diff options
author | Dana Zlotnik <dana.zlotnik@arm.com> | 2022-01-03 10:59:41 +0200 |
---|---|---|
committer | Dana Zlotnik <dana.zlotnik@arm.com> | 2022-01-12 12:31:14 +0000 |
commit | d7e2ec51239e2075f931e0a9364e0a68534676f1 (patch) | |
tree | fa2e37fdd125caba983d3b9cdabdb8e03921f6a6 /src/cpu/kernels/instancenorm/generic/neon/impl.cpp | |
parent | 5ae8d804d67f57fbfa793800ddcc21a5aff954dd (diff) | |
download | ComputeLibrary-d7e2ec51239e2075f931e0a9364e0a68534676f1.tar.gz |
Decouple NEInstanceNormalizationLayerKernel
Resolves COMPMID-4620
Signed-off-by: Dana Zlotnik <dana.zlotnik@arm.com>
Change-Id: I22c285339840493c9cfd4c1abfbc3768ad4db824
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6871
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Comments-Addressed: 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 | 172 |
1 files changed, 172 insertions, 0 deletions
diff --git a/src/cpu/kernels/instancenorm/generic/neon/impl.cpp b/src/cpu/kernels/instancenorm/generic/neon/impl.cpp new file mode 100644 index 0000000000..e35cf97608 --- /dev/null +++ b/src/cpu/kernels/instancenorm/generic/neon/impl.cpp @@ -0,0 +1,172 @@ +/* + * Copyright (c) 2019-2022 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 "src/cpu/kernels/instancenorm/generic/neon/impl.h" +#include "src/core/NEON/wrapper/wrapper.h" + +namespace arm_compute +{ +class ITensor; +class Window; +namespace cpu +{ +template <typename InputType, typename AccType> +void vector_float_sum(AccType &result, AccType &result_square, const InputType &inputs) +{ + result = wrapper::vadd(result, inputs); + 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) +{ + /** SIMD vector tag type. */ + using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>; + + // Clear X/Y dimensions on execution window as we handle the planes manually + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + win.set(Window::DimY, Window::Dimension(0, 1, 1)); + + constexpr int window_step_x = 16 / sizeof(T); + const unsigned int elements_plane = input->info()->dimension(0) * output->info()->dimension(1); + + Iterator input_it(input, win); + execute_window_loop(win, [&](const Coordinates & id) + { + Window win_plane = window; + win_plane.set(Window::DimX, Window::Dimension(0, 1, 1)); + win_plane.set(Window::DimZ, Window::Dimension(id[2], id[2] + 1, 1)); + win_plane.set(3, Window::Dimension(id[3], id[3] + 1, 1)); + + Iterator input_plane_it(input, win_plane); + Iterator output_plane_it(output, win_plane); + + auto sum_h_w = static_cast<AccType>(0.f); + auto sum_squares_h_w = static_cast<AccType>(0.f); + + execute_window_loop(win_plane, [&](const Coordinates &) + { + const auto input_ptr = reinterpret_cast<const T *>(input_plane_it.ptr()); + + auto vec_sum_h_w = wrapper::vdup_n(static_cast<AccType>(0.f), ExactTagType{}); + auto vec_sum_squares_h_w = wrapper::vdup_n(static_cast<AccType>(0.f), ExactTagType{}); + + // Compute S elements per iteration + int x = window.x().start(); + for(; x <= (window.x().end() - window_step_x); x += window_step_x) + { + auto vec_input_val = wrapper::vloadq(input_ptr + x); + vector_float_sum(vec_sum_h_w, vec_sum_squares_h_w, vec_input_val); + } + + auto vec2_sum_h_w = wrapper::vpadd(wrapper::vgethigh(vec_sum_h_w), wrapper::vgetlow(vec_sum_h_w)); + auto vec2_sum_squares_h_w = wrapper::vpadd(wrapper::vgethigh(vec_sum_squares_h_w), wrapper::vgetlow(vec_sum_squares_h_w)); + + vec2_sum_h_w = wrapper::vpadd(vec2_sum_h_w, vec2_sum_h_w); + vec2_sum_squares_h_w = wrapper::vpadd(vec2_sum_squares_h_w, vec2_sum_squares_h_w); + + sum_h_w += wrapper::vgetlane(vec2_sum_h_w, 0); + sum_squares_h_w += wrapper::vgetlane(vec2_sum_squares_h_w, 0); + + // Compute left-over elements + for(; x < window.x().end(); ++x) + { + const auto value = static_cast<AccType>(*(input_ptr + x)); + sum_h_w += value; + sum_squares_h_w += value * value; + } + }, + input_plane_it, output_plane_it); + + const auto mean_h_w = sum_h_w / elements_plane; + const auto var_h_w = sum_squares_h_w / elements_plane - mean_h_w * mean_h_w; + + const auto multip_h_w = gamma / std::sqrt(var_h_w + epsilon); + const auto vec_mean_h_w = wrapper::vdup_n(static_cast<AccType>(mean_h_w), ExactTagType{}); + const auto vec_multip_h_w = wrapper::vdup_n(static_cast<AccType>(multip_h_w), ExactTagType{}); + const auto vec_beta = wrapper::vdup_n(static_cast<AccType>(beta), ExactTagType{}); + + execute_window_loop(win_plane, [&](const Coordinates &) + { + auto input_ptr = reinterpret_cast<T *>(input_plane_it.ptr()); + auto output_ptr = reinterpret_cast<T *>(output_plane_it.ptr()); + + // Compute S elements per iteration + int x = window.x().start(); + //auto vec_val = wrapper::vdup_n(static_cast<T>(0.0f), ExactTagType{}); + for(; x <= (window.x().end() - window_step_x); x += window_step_x) + { + const auto vec_val = wrapper::vloadq(input_ptr + x); + const auto normalized_vec = vector_float_norm(vec_val, vec_mean_h_w, vec_multip_h_w, vec_beta); + wrapper::vstore(output_ptr + x, normalized_vec); + } + + // Compute left-over elements + for(; x < window.x().end(); ++x) + { + const auto val = static_cast<AccType>(*(input_ptr + x)); + *(output_ptr + x) = static_cast<T>((val - mean_h_w) * multip_h_w + beta); + } + }, + input_plane_it, output_plane_it); + }, + input_it); +} + +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 |