-platforms/123694?page=2

@fengyuentau - thank you! What is the best way to enable the `dnn` module? Simply adding it to t BUILD_LIST didn't seem to do that.

I'm not quite sure how to get objdetect to build anymore. I'm building on Intel Macos, and seeing this output: ``` conanfile.py: Calling build() -- The CXX compiler identification is...

@alalek - thanks, this helped! Protobuf was being disabled.

Well -- I don't have a first hand experience with mgzip (https://pypi.org/project/mgzip/) should do just that -- it's a drop in replacement for gzip, while advertising performance gain. Don't have...

Oh, and I've also attempted to build the Select framework with `bazel build --config=ios_fat -c opt --copt -Os --cpu=ios_arm64 --apple_platform_type=ios --cxxopt=-std=c++17 --host_cxxopt=-std=c++17 //tensorflow/lite/ios:TensorFlowLiteSelectTfOps_framework --define tflite_with_xnnpack_enabled=false` ... didn't make a difference.