diff options
Diffstat (limited to 'src/core/NEON/kernels/batchnormalization/impl')
5 files changed, 676 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp new file mode 100644 index 0000000000..e3d9b670b3 --- /dev/null +++ b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp @@ -0,0 +1,179 @@ +/* + * Copyright (c) 2020-2021 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. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensorPack.h" +#include "arm_compute/core/Window.h" + +#include "src/core/NEON/kernels/detail/NEActivationFunctionDetail.h" +#include "src/core/NEON/NEMath.h" +#include "src/core/NEON/wrapper/wrapper.h" + +#include <arm_neon.h> +#include <cmath> +#include <cstddef> + +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) +namespace arm_compute +{ +namespace +{ +using BatchNomalizationPtr = void (*)(ITensor *src, + ITensor *dst, + const ITensor *mean, + const ITensor *var, + const ITensor *beta, + const ITensor *gamma, + float epsilon, + ActivationLayerInfo &act_info, + const Window &window); + +template <typename T> +void batch_normalization(ITensor *src, + ITensor *dst, + const ITensor *mean, + const ITensor *var, + const ITensor *beta, + const ITensor *gamma, + float epsilon, + ActivationLayerInfo &act_info, + const Window &window) +{ + /** SIMD vector tag type. */ + using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<float16_t, wrapper::traits::BitWidth::W128>; + + const int window_step_x = 8; + const auto window_start_x = static_cast<int>(window.x().start()); + const auto window_end_x = static_cast<int>(window.x().end()); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + const auto input_mean = reinterpret_cast<const float16_t *>(mean->ptr_to_element(Coordinates(0, 0))); + const auto input_var = reinterpret_cast<const float16_t *>(var->ptr_to_element(Coordinates(0, 0))); + const auto input_gamma = + (gamma != nullptr) ? reinterpret_cast<const float16_t *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr; + const auto input_beta = + (beta != nullptr) ? reinterpret_cast<const float16_t *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr; + + T activation_functor(act_info); + + const auto epsilon_vec = wrapper::vdup_n(static_cast<float16_t>(epsilon), ExactTagType{}); + execute_window_loop( + win_collapsed, + [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast<const float16_t *>(input.ptr()); + const auto output_ptr = reinterpret_cast<float16_t *>(output.ptr()); + + // Perform core calculations using vector operations + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) + { + // Conctruct vectors + const auto mean_vec = wrapper::vloadq(input_mean + x); + const auto var_vec = wrapper::vloadq(input_var + x); + const auto gamma_vec = (input_gamma != nullptr) + ? wrapper::vloadq(input_gamma + x) + : wrapper::vdup_n(static_cast<float16_t>(1.f), ExactTagType{}); + const auto beta_vec = (input_beta != nullptr) + ? wrapper::vloadq(input_beta + x) + : wrapper::vdup_n(static_cast<float16_t>(0.f), ExactTagType{}); + + // Calculate denominator + const auto denominator = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec)); + + // Calculate x bar + const auto numerator = wrapper::vsub(wrapper::vloadq(input_ptr + x), mean_vec); + const auto x_bar = wrapper::vmul(numerator, denominator); + auto res = wrapper::vmla(beta_vec, x_bar, gamma_vec); + + // Perform fused activation + if (act_info.enabled()) + { + activation_functor(res); + } + + // Store results + wrapper::vstore(output_ptr + x, res); + } + + // Compute left-over elements + for (; x < window_end_x; ++x) + { + // Conctruct vectors + const float16_t gamma = (input_gamma != nullptr) ? input_gamma[x] : 1.f; + const float16_t beta = (input_beta != nullptr) ? input_beta[x] : 0.f; + + const float16_t denominator = sqrt(input_var[x] + epsilon); + const float16_t numerator = input_ptr[x] - input_mean[x]; + const float16_t x_bar = numerator / denominator; + float16_t res = beta + x_bar * gamma; + + // Perform fused activation + if (act_info.enabled()) + { + activation_functor(res); + } + + // Store results + *reinterpret_cast<float16_t *>(output_ptr + x) = res; + } + }, + input, output); +} + +// Fused Batched Normalization with activation functions +static std::map<ActivationLayerInfo::ActivationFunction, BatchNomalizationPtr> fused_map = { + {ActivationLayerInfo::ActivationFunction::RELU, &batch_normalization<detail::relu<float16_t, 8>>}, + {ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, &batch_normalization<detail::brelu<float16_t, 8>>}, + {ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, &batch_normalization<detail::lubrelu<float16_t, 8>>}}; +} // namespace +namespace cpu +{ +void fp16_neon_batch_normalization(ITensor *src, + ITensor *dst, + const ITensor *mean, + const ITensor *var, + const ITensor *beta, + const ITensor *gamma, + float epsilon, + ActivationLayerInfo &act_info, + const Window &window) +{ + if (act_info.enabled()) + { + fused_map[act_info.activation()](src, dst, mean, var, beta, gamma, epsilon, act_info, window); + } + else + { + batch_normalization<detail::dummy<float16_t, 8>>(src, dst, mean, var, beta, gamma, epsilon, act_info, window); + } +} +} // namespace cpu +} // namespace arm_compute + +#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ diff --git a/src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp new file mode 100644 index 0000000000..4e1654ee6b --- /dev/null +++ b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp32.cpp @@ -0,0 +1,176 @@ +/* + * Copyright (c) 2020-2021 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. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensorPack.h" +#include "arm_compute/core/Window.h" + +#include "src/core/NEON/kernels/detail/NEActivationFunctionDetail.h" +#include "src/core/NEON/NEMath.h" +#include "src/core/NEON/wrapper/wrapper.h" + +#include <arm_neon.h> +#include <cmath> +#include <cstddef> + +namespace arm_compute +{ +namespace +{ +using BatchNomalizationPtr = void (*)(ITensor *src, + ITensor *dst, + const ITensor *mean, + const ITensor *var, + const ITensor *beta, + const ITensor *gamma, + float epsilon, + ActivationLayerInfo &act_info, + const Window &window); + +template <typename T> +void batch_normalization(ITensor *src, + ITensor *dst, + const ITensor *mean, + const ITensor *var, + const ITensor *beta, + const ITensor *gamma, + float epsilon, + ActivationLayerInfo &act_info, + const Window &window) +{ + /** SIMD vector tag type. */ + using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<float, wrapper::traits::BitWidth::W128>; + + const int window_step_x = 4; + const auto window_start_x = static_cast<int>(window.x().start()); + const auto window_end_x = static_cast<int>(window.x().end()); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + const auto input_mean = reinterpret_cast<const float *>(mean->ptr_to_element(Coordinates(0, 0))); + const auto input_var = reinterpret_cast<const float *>(var->ptr_to_element(Coordinates(0, 0))); + const auto input_gamma = + (gamma != nullptr) ? reinterpret_cast<const float *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr; + const auto input_beta = + (beta != nullptr) ? reinterpret_cast<const float *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr; + + T activation_functor(act_info); + + const auto epsilon_vec = wrapper::vdup_n(static_cast<float>(epsilon), ExactTagType{}); + execute_window_loop( + win_collapsed, + [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast<const float *>(input.ptr()); + const auto output_ptr = reinterpret_cast<float *>(output.ptr()); + + // Perform core calculations using vector operations + int x = window_start_x; + for (; x <= (window_end_x - window_step_x); x += window_step_x) + { + // Conctruct vectors + const auto mean_vec = wrapper::vloadq(input_mean + x); + const auto var_vec = wrapper::vloadq(input_var + x); + const auto gamma_vec = (input_gamma != nullptr) + ? wrapper::vloadq(input_gamma + x) + : wrapper::vdup_n(static_cast<float>(1.f), ExactTagType{}); + const auto beta_vec = (input_beta != nullptr) + ? wrapper::vloadq(input_beta + x) + : wrapper::vdup_n(static_cast<float>(0.f), ExactTagType{}); + + // Calculate denominator + const auto denominator = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec)); + + // Calculate x bar + const auto numerator = wrapper::vsub(wrapper::vloadq(input_ptr + x), mean_vec); + const auto x_bar = wrapper::vmul(numerator, denominator); + auto res = wrapper::vmla(beta_vec, x_bar, gamma_vec); + + // Perform fused activation + if (act_info.enabled()) + { + activation_functor(res); + } + + // Store results + wrapper::vstore(output_ptr + x, res); + } + + // Compute left-over elements + for (; x < window_end_x; ++x) + { + // Conctruct vectors + const float gamma = (input_gamma != nullptr) ? input_gamma[x] : 1.f; + const float beta = (input_beta != nullptr) ? input_beta[x] : 0.f; + + const float denominator = sqrt(input_var[x] + epsilon); + const float numerator = input_ptr[x] - input_mean[x]; + const float x_bar = numerator / denominator; + float res = beta + x_bar * gamma; + + // Perform fused activation + if (act_info.enabled()) + { + activation_functor(res); + } + + // Store results + *reinterpret_cast<float *>(output_ptr + x) = res; + } + }, + input, output); +} + +// Fused Batched Normalization with activation functions +static std::map<ActivationLayerInfo::ActivationFunction, BatchNomalizationPtr> fused_map = { + {ActivationLayerInfo::ActivationFunction::RELU, &batch_normalization<detail::relu<float, 4>>}, + {ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, &batch_normalization<detail::brelu<float, 4>>}, + {ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, &batch_normalization<detail::lubrelu<float, 4>>}}; +} // namespace +namespace cpu +{ +void fp32_neon_batch_normalization(ITensor *src, + ITensor *dst, + const ITensor *mean, + const ITensor *var, + const ITensor *beta, + const ITensor *gamma, + float epsilon, + ActivationLayerInfo &act_info, + const Window &window) +{ + if (act_info.enabled()) + { + fused_map[act_info.activation()](src, dst, mean, var, beta, gamma, epsilon, act_info, window); + } + else + { + batch_normalization<detail::dummy<float, 4>>(src, dst, mean, var, beta, gamma, epsilon, act_info, window); + } +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp b/src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp new file mode 100644 index 0000000000..48caaa3e63 --- /dev/null +++ b/src/core/NEON/kernels/batchnormalization/impl/SVE/fp16.cpp @@ -0,0 +1,131 @@ +/* + * Copyright (c) 2020-2021,2023 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. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensorPack.h" +#include "arm_compute/core/Window.h" +#include "arm_compute/function_info/ActivationLayerInfo.h" + +#include "src/core/NEON/SVEMath.h" + +#include <cmath> +#include <cstddef> + +#if defined(ARM_COMPUTE_ENABLE_SVE) +#include <arm_sve.h> + +namespace arm_compute +{ +namespace cpu +{ +void fp16_sve_batch_normalization(ITensor *src, + ITensor *dst, + const ITensor *mean, + const ITensor *var, + const ITensor *beta, + const ITensor *gamma, + float epsilon, + ActivationLayerInfo &act_info, + const Window &window) +{ + const auto window_start_x = static_cast<int>(window.x().start()); + const auto window_end_x = static_cast<int>(window.x().end()); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + const auto input_mean = reinterpret_cast<const float16_t *>(mean->ptr_to_element(Coordinates(0, 0))); + const auto input_var = reinterpret_cast<const float16_t *>(var->ptr_to_element(Coordinates(0, 0))); + const auto input_gamma = + (gamma != nullptr) ? reinterpret_cast<const float16_t *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr; + const auto input_beta = + (beta != nullptr) ? reinterpret_cast<const float16_t *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr; + + const auto epsilon_vec = svdup_n_f16(epsilon); + const auto const_1 = svdup_n_f16(1.f); + const auto const_0 = svdup_n_f16(0.f); + const auto va = svdup_n_f16(act_info.a()); + const auto vb = svdup_n_f16(act_info.b()); + execute_window_loop( + win_collapsed, + [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast<const float16_t *>(input.ptr()); + const auto output_ptr = reinterpret_cast<float16_t *>(output.ptr()); + + // Compute S elements per iteration + int x = window_start_x; + svbool_t pg = svwhilelt_b16(x, window_end_x); + do + { + // Conctruct vectors + const auto mean_vec = svld1_f16(pg, input_mean + x); + const auto var_vec = svld1_f16(pg, input_var + x); + const auto gamma_vec = (input_gamma != nullptr) ? svld1_f16(pg, input_gamma + x) : const_1; + const auto beta_vec = (input_beta != nullptr) ? svld1_f16(pg, input_beta + x) : const_0; + + // Calculate denominator + const auto tmp = svadd_f16_z(pg, var_vec, epsilon_vec); + auto denominator = svrsqrte_f16(tmp); + denominator = + svmul_f16_z(pg, svrsqrts_f16(svmul_f16_z(pg, tmp, denominator), denominator), denominator); + denominator = + svmul_f16_z(pg, svrsqrts_f16(svmul_f16_z(pg, tmp, denominator), denominator), denominator); + + // Calculate x bar + const auto numerator = svsub_f16_z(pg, svld1_f16(pg, input_ptr + x), mean_vec); + const auto x_bar = svmul_f16_z(pg, numerator, denominator); + auto res = svmla_f16_z(pg, beta_vec, x_bar, gamma_vec); + + // Perform fused activation + if (act_info.enabled()) + { + if (act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU) + { + res = svmax_f16_z(pg, const_0, res); + } + else if (act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) + { + res = svmin_f16_z(pg, va, svmax_f16_z(pg, const_0, res)); + } + else if (act_info.activation() == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + res = svmin_f16_z(pg, va, svmax_f16_z(pg, vb, res)); + } + } + + // Store results + svst1_f16(pg, output_ptr + x, res); + + x += svcntw(); + pg = svwhilelt_b16(x, window_end_x); + } while (svptest_any(svptrue_b16(), pg)); + }, + input, output); +} +} // namespace cpu +} // namespace arm_compute +#endif // ENABLE_SVE diff --git a/src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp b/src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp new file mode 100644 index 0000000000..df4fbfe607 --- /dev/null +++ b/src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp @@ -0,0 +1,131 @@ +/* + * Copyright (c) 2020-2021,2023 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. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensorPack.h" +#include "arm_compute/core/Window.h" +#include "arm_compute/function_info/ActivationLayerInfo.h" + +#include "src/core/NEON/SVEMath.h" + +#include <cmath> +#include <cstddef> + +#if defined(ARM_COMPUTE_ENABLE_SVE) +#include <arm_sve.h> + +namespace arm_compute +{ +namespace cpu +{ +void fp32_sve_batch_normalization(ITensor *src, + ITensor *dst, + const ITensor *mean, + const ITensor *var, + const ITensor *beta, + const ITensor *gamma, + float epsilon, + ActivationLayerInfo &act_info, + const Window &window) +{ + const auto window_start_x = static_cast<int>(window.x().start()); + const auto window_end_x = static_cast<int>(window.x().end()); + + Window win_collapsed = window.collapse_if_possible(window, Window::DimZ); + win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(src, win_collapsed); + Iterator output(dst, win_collapsed); + + const auto input_mean = reinterpret_cast<const float *>(mean->ptr_to_element(Coordinates(0, 0))); + const auto input_var = reinterpret_cast<const float *>(var->ptr_to_element(Coordinates(0, 0))); + const auto input_gamma = + (gamma != nullptr) ? reinterpret_cast<const float *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr; + const auto input_beta = + (beta != nullptr) ? reinterpret_cast<const float *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr; + + const auto epsilon_vec = svdup_n_f32(epsilon); + const auto const_1 = svdup_n_f32(1.f); + const auto const_0 = svdup_n_f32(0.f); + const auto va = svdup_n_f32(act_info.a()); + const auto vb = svdup_n_f32(act_info.b()); + execute_window_loop( + win_collapsed, + [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast<const