diff options
author | Pablo Tello <pablo.tello@arm.com> | 2017-07-06 16:43:14 +0100 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-09-17 14:16:42 +0100 |
commit | 0d176141ca759f0f45b47ed32547f1e44fd875fb (patch) | |
tree | a6c92a40567347e731d9d5efbe2a644bf4afb94b /tests | |
parent | b7c2a99f847d3baef1710be5cf34f978514101dd (diff) | |
download | ComputeLibrary-0d176141ca759f0f45b47ed32547f1e44fd875fb.tar.gz |
COMPMID-421: Added FP16 support to the NEON Direct Convolution function.
Change-Id: I3a1aa2ce985ecf95fc5f441a6e6d43b4935306ee
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/79965
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Diffstat (limited to 'tests')
-rw-r--r-- | tests/benchmark_new/NEON/DirectConvolutionLayer.cpp | 13 | ||||
-rw-r--r-- | tests/validation/NEON/ConvolutionLayerDirect.cpp | 56 |
2 files changed, 62 insertions, 7 deletions
diff --git a/tests/benchmark_new/NEON/DirectConvolutionLayer.cpp b/tests/benchmark_new/NEON/DirectConvolutionLayer.cpp index 5588321cc6..dcefbc7512 100644 --- a/tests/benchmark_new/NEON/DirectConvolutionLayer.cpp +++ b/tests/benchmark_new/NEON/DirectConvolutionLayer.cpp @@ -37,14 +37,21 @@ namespace arm_compute { namespace test { +namespace +{ +#ifdef ARM_COMPUTE_ENABLE_FP16 +const auto data_types = framework::dataset::make("DataType", { DataType::QS8, DataType::F16, DataType::F32 }); +#else /* ARM_COMPUTE_ENABLE_FP16 */ +const auto data_types = framework::dataset::make("DataType", { DataType::QS8, DataType::F32 }); +#endif /* ARM_COMPUTE_ENABLE_FP16 */ +} // namespace + using NEDirectConvolutionLayerFixture = ConvolutionLayerFixture<Tensor, NEDirectConvolutionLayer, Accessor>; TEST_SUITE(NEON) REGISTER_FIXTURE_DATA_TEST_CASE(DirectConvolutionLayer, NEDirectConvolutionLayerFixture, framework::DatasetMode::ALL, - framework::dataset::combine(framework::dataset::combine(datasets::DirectConvolutionLayerDataset(), - framework::dataset::make("DataType", { DataType::F32, DataType::QS8 })), - framework::dataset::make("Batches", { 1, 4, 8 }))); + framework::dataset::combine(framework::dataset::combine(datasets::DirectConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", { 1, 4, 8 }))); TEST_SUITE_END() } // namespace test diff --git a/tests/validation/NEON/ConvolutionLayerDirect.cpp b/tests/validation/NEON/ConvolutionLayerDirect.cpp index 4949f387f7..034a8b2045 100644 --- a/tests/validation/NEON/ConvolutionLayerDirect.cpp +++ b/tests/validation/NEON/ConvolutionLayerDirect.cpp @@ -48,8 +48,11 @@ using namespace arm_compute::test::validation; namespace { -const float tolerance_fp = 1e-3f; /**< Tolerance for floating point tests */ -const float tolerance_qs8 = 1; /**< Tolerance for fixed point tests */ +const float tolerance_fp32 = 1e-3f; /**< Tolerance for floating point tests */ +#ifdef ARM_COMPUTE_ENABLE_FP16 +const float tolerance_fp16 = 0.01f; /**< Tolerance for half precision floating point tests */ +#endif /* ARM_COMPUTE_ENABLE_FP16 */ +const float tolerance_qs8 = 1; /**< Tolerance for fixed point tests */ /** Compute NEON direct convolution layer function. * @@ -88,7 +91,7 @@ Tensor compute_convolution_layer(const