jxt1234
jxt1234
2. 没明白意思。是指生成 cache 之后，buffer 启动速度慢于image 么？
已知问题，暂时先修改 opencl 的 cmakelists 规避一下 source/backend/opencl/CMakeLists.txt 32 行去掉： # add_definitions(-DMNN_SUPPORT_INTEL_SUBGROUP) 亦即不要加 -DMNN_SUPPORT_INTEL_SUBGROUP
Use testMNNFromOnnx.py to test firstly.
testMNNFromOnnx.py TEST_SUCCESS means MNN is right. You can convert MNN with --keepInputFormat and don't call MNN.expr.convert(input,MNN.expr.NC4HW4).
For your case np.zeros / np.ones will cause input data is all same. So var is very small. It will cause sqrt function Disparity. It's not common case. Please use...
np.zeros() and np.ones() isn't valid input for this model because of instancenorm. It will result to compute sqrt(0.0 + 0.000000000001) , thus cause [Disparity]
MNN-TensorRT support TensorRT 7.0. In future MNN will develop on MNN-CUDA with using cutlass instead of MNN-TensorRT.
有更详细的报错信息么？具体到代码中报错行号?
mnn 版本是什么？
mnn 版本是什么？大概率是 ios 上的 mnn 版本较低