diff options
| author | Pablo Marquez Tello <pablo.tello@arm.com> | 2023-09-04 16:08:33 +0100 |
|---|---|---|
| committer | Pablo Marquez Tello <pablo.tello@arm.com> | 2023-09-13 09:16:23 +0000 |
| commit | 145e82e74916a801bd12720564c837c8286042d0 (patch) | |
| tree | a66edf213bde752b572376dfa636191cb1e45880 /src/cpu/kernels/instancenorm/generic/neon/impl.cpp | |
| parent | cf219a4be6e9e9637193b5c9aa4f1eedd0a23900 (diff) | |
| download | ComputeLibrary-145e82e74916a801bd12720564c837c8286042d0.tar.gz | |
Changes to InstanceNrom to enable fp16 in armv8a multi_isa builds
* Code guarded with __ARM_FEATURE_FP16_VECTOR_ARITHMETIC needs
to be moved to an fp16.cpp file to allow compilation with
-march=armv8.2-a+fp16
* Partially resolves MLCE-1102
Change-Id: If53ff1927948b3ad7c9e3c9347bc2af38764e342
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10243
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/cpu/kernels/instancenorm/generic/neon/impl.cpp')
| -rw-r--r-- | src/cpu/kernels/instancenorm/generic/neon/impl.cpp | 29 |
1 files changed, 1 insertions, 28 deletions
diff --git a/src/cpu/kernels/instancenorm/generic/neon/impl.cpp b/src/cpu/kernels/instancenorm/generic/neon/impl.cpp index e35cf97608..483b6f568b 100644 --- a/src/cpu/kernels/instancenorm/generic/neon/impl.cpp +++ b/src/cpu/kernels/instancenorm/generic/neon/impl.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2022 Arm Limited. + * Copyright (c) 2019-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -37,35 +37,12 @@ void vector_float_sum(AccType &result, AccType &result_square, const InputType & result_square = wrapper::vadd(result_square, wrapper::vmul(inputs, inputs)); } -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -template <> -inline void vector_float_sum(float32x4_t &result, float32x4_t &result_square, const float16x8_t &inputs) -{ - vector_float_sum(result, result_square, wrapper::vcvt<float>(wrapper::vgetlow(inputs))); - vector_float_sum(result, result_square, wrapper::vcvt<float>(wrapper::vgethigh(inputs))); -} -template <> -inline float16x8_t vector_float_norm(const float16x8_t &inputs, const float32x4_t &vec_mean, const float32x4_t &vec_multip, const float32x4_t &vec_beta) -{ - const auto input_low = wrapper::vcvt<float>(wrapper::vgetlow(inputs)); - const auto input_high = wrapper::vcvt<float>(wrapper::vgethigh(inputs)); - const auto result_low = wrapper::vcvt<float16_t>(vector_float_norm(input_low, vec_mean, vec_multip, vec_beta)); - const auto result_high = wrapper::vcvt<float16_t>(vector_float_norm(input_high, vec_mean, vec_multip, vec_beta)); - float16x8_t result = wrapper::vcombine(result_low, result_high); - - return result; -} -#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - template <typename InputType, typename AccType> InputType vector_float_norm(const InputType &inputs, const AccType &vec_mean, const AccType &vec_multip, const AccType &vec_beta) { return wrapper::vadd(wrapper::vmul(wrapper::vsub(inputs, vec_mean), vec_multip), vec_beta); } -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - -#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC template <typename T, typename AccType> void instance_normalization_nchw(ITensor *input, ITensor *output, float gamma, float beta, float epsilon, const Window &window) { @@ -164,9 +141,5 @@ void instance_normalization_nchw(ITensor *input, ITensor *output, float gamma, f } template void instance_normalization_nchw<float>(ITensor *input, ITensor *output, float gamma, float beta, float epsilon, const Window &window); -#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) -template void instance_normalization_nchw<float16_t, float>(ITensor *input, ITensor *output, float gamma, float beta, float epsilon, const Window &window); -template void instance_normalization_nchw<float16_t>(ITensor *input, ITensor *output, float gamma, float beta, float epsilon, const Window &window); -#endif //defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) } // namespace cpu } // namespace arm_compute |