convolution-input-gradient-overfeat-test failed on arm64

[Danliran]

convolution-input-gradient-overfeat-test failed on arm64

TEST(FT16x16, conv2) {
OverFeat_Fast::conv2()
.batchSize(128)
.errorLimit(1.0e-5)
.testInputGradient(nnp_convolution_algorithm_ft16x16);
}

[----------] 4 tests from FT16x16
[ RUN ] FT16x16.conv2
/home/taishan/master/NNPACK/test/testers/convolution.h:316: Failure
Expected: (median(maxErrors)) < (errorLimit()), actual: 1.11966e-05 vs 1e-05
[ FAILED ] FT16x16.conv2 (10959 ms)
[ RUN ] FT16x16.conv3
[ OK ] FT16x16.conv3 (10753 ms)
[ RUN ] FT16x16.conv4

[Danliran]

@Maratyszcza could you please help to investigate this issue?

[Marat Dukhan]

The error is just above the limit, looks like the error tolerance is just too low in the test