diff options
| author | Vidhya Sudhan Loganathan <vidhyasudhan.loganathan@arm.com> | 2018-11-16 11:33:12 +0000 |
|---|---|---|
| committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2018-11-16 17:37:40 +0000 |
| commit | a25d16c86f0d870408bc8b941aa755093417b0f0 (patch) | |
| tree | b62d145a4e5009d894262a7ffa66cdba8260bb03 /tests/validation/CL/Winograd.cpp | |
| parent | a7b54f44e2bf133179f24a34007bc93237dd2265 (diff) | |
| download | ComputeLibrary-a25d16c86f0d870408bc8b941aa755093417b0f0.tar.gz | |
COMPMID-1266 : Add support for FP16 in CLWinogradConvolutionLayer: 5x5 kernels
Introduced F32 accumulation for F16 winograd gemm and output transform
WinogradConvolution will be available for F16 only if fast math flag is enabled
Change-Id: I215593c205236a0f9669218437bb40b184ec6a4f
Diffstat (limited to 'tests/validation/CL/Winograd.cpp')
| -rw-r--r-- | tests/validation/CL/Winograd.cpp | 21 |
1 files changed, 12 insertions, 9 deletions
diff --git a/tests/validation/CL/Winograd.cpp b/tests/validation/CL/Winograd.cpp index 930f7aa8ce..f7f06b7f79 100644 --- a/tests/validation/CL/Winograd.cpp +++ b/tests/validation/CL/Winograd.cpp @@ -58,6 +58,9 @@ constexpr AbsoluteTolerance<float> tolerance_f32(0.001f); const AbsoluteTolerance<half> tolerance_f16(half(0.5f)); constexpr AbsoluteTolerance<float> tolerance_convolution_layer_f32(0.1f); const AbsoluteTolerance<half> tolerance_convolution_layer_f16(half(0.4f)); +RelativeTolerance<half_float::half> rel_tolerance_f16(half(0.2)); /**< Tolerance value for comparing reference's output against implementation's output for FP16 data types */ +constexpr float tolerance_num = 0.05f; /**< Tolerance number */ +constexpr float abs_tolerance_convolution_layer_f16 = 2.5f; /**< Tolerance number */ // Input transform const auto SmallWinogradInputTransformDatasetNCHW = @@ -834,10 +837,10 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture, fram TEST_SUITE_END() // Conv1x5 TEST_SUITE_END() // FP32 -#ifdef WINOGRAD_F16_SUPPORT //to be reintroduced after COMPMID-1266 is resolved + TEST_SUITE(FP16) -using CLWinogradConvolutionLayerFastMathFixture16 = WinogradConvolutionLayerFastMathValidationFixture<CLTensor, CLAccessor, CLWinogradConvolutionLayer, half>; +using CLWinogradConvolutionLayerFastMathFixture16 = WinogradConvolutionLayerFastMathValidationFixture<CLTensor, CLAccessor, CLWinogradConvolutionLayer, half, float>; TEST_SUITE(Conv3x3) FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallWinogradConvolutionLayer3x3Dataset(), @@ -856,7 +859,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture16, fr framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output - validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f16); + validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_convolution_layer_f16); } TEST_SUITE_END() // Conv3x3 @@ -878,7 +881,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture16, fr framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output - validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f16); + validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_convolution_layer_f16); } TEST_SUITE_END() // Conv3x1 @@ -900,7 +903,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture16, fr framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output - validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f16); + validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_convolution_layer_f16); } TEST_SUITE_END() // Conv1x3 @@ -924,7 +927,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture16, fr { // Validate output - validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f16); + validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_convolution_layer_f16); } TEST_SUITE_END() // Conv5x5 @@ -948,7 +951,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture16, fr { // Validate output - validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f16); + validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_convolution_layer_f16); } TEST_SUITE_END() // Conv5x1 @@ -972,12 +975,12 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture16, fr { // Validate output - validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f16); + validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_convolution_layer_f16); } TEST_SUITE_END() // Conv1x5 TEST_SUITE_END() // FP16 -#endif /*#ifdef WINOGRAD_F16_SUPPORT*/ + TEST_SUITE_END() // ConvolutionLayer TEST_SUITE_END() // Winograd TEST_SUITE_END() // CL |