diff options
| -rw-r--r-- | src/core/NEON/kernels/NESoftmaxLayerKernel.cpp | 24 | ||||
| -rw-r--r-- | tests/validation/NEON/SoftmaxLayer.cpp | 7 |
2 files changed, 24 insertions, 7 deletions
diff --git a/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp b/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp index 484e58b79b..0f416defab 100644 --- a/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp +++ b/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp @@ -210,10 +210,6 @@ T sqmul(T a, T b); { \ return vsubq_##TAG(a, b); \ } \ - inline vec_16_byte_t<TYPE> vexp(vec_16_byte_t<TYPE> vec) \ - { \ - return vexpq_##TAG(vec); \ - } \ inline vec_16_byte_t<TYPE> vmul_n(vec_16_byte_t<TYPE> vec, TYPE val) \ { \ return vmulq_n_##TAG(vec, val); \ @@ -280,6 +276,26 @@ float32x4x4_t vexp(float32x4x4_t vec) return res; } +float32x4_t vexp(const float32x4_t &vec) +{ + return vexpq_f32(vec); +} + +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +// TODO (COMPMID-1535) : Revisit FP16 approximations +float16x8_t vexp(const float16x8_t &vec) +{ + float16x4x2_t res = + { + { + vcvt_f16_f32(vexpq_f32(vcvt_f32_f16(vget_low_f16(vec)))), + vcvt_f16_f32(vexpq_f32(vcvt_f32_f16(vget_high_f16(vec)))) + } + }; + return vcombine_f16(res.val[0], res.val[1]); +} +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + template <> float32x4x4_t vdup_n<float32x4x4_t>(float val) { diff --git a/tests/validation/NEON/SoftmaxLayer.cpp b/tests/validation/NEON/SoftmaxLayer.cpp index 8c0d46bc41..a5d6344423 100644 --- a/tests/validation/NEON/SoftmaxLayer.cpp +++ b/tests/validation/NEON/SoftmaxLayer.cpp @@ -45,7 +45,8 @@ namespace /** Tolerance for float operations */ constexpr AbsoluteTolerance<float> tolerance_f32(0.000001f); #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -constexpr AbsoluteTolerance<float> tolerance_f16(0.0001f); +constexpr RelativeTolerance<float> rel_tolerance_f16(0.1f); +constexpr AbsoluteTolerance<float> abs_tolerance_f16(0.01f); #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC*/ /** Tolerance for quantized operations */ @@ -122,14 +123,14 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NESoftmaxLayerFixture<half>, framework::Dataset framework::dataset::make("Beta", { 1.0f, 2.0f }))) { // Validate output - validate(Accessor(_target), _reference, tolerance_f16); + validate(Accessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); } FIXTURE_DATA_TEST_CASE(RunLarge, NESoftmaxLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SoftmaxLayerSmallShapes(), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("Beta", { 1.0f, 2.0f }))) { // Validate output - validate(Accessor(_target), _reference, tolerance_f16); + validate(Accessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); } TEST_SUITE_END() #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |