diff options
Diffstat (limited to 'tests')
-rw-r--r-- | tests/validation/CPP/GEMMLowp.cpp | 6 | ||||
-rw-r--r-- | tests/validation/CPP/GEMMLowp.h | 3 | ||||
-rw-r--r-- | tests/validation/NEON/GEMMLowp.cpp | 17 | ||||
-rw-r--r-- | tests/validation/fixtures/GEMMLowpAssemblyFixture.h | 121 |
4 files changed, 145 insertions, 2 deletions
diff --git a/tests/validation/CPP/GEMMLowp.cpp b/tests/validation/CPP/GEMMLowp.cpp index bac3a20c8e..8670a22a66 100644 --- a/tests/validation/CPP/GEMMLowp.cpp +++ b/tests/validation/CPP/GEMMLowp.cpp @@ -73,6 +73,12 @@ SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<T> &a, co return c; } +// used to validate assembly kernels which don't know anything about offsets +SimpleTensor<int32_t> gemmlowp(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b) +{ + return gemmlowp_matrix_multiply_core(a, b, 0, 0); +} + template <typename T> SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<T> &in, int32_t result_offset, int32_t result_mult_int, int32_t result_shift) { diff --git a/tests/validation/CPP/GEMMLowp.h b/tests/validation/CPP/GEMMLowp.h index c09d8f6176..cbed2206e3 100644 --- a/tests/validation/CPP/GEMMLowp.h +++ b/tests/validation/CPP/GEMMLowp.h @@ -40,6 +40,9 @@ SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<T> &a, co template <typename T> SimpleTensor<uint8_t> gemmlowp_quantize_down_int32_to_uint8_scale(const SimpleTensor<T> &in, int32_t result_offset, int32_t result_mult_int, int32_t result_shift); + +SimpleTensor<int32_t> gemmlowp(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b); + } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/NEON/GEMMLowp.cpp b/tests/validation/NEON/GEMMLowp.cpp index 4407eff060..ba91ced443 100644 --- a/tests/validation/NEON/GEMMLowp.cpp +++ b/tests/validation/NEON/GEMMLowp.cpp @@ -22,7 +22,9 @@ * SOFTWARE. */ #include "arm_compute/core/NEON/kernels/NEGEMMInterleaveBlockedKernel.h" +#include "arm_compute/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.h" #include "arm_compute/core/Types.h" +#include "arm_compute/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.h" #include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h" #include "arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h" #include "arm_compute/runtime/Tensor.h" @@ -38,6 +40,7 @@ #include "tests/framework/datasets/Datasets.h" #include "tests/validation/Validation.h" #include "tests/validation/fixtures/GEMMInterleaveBlockedFixture.h" +#include "tests/validation/fixtures/GEMMLowpAssemblyFixture.h" #include "tests/validation/fixtures/GEMMLowpFixture.h" namespace arm_compute @@ -48,11 +51,21 @@ namespace validation { namespace { -const auto data_int_blk = framework::dataset::make("M", 8, 12) * framework::dataset::make("N", 8, 12) * framework::dataset::make("by", 8, 13) * framework::dataset::make("block", 4, 9); -const auto data_int_blk_tr = framework::dataset::make("M", 8, 17) * framework::dataset::make("N", 8, 14) * framework::dataset::make("by", 12) * framework::dataset::make("block", 4); +const auto data_int_blk = framework::dataset::make("M", 8, 12) * framework::dataset::make("N", 8, 12) * framework::dataset::make("by", 8, 13) * framework::dataset::make("block", 4, 9); +const auto data_int_blk_tr = framework::dataset::make("M", 8, 17) * framework::dataset::make("N", 8, 14) * framework::dataset::make("by", 12) * framework::dataset::make("block", 4); +const auto data_matrix_multiply = framework::dataset::make("M", 12, 20) * framework::dataset::make("N", 12, 20) * framework::dataset::make("K", 16); } // namespace TEST_SUITE(NEON) +TEST_SUITE(ASSEMBLY_MATRIX_MULTIPLY) +using NEGEMMAssemblyFixture = GEMMLowpAssemblyFixture<Tensor, Accessor, NEGEMMLowpAssemblyMatrixMultiplyCore>; +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMAssemblyFixture, framework::DatasetMode::PRECOMMIT, data_matrix_multiply) +{ + // Validate output + validate(Accessor(_target), _reference); +} +TEST_SUITE_END() + TEST_SUITE(GEMMLowp) TEST_SUITE(INTERLEAVE_BLOCKED) diff --git a/tests/validation/fixtures/GEMMLowpAssemblyFixture.h b/tests/validation/fixtures/GEMMLowpAssemblyFixture.h new file mode 100644 index 0000000000..a2587440fb --- /dev/null +++ b/tests/validation/fixtures/GEMMLowpAssemblyFixture.h @@ -0,0 +1,121 @@ +/* + * Copyright (c) 2017 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ARM_COMPUTE_TEST_GEMMLOWP_ASSEMBLY_FIXTURE +#define ARM_COMPUTE_TEST_GEMMLOWP_ASSEMBLY_FIXTURE + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.h" +#include "tests/AssetsLibrary.h" +#include "tests/Globals.h" +#include "tests/IAccessor.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Fixture.h" +#include "tests/validation/CPP/GEMMLowp.h" +#include "tests/validation/Helpers.h" + +#include <random> + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType> +class GEMMLowpAssemblyFixture : public framework::Fixture +{ +public: + template <typename...> + void setup(size_t m, size_t n, size_t k) + { + const TensorShape shape_a(k, m); + const TensorShape shape_b(n, k); + const TensorShape shape_c(n, m); + _target = compute_target(shape_a, shape_b, shape_c); + _reference = compute_reference(shape_a, shape_b, shape_c); + } + +protected: + template <typename U> + void fill(U &&tensor, int i, int lo, int hi) + { + std::uniform_int_distribution<> distribution(lo, hi); + library->fill(tensor, distribution, i); + } + + TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c) + { + // Create tensors + TensorType a = create_tensor<TensorType>(shape_a, DataType::S8, 1); + TensorType b = create_tensor<TensorType>(shape_b, DataType::S8, 1); + TensorType c = create_tensor<TensorType>(shape_c, DataType::S32, 1); + + // Create and configure function + FunctionType gemmlowp; + gemmlowp.configure(&a, &b, &c); + + ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Allocate tensors + a.allocator()->allocate(); + b.allocator()->allocate(); + c.allocator()->allocate(); + + ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Fill tensors + fill(AccessorType(a), 0, -128, 127); + fill(AccessorType(b), 1, -128, 127); + fill(AccessorType(c), 2, 0, 0); + + // Compute GEMM function + gemmlowp.run(); + return c; + } + + SimpleTensor<int32_t> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c) + { + // Create reference + SimpleTensor<int8_t> a{ shape_a, DataType::S8, 1 }; + SimpleTensor<int8_t> b{ shape_b, DataType::S8, 1 }; + + // Fill reference + fill(a, 0, -128, 127); + fill(b, 1, -128, 127); + + return reference::gemmlowp(a, b); + } + + TensorType _target{}; + SimpleTensor<int32_t> _reference{}; +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_GEMMLOWP_FIXTURE */ |