From 6124390be4690ba06c404d56449f7e5d390cef53 Mon Sep 17 00:00:00 2001 From: Sheri Zhang Date: Tue, 12 Jan 2021 18:25:16 +0000 Subject: Make Add kernel and operator stateless - Rename NEArithmeticAdditionKernel to CpuAddKernel Cpu and move files appropriately - Add CpuAdd under src/runtime/cpu/operators Partially resolves: COMPMID-4005 Signed-off-by: Sheri Zhang Change-Id: I1d8d406df9773fea198899f50327407d7125c38d Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4867 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Gian Marco Iodice Reviewed-by: Georgios Pinitas --- src/core/cpu/kernels/add/neon/integer.cpp | 170 ++++++++++++++++++++++++++++++ 1 file changed, 170 insertions(+) create mode 100644 src/core/cpu/kernels/add/neon/integer.cpp (limited to 'src/core/cpu/kernels/add/neon/integer.cpp') diff --git a/src/core/cpu/kernels/add/neon/integer.cpp b/src/core/cpu/kernels/add/neon/integer.cpp new file mode 100644 index 0000000000..24a0ac3b7c --- /dev/null +++ b/src/core/cpu/kernels/add/neon/integer.cpp @@ -0,0 +1,170 @@ +/* + * 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/ITensor.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/Traits.h" +#include "src/core/NEON/wrapper/wrapper.h" +#include "src/core/helpers/WindowHelpers.h" + +namespace arm_compute +{ +namespace cpu +{ +void add_u8_u8_s16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) +{ + // Create input windows + Window win = window; + Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape()); + Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()); + + // Clear X Dimension on execution window as we handle manually + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input1(src0, input1_win); + Iterator input2(src1, input2_win); + Iterator output(dst, win); + + const int window_step_x = 8; + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + + execute_window_loop(win, [&](const Coordinates &) + { + const auto input1_ptr = reinterpret_cast(input1.ptr()); + const auto input2_ptr = reinterpret_cast(input2.ptr()); + const auto output_ptr = reinterpret_cast(output.ptr()); + + if(policy == ConvertPolicy::WRAP) + { + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin1 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input1_ptr + x))); + const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x))); + wrapper::vstore(output_ptr + x, wrapper::vadd(vin1, vin2)); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + *(output_ptr + x) = static_cast(*(input1_ptr + x)) + static_cast(*(input2_ptr + x)); + } + } + else + { + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin1 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input1_ptr + x))); + const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x))); + wrapper::vstore(output_ptr + x, wrapper::vqadd(vin1, vin2)); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + *(output_ptr + x) = wrapper::add_sat(static_cast(*(input1_ptr + x)), + static_cast(*(input2_ptr + x))); + } + } + }, + input1, input2, output); +} + +void add_s16_u8_s16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) +{ + // Create input windows + Window win = window; + Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape()); + Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()); + + // Clear X Dimension on execution window as we handle manually + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + input1_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + input2_win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input1(src0, input1_win); + Iterator input2(src1, input2_win); + Iterator output(dst, win); + + const int window_step_x = 8; + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + + execute_window_loop(win, [&](const Coordinates &) + { + const auto input1_ptr = reinterpret_cast(input1.ptr()); + const auto input2_ptr = reinterpret_cast(input2.ptr()); + const auto output_ptr = reinterpret_cast(output.ptr()); + + if(policy == ConvertPolicy::WRAP) + { + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin1 = wrapper::vloadq(input1_ptr + x); + const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x))); + wrapper::vstore(output_ptr + x, wrapper::vadd(vin1, vin2)); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + *(output_ptr + x) = *(input1_ptr + x) + static_cast(*(input2_ptr + x)); + } + } + else + { + // Compute S elements per iteration + int x = window_start_x; + for(; x <= (window_end_x - window_step_x); x += window_step_x) + { + const auto vin1 = wrapper::vloadq(input1_ptr + x); + const auto vin2 = vreinterpretq_s16_u16(wrapper::vmovl(wrapper::vload(input2_ptr + x))); + wrapper::vstore(output_ptr + x, wrapper::vqadd(vin1, vin2)); + } + + // Compute left-over elements + for(; x < window_end_x; ++x) + { + *(output_ptr + x) = wrapper::add_sat(*(input1_ptr + x), static_cast(*(input2_ptr + x))); + } + } + }, + input1, input2, output); +} + +void add_u8_s16_s16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window) +{ + // Simply swap the two input buffers: + add_s16_u8_s16_neon(src1, src0, dst, policy, window); +} +} // namespace cpu +} // namespace arm_compute \ No newline at end of file -- cgit v1.2.1