TensorShape &src_shape, const TensorShape BOOST_TEST(!dst.info()->is_resizable()); // Fill tensors - if(dt == DataType::F32) + if(dt == DataType::F16 || dt == DataType::F32) { std::uniform_real_distribution<> distribution(-1.f, 1.f); library->fill(Accessor(src), distribution, 0); @@ -129,6 +132,51 @@ BOOST_AUTO_TEST_SUITE(NEON) BOOST_AUTO_TEST_SUITE(ConvolutionLayer) BOOST_AUTO_TEST_SUITE(Direct) +#ifdef ARM_COMPUTE_ENABLE_FP16 +BOOST_AUTO_TEST_SUITE(Float16) +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit")) +BOOST_DATA_TEST_CASE(W1x1, + DirectConvolutionShapes() * boost::unit_test::data::make(DataType::F16) * boost::unit_test::data::xrange(1, 3, 1) * boost::unit_test::data::xrange(1, 3, 1) * boost::unit_test::data::make({ 1, 4, 8, 16 }), + input_shape, dt, sx, sy, num_kernels) +{ + const unsigned int kernel_size = 1; + const PadStrideInfo conv_info(sx, sy, 0, 0, DimensionRoundingType::FLOOR); + const TensorShape w_shape(kernel_size, kernel_size, input_shape.z(), static_cast<unsigned int>(num_kernels)); + const TensorShape b_shape(static_cast<unsigned int>(num_kernels)); + const TensorShape d_shape(get_output_shape(input_shape, w_shape, conv_info)); + + Tensor dst = compute_convolution_layer(input_shape, w_shape, b_shape, d_shape, dt, conv_info); + + RawTensor ref = Reference::compute_reference_convolution_layer(input_shape, w_shape, b_shape, d_shape, dt, conv_info, 0); + + // Validate output + validate(NEAccessor(dst), ref); +} + +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit")) +BOOST_DATA_TEST_CASE(W3x3, DirectConvolutionShapes() * boost::unit_test::data::make(DataType::F16) * boost::unit_test::data::xrange(1, 3, 1) * boost::unit_test::data::xrange(1, 3, + 1) + * boost::unit_test::data::xrange(0, 2, + 1) + * boost::unit_test::data::xrange(0, 2, 1) * boost::unit_test::data::make({ 1, 4, 8, 16 }), + input_shape, dt, sx, sy, px, py, num_kernels) +{ + const unsigned int kernel_size = 3; + const PadStrideInfo conv_info(sx, sy, px, py, DimensionRoundingType::FLOOR); + const TensorShape w_shape(kernel_size, kernel_size, input_shape.z(), static_cast<unsigned int>(num_kernels)); + const TensorShape b_shape(static_cast<unsigned int>(num_kernels)); + const TensorShape d_shape(get_output_shape(input_shape, w_shape, conv_info)); + + Tensor dst = compute_convolution_layer(input_shape, w_shape, b_shape, d_shape, dt, conv_info); + + RawTensor ref = Reference::compute_reference_convolution_layer(input_shape, w_shape, b_shape, d_shape, dt, conv_info, 0); + + // Validate output + validate(NEAccessor(dst), ref, tolerance_fp16); +} +BOOST_AUTO_TEST_SUITE_END() +#endif /* ARM_COMPUTE_ENABLE_FP16 */ + BOOST_AUTO_TEST_SUITE(Float) BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit")) BOOST_DATA_TEST_CASE(W1x1, @@ -166,7 +214,7 @@ BOOST_DATA_TEST_CASE(W3x3, DirectConvolutionShapes() * CNNFloatDataTypes() * boo RawTensor ref = Reference::compute_reference_convolution_layer(input_shape, w_shape, b_shape, d_shape, dt, conv_info, 0); // Validate output - validate(Accessor(dst), ref, tolerance_fp); + validate(Accessor(dst), ref, tolerance_fp32); } BOOST_AUTO_TEST_SUITE_END() |