From 027bcef3dbf63d4c9f56149925be4b03e39274b5 Mon Sep 17 00:00:00 2001 From: Dana Zlotnik Date: Mon, 27 Dec 2021 17:35:00 +0200 Subject: Decouple NEMeanStdDevNormalizationKernel Resolves COMPMID-4617 Change-Id: Ic8793aaf64c6137f848f39c62e33b44ae79ad21d Signed-off-by: Dana Zlotnik Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6870 Tested-by: Arm Jenkins Reviewed-by: Giorgio Arena Comments-Addressed: Arm Jenkins --- .../kernels/meanstddevnorm/generic/neon/fp16.cpp | 37 +++++++ .../kernels/meanstddevnorm/generic/neon/fp32.cpp | 35 +++++++ .../kernels/meanstddevnorm/generic/neon/impl.cpp | 110 +++++++++++++++++++++ src/cpu/kernels/meanstddevnorm/generic/neon/impl.h | 38 +++++++ src/cpu/kernels/meanstddevnorm/list.h | 39 ++++++++ 5 files changed, 259 insertions(+) create mode 100644 src/cpu/kernels/meanstddevnorm/generic/neon/fp16.cpp create mode 100644 src/cpu/kernels/meanstddevnorm/generic/neon/fp32.cpp create mode 100644 src/cpu/kernels/meanstddevnorm/generic/neon/impl.cpp create mode 100644 src/cpu/kernels/meanstddevnorm/generic/neon/impl.h create mode 100644 src/cpu/kernels/meanstddevnorm/list.h (limited to 'src/cpu/kernels/meanstddevnorm') diff --git a/src/cpu/kernels/meanstddevnorm/generic/neon/fp16.cpp b/src/cpu/kernels/meanstddevnorm/generic/neon/fp16.cpp new file mode 100644 index 0000000000..3e712b5593 --- /dev/null +++ b/src/cpu/kernels/meanstddevnorm/generic/neon/fp16.cpp @@ -0,0 +1,37 @@ +/* + * Copyright (c) 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. + */ +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) +#include "src/cpu/kernels/meanstddevnorm/generic/neon/impl.h" + +namespace arm_compute +{ +namespace cpu +{ +void neon_fp16_meanstddevnorm(ITensor *input, ITensor *output, float epsilon, const Window &window) +{ + return mean_stddev_normalization(input, output, epsilon, window); +} +} // namespace cpu +} // namespace arm_compute +#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ diff --git a/src/cpu/kernels/meanstddevnorm/generic/neon/fp32.cpp b/src/cpu/kernels/meanstddevnorm/generic/neon/fp32.cpp new file mode 100644 index 0000000000..4bff26b036 --- /dev/null +++ b/src/cpu/kernels/meanstddevnorm/generic/neon/fp32.cpp @@ -0,0 +1,35 @@ +/* + * Copyright (c) 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/meanstddevnorm/generic/neon/impl.h" + +namespace arm_compute +{ +namespace cpu +{ +void neon_fp32_meanstddevnorm(ITensor *input, ITensor *output, float epsilon, const Window &window) +{ + return mean_stddev_normalization(input, output, epsilon, window); +} +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/meanstddevnorm/generic/neon/impl.cpp b/src/cpu/kernels/meanstddevnorm/generic/neon/impl.cpp new file mode 100644 index 0000000000..be07ea78e4 --- /dev/null +++ b/src/cpu/kernels/meanstddevnorm/generic/neon/impl.cpp @@ -0,0 +1,110 @@ +/* + * 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/meanstddevnorm/generic/neon/impl.h" +#include "src/core/NEON/wrapper/wrapper.h" + +namespace arm_compute +{ +namespace cpu +{ +template +void mean_stddev_normalization(ITensor *input, ITensor *output, float epsilon, const Window &window) +{ + using ExactTagType = typename wrapper::traits::neon_vector::tag_type; + + // Set build options + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + const int window_step_x = size; + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + + Iterator input_itr(input, win); + Iterator output_itr(output, win); + + execute_window_loop(win, [&](const Coordinates &) + { + int x = window_start_x; + auto in_ptr = reinterpret_cast(input_itr.ptr()); + auto out_ptr = reinterpret_cast(output_itr.ptr()); + + auto sum_vec = wrapper::vdup_n(static_cast(0.f), ExactTagType{}); + auto sum_sq_vec = wrapper::vdup_n(static_cast(0.f), ExactTagType{}); + + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + auto data = wrapper::vloadq(in_ptr + x); + sum_vec = wrapper::vadd(sum_vec, data); + sum_sq_vec = wrapper::vadd(sum_sq_vec, wrapper::vmul(data, data)); + } + + auto sum_carry_res = wrapper::vpadd(wrapper::vgethigh(sum_vec), wrapper::vgetlow(sum_vec)); + auto sum_sq_carry_res = wrapper::vpadd(wrapper::vgethigh(sum_sq_vec), wrapper::vgetlow(sum_sq_vec)); + for(int i = 0; i < size / 4; ++i) + { + sum_carry_res = wrapper::vpadd(sum_carry_res, sum_carry_res); + sum_sq_carry_res = wrapper::vpadd(sum_sq_carry_res, sum_sq_carry_res); + } + + auto sum = wrapper::vgetlane(sum_carry_res, 0); + auto sum_sq = wrapper::vgetlane(sum_sq_carry_res, 0); + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + ScalarType data = *(in_ptr + x); + sum += data; + sum_sq += data * data; + } + + ScalarType mean = sum / input->info()->dimension(0); + ScalarType var = (sum_sq / input->info()->dimension(0)) - (mean * mean); + ScalarType stddev_inv = 1.f / sqrt(var + epsilon); + + auto mean_vec = wrapper::vdup_n(mean, ExactTagType{}); + auto stddev_inv_vec = wrapper::vdup_n(stddev_inv, ExactTagType{}); + for(x = window_start_x; x <= (window_end_x - window_step_x); x += window_step_x) + { + auto data = wrapper::vloadq(in_ptr + x); + auto res = wrapper::vmul(wrapper::vsub(data, mean_vec), stddev_inv_vec); + // Store results + wrapper::vstore(out_ptr + x, res); + } + for(; x < window_end_x; ++x) + { + *(out_ptr + x) = (*(in_ptr + x) - mean) * stddev_inv; + } + }, + input_itr, output_itr); +} +template void mean_stddev_normalization(ITensor *input, ITensor *output, float epsilon, const Window &window); + +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) +template void mean_stddev_normalization(ITensor *input, ITensor *output, float epsilon, const Window &window); +#endif //defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) + +} // namespace cpu +} // namespace arm_compute diff --git a/src/cpu/kernels/meanstddevnorm/generic/neon/impl.h b/src/cpu/kernels/meanstddevnorm/generic/neon/impl.h new file mode 100644 index 0000000000..6466506f06 --- /dev/null +++ b/src/cpu/kernels/meanstddevnorm/generic/neon/impl.h @@ -0,0 +1,38 @@ +/* + * Copyright (c) 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. + */ +#ifndef SRC_CORE_SVE_KERNELS_MEANSTDDEVNORM_IMPL_H +#define SRC_CORE_SVE_KERNELS_MEANSTDDEVNORM_IMPL_H +#include "arm_compute/core/Helpers.h" + +namespace arm_compute +{ +namespace cpu +{ +template +void mean_stddev_normalization(ITensor *_input, ITensor *_output, float _epsilon, const Window &window); + +} // namespace cpu +} // namespace arm_compute + +#endif //define SRC_CORE_SVE_KERNELS_MEANSTDDEVNORM_IMPL_H diff --git a/src/cpu/kernels/meanstddevnorm/list.h b/src/cpu/kernels/meanstddevnorm/list.h new file mode 100644 index 0000000000..ac9cb37d23 --- /dev/null +++ b/src/cpu/kernels/meanstddevnorm/list.h @@ -0,0 +1,39 @@ +/* + * Copyright (c) 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. + */ +#ifndef SRC_CORE_NEON_KERNELS_MEANSTDDEVNORM_LIST_H +#define SRC_CORE_NEON_KERNELS_MEANSTDDEVNORM_LIST_H +namespace arm_compute +{ +namespace cpu +{ +#define DECLARE_MEANSTDDEVNORM_KERNEL(func_name) \ + void func_name(ITensor *input, ITensor *output, float epsilon, const Window &window) + +DECLARE_MEANSTDDEVNORM_KERNEL(neon_fp32_meanstddevnorm); +DECLARE_MEANSTDDEVNORM_KERNEL(neon_fp16_meanstddevnorm); + +#undef DECLARE_MEANSTDDEVNORM_KERNEL +} // namespace cpu +} // namespace arm_compute +#endif //SRC_CORE_NEON_KERNELS_MEANSTDDEVNORM_LIST_H -- cgit v1.2.1