cv::dnn::net::forward fails because of the latest Winograd optimization in OpenCV

[lioneldaniel]

cv::dnn::net::forward fails because of the latest Winograd optimization in OpenCV

[UNeedCryDear]

I got the same error in forward.
I'm not sure whether it is caused by this pr, because I run it in Python(opencv-python 4.7.0.68) without error.

import cv2
import numpy as np
print(cv2.__version__)
net =cv2.dnn.readNetFromONNX("./yolov5s-seg.onnx")
size=(640,640,3)
img=np.zeros(size,dtype=np.uint8)
blob=cv2.dnn.blobFromImage(img,1/255.0,size=size[0:2],mean=(0,0,0),swapRB=True)
net.setInput(blob)
out=net.getUnconnectedOutLayersNames()
print(out)
pred1,pred2=net.forward(out)
print(pred1.shape,pred2.shape)

At present, it can be determined that there will be problems under the cpp, so I plan to cmake the source code and debug it.