NCNN has different results based on CPU vs vulkan (vulkan is incorrect)

[koush]

The inferences sometimes works correctly but more often than not the vulkan inference seems has corrupt/invalid results. Inference always works correctly on CPU.

1. Run model https://github.com/koush/ncnn-models/tree/main/scrypted_yolov9t_320
2. Compare results from CPU to Vulkan

I also tried using the various tips listed here:: https://github.com/Tencent/ncnn/wiki/FAQ-ncnn-produce-wrong-result#disable-fp16

[apanand14]

Hello all,
@nihui, I'm having the same issue. I have yolov8-seg based model which I run with NCNN CPU successfully. Great performace. I run the same model on vulkan available device and I observed that performance degraded. With vulkan compute, many objects were not detected due to low confidence which was detected with CPU. It could be related to vulkan, not optimzed vulkan drivers but I would appreciate if there's some solution in this direction available. Thank you in advance.

I'm running it on android