cc @joy2myself and @asmorkalov

Remaining failed test on RISC-V:

[  FAILED  ] 13 tests, listed below:
[  FAILED  ] Reproducibility_FCN.Accuracy
[  FAILED  ] Layer_GRU_Test_Accuracy_with_.Pytorch
[  FAILED  ] Test_Int8_nets.FasterRCNN_inceptionv2/0, where GetParam() = OCV/CPU
[  FAILED  ] Test_Int8_nets.FasterRCNN_vgg16/0, where GetParam() = OCV/CPU
[  FAILED  ] Test_Int8_nets.YOLOv3/0, where GetParam() = OCV/CPU
[  FAILED  ] Test_ONNX_layers.LSTM_Activations/0, where GetParam() = OCV/CPU
[  FAILED  ] Test_ONNX_layers.GRU/0, where GetParam() = OCV/CPU
[  FAILED  ] Test_ONNX_layers.GRU_bidirectional/0, where GetParam() = OCV/CPU
[  FAILED  ] Test_ONNX_layers.CumSum/0, where GetParam() = OCV/CPU
[  FAILED  ] Test_TensorFlow_layers.reduce_max/0, where GetParam() = OCV/CPU
[  FAILED  ] Test_TensorFlow_layers.reduce_max_channel/0, where GetParam() = OCV/CPU
[  FAILED  ] Test_TensorFlow_layers.reduce_max_channel_keep_dims/0, where GetParam() = OCV/CPU
[  FAILED  ] Test_TensorFlow_layers.pooling_reduce_max/0, where GetParam() = OCV/CPU

@asmorkalov Thanks for your information! I can reproduce 4/13 failed test cases as below.

[  FAILED  ] Reproducibility_FCN.Accuracy
[  FAILED  ] Test_Int8_nets.FasterRCNN_inceptionv2/0, where GetParam() = OCV/CPU
[  FAILED  ] Test_Int8_nets.FasterRCNN_vgg16/0, where GetParam() = OCV/CPU
[  FAILED  ] Test_Int8_nets.YOLOv3/0, where GetParam() = OCV/CPU

And then, I test the 4 failed cases on master, both of them failed too. So I think it may not be caused by this PR.

Others 9 test cases are pass on my server but I notice that all of those test cases are added to opencv_extra in the last several weeks. So if I guess right, update the opencv_extra repository to the current master can resolve the problem.

Tests Test_Int8_nets_XXX are the new thing introduced in OpenCV week ago or so. Please ignore it for now. It looks like I have outdated extra or some inconsistency with my Docker containers.

Could you add small test for the fixed universal intrinsics to https://github.com/opencv/opencv/blob/master/modules/core/test/test_intrin128.simd.hpp or https://github.com/opencv/opencv/blob/master/modules/core/test/test_intrin.cpp.

Oh, I found that there is already a function to check the result of reduce_sum in "test_intrin_utils.hpp". And the reason of why we miss it before is we usually run the test with VLEN=128 on QEMU.

We can just run qemu-riscv64 -cpu rv64,x-v=true,vlen=256 ./bin/opencv_test_core --gtest_filter="hal_intrin128.float32x4_BASELINE" to reproduces the issue.

The google test report the Failure on line 1467 as below

/root/opencv/modules/core/test/test_intrin_utils.hpp:231: Failure
Expected equality of these values:
  a
    Which is: 10
  b
    Which is: 14.4375
Google Test trace:
/root/opencv/modules/core/test/test_intrin_utils.hpp:1467: i=0

Do you think we need new test cases? Or we just need a new way (with VLEN>128, VLEN=256 as an example) to run the existing tests?

Unfortunately, I also found that this PR does not fix all the failures in hal_intrin128.float32x4_BASELINE. It does fix the v_reduce_sum4 but looks like there are other errors on v_rotate_right as the report by google test.

No new test case is not required then. I'll add vlen=256 case to CI.

Hi @asmorkalov @hanliutong. Can you please specify what is the exact error in the failed Test_Int8_Nets.XXX? Is it related to stricter thresholds for scoreDiff and IoUDiff?