diff options
Diffstat (limited to 'tests/validation/NEON/ConvolutionLayer.cpp')
| -rw-r--r-- | tests/validation/NEON/ConvolutionLayer.cpp | 10 |
1 files changed, 6 insertions, 4 deletions
diff --git a/tests/validation/NEON/ConvolutionLayer.cpp b/tests/validation/NEON/ConvolutionLayer.cpp index 72e67a9acc..d216d9db86 100644 --- a/tests/validation/NEON/ConvolutionLayer.cpp +++ b/tests/validation/NEON/ConvolutionLayer.cpp @@ -47,9 +47,10 @@ namespace validation { namespace { -RelativeTolerance<float> rel_tolerance_f32(0.01f); /**< Relative tolerance for FP32 types */ -const AbsoluteTolerance<float> abs_tolerance_f32(0.002f); /**< Absolute tolerance for FP32 types */ -const AbsoluteTolerance<float> abs_tolerance_1xN_f32(0.0041f); /**< Absolute tolerance for FP32 types */ +const RelativeTolerance<float> rel_tolerance_f32(0.01f); /**< Relative tolerance for FP32 types */ +const RelativeTolerance<float> rel_tolerance_winograd_3x3_f32(0.05f); /**< Relative tolerance for FP32 types */ +const AbsoluteTolerance<float> abs_tolerance_f32(0.002f); /**< Absolute tolerance for FP32 types */ +const AbsoluteTolerance<float> abs_tolerance_1xN_f32(0.0041f); /**< Absolute tolerance for FP32 types */ #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC const RelativeTolerance<half_float::half> rel_tolerance_f16(half_float::half(0.2f)); /**< Relative tolerance value for FP16 types */ @@ -274,7 +275,8 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEWinogradConvolutionLayerFixture<float>, frame { // Validate output - validate(Accessor(_target), _reference, abs_tolerance_f32); + // floating point arithmetic the Winograd results will not be exactly the same as direct convolution, especially for big shapes + validate(Accessor(_target), _reference, rel_tolerance_winograd_3x3_f32, 0.f, float(abs_tolerance_f32)); } TEST_SUITE_END() // Conv3x3 |