float *>(input.ptr()); + const auto output_ptr = reinterpret_cast<float *>(output.ptr()); + + // Compute S elements per iteration + int x = window_start_x; + svbool_t pg = svwhilelt_b32(x, window_end_x); + do + { + // Conctruct vectors + const auto mean_vec = svld1_f32(pg, input_mean + x); + const auto var_vec = svld1_f32(pg, input_var + x); + const auto gamma_vec = (input_gamma != nullptr) ? svld1_f32(pg, input_gamma + x) : const_1; + const auto beta_vec = (input_beta != nullptr) ? svld1_f32(pg, input_beta + x) : const_0; + + // Calculate denominator + const auto tmp = svadd_f32_z(pg, var_vec, epsilon_vec); + auto denominator = svrsqrte_f32(tmp); + denominator = + svmul_f32_z(pg, svrsqrts_f32(svmul_f32_z(pg, tmp, denominator), denominator), denominator); + denominator = + svmul_f32_z(pg, svrsqrts_f32(svmul_f32_z(pg, tmp, denominator), denominator), denominator); + + // Calculate x bar + const auto numerator = svsub_f32_z(pg, svld1_f32(pg, input_ptr + x), mean_vec); + const auto x_bar = svmul_f32_z(pg, numerator, denominator); + auto res = svmla_f32_z(pg, beta_vec, x_bar, gamma_vec); + + // Perform fused activation + if (act_info.enabled()) + { + if (act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU) + { + res = svmax_f32_z(pg, const_0, res); + } + else if (act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU) + { + res = svmin_f32_z(pg, va, svmax_f32_z(pg, const_0, res)); + } + else if (act_info.activation() == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + { + res = svmin_f32_z(pg, va, svmax_f32_z(pg, vb, res)); + } + } + + // Store results + svst1_f32(pg, output_ptr + x, res); + + x += svcntw(); + pg = svwhilelt_b32(x, window_end_x); + } while (svptest_any(svptrue_b32(), pg)); + }, + input, output); +} +} // namespace cpu +} // namespace arm_compute +#endif // ENABLE_SVE diff --git a/src/core/NEON/kernels/batchnormalization/impl/list.h b/src/core/NEON/kernels/batchnormalization/impl/list.h new file mode 100644 index 0000000000..c619788125 --- /dev/null +++ b/src/core/NEON/kernels/batchnormalization/impl/list.h @@ -0,0 +1,59 @@ +/* + * Copyright (c) 2020, 2023 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 ACL_SRC_CORE_NEON_KERNELS_BATCHNORMALIZATION_IMPL_LIST_H +#define ACL_SRC_CORE_NEON_KERNELS_BATCHNORMALIZATION_IMPL_LIST_H + +namespace arm_compute +{ +namespace cpu +{ +#define DECLARE_BATCH_NORMALIZATION_KERNEL(func_name) \ + void func_name(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, \ + const ITensor *gamma, float epsilon, ActivationLayerInfo &act_info, const Window &window) + +DECLARE_BATCH_NORMALIZATION_KERNEL(fp16_neon_batch_normalization); +DECLARE_BATCH_NORMALIZATION_KERNEL(fp16_sve_batch_normalization); +DECLARE_BATCH_NORMALIZATION_KERNEL(fp32_neon_batch_normalization); +DECLARE_BATCH_NORMALIZATION_KERNEL(fp32_sve_batch_normalization); + +#define DECLARE_BATCH_NORMALIZATION_NCHW_KERNEL(func_name) \ + void func_name(const Window &window, ITensor *input, ITensor *output, const ITensor *mean, const ITensor *var, \ + const ITensor *beta, const ITensor *gamma, float epsilon, ActivationLayerInfo act_info) + +DECLARE_BATCH_NORMALIZATION_NCHW_KERNEL(fp16_batch_normalization_nchw_non_fused); +DECLARE_BATCH_NORMALIZATION_NCHW_KERNEL(fp32_batch_normalization_nchw_non_fused); +DECLARE_BATCH_NORMALIZATION_NCHW_KERNEL(fp16_batch_normalization_nchw_non_fused_relu); +DECLARE_BATCH_NORMALIZATION_NCHW_KERNEL(fp16_batch_normalization_nchw_non_fused_brelu); +DECLARE_BATCH_NORMALIZATION_NCHW_KERNEL(fp16_batch_normalization_nchw_non_fused_lubrelu); +DECLARE_BATCH_NORMALIZATION_NCHW_KERNEL(fp32_batch_normalization_nchw_non_fused_relu); +DECLARE_BATCH_NORMALIZATION_NCHW_KERNEL(fp32_batch_normalization_nchw_non_fused_brelu); +DECLARE_BATCH_NORMALIZATION_NCHW_KERNEL(fp32_batch_normalization_nchw_non_fused_lubrelu); + +#undef DECLARE_BATCH_NORMALIZATION_KERNEL +#undef DECLARE_BATCH_NORMALIZATION_NCHW_KERNEL + +} // namespace cpu +} // namespace arm_compute + +#endif // ACL_SRC_CORE_NEON_KERNELS_BATCHNORMALIZATION_IMPL_LIST_H |