From 5a97b28d483eefa5810f5cf57356086090c8c894 Mon Sep 17 00:00:00 2001 From: George Wort Date: Fri, 21 Dec 2018 16:21:04 +0000 Subject: COMPMID-1756: NEON: Implement RSqrt, Exp Change-Id: I6b140b8868b04f7d3032a51831a80829e8e1560e Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/165590 Tested-by: bsgcomp Reviewed-by: Anthony Barbier Reviewed-on: https://review.mlplatform.org/470 Reviewed-by: Georgios Pinitas Tested-by: Arm Jenkins --- src/core/NEON/kernels/NEElementwiseUnaryKernel.cpp | 206 +++++++++++++++++++++ .../NEON/functions/NEElementwiseUnaryLayer.cpp | 54 ++++++ .../NEON/functions/NEFuseBatchNormalization.cpp | 6 +- src/runtime/NEON/functions/NEReduceMean.cpp | 6 +- 4 files changed, 266 insertions(+), 6 deletions(-) create mode 100644 src/core/NEON/kernels/NEElementwiseUnaryKernel.cpp create mode 100644 src/runtime/NEON/functions/NEElementwiseUnaryLayer.cpp (limited to 'src') diff --git a/src/core/NEON/kernels/NEElementwiseUnaryKernel.cpp b/src/core/NEON/kernels/NEElementwiseUnaryKernel.cpp new file mode 100644 index 0000000000..7ecc4d1c44 --- /dev/null +++ b/src/core/NEON/kernels/NEElementwiseUnaryKernel.cpp @@ -0,0 +1,206 @@ +/* + * Copyright (c) 2018-2019 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/NEON/kernels/NEElementwiseUnaryKernel.h" + +#include "arm_compute/core/CPP/Validate.h" +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/IAccessWindow.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/NEON/NEAsymm.h" +#include "arm_compute/core/NEON/NEFixedPoint.h" +#include "arm_compute/core/NEON/wrapper/wrapper.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Validate.h" + +#include +#include +#include +#include +#include + +namespace arm_compute +{ +class Coordinates; + +namespace +{ +template +inline ScalarType elementwise_op_scalar(const ScalarType &a) +{ + switch(op) + { + case ElementWiseUnary::RSQRT: + return 1 / sqrt(a); + case ElementWiseUnary::EXP: + return std::exp(a); + default: + ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); + } +} + +template +inline VectorType elementwise_op(const VectorType &a) +{ + switch(op) + { + case ElementWiseUnary::RSQRT: + return wrapper::vinvsqrt(a); + case ElementWiseUnary::EXP: + return wrapper::vexpq(a); + default: + ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); + } +} + +template +void elementwise_op(const ITensor *in, ITensor *out, const Window &window) +{ + const int window_step_x = 16 / sizeof(ScalarType); + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + Iterator input(in, win); + Iterator output(out, win); + + execute_window_loop(win, [&](const Coordinates & id) + { + auto output_ptr = reinterpret_cast(output.ptr()); + const auto input_ptr = reinterpret_cast(input.ptr()); + + int x = window_start_x; + for(; x <= window_end_x - window_step_x; x += window_step_x) + { + wrapper::vstore(output_ptr + x, elementwise_op(wrapper::vloadq(input_ptr + x))); + } + for(; x < window_end_x; ++x) + { + *(output_ptr + x) = elementwise_op_scalar(*(input_ptr + x)); + } + }, + input, output); +} + +template +std::function +configure_func(const ITensor *input, ITensor *output) +{ + std::string function_to_call("op_"); + function_to_call += string_from_data_type(input->info()->data_type()) + "_"; + function_to_call += string_from_data_type(output->info()->data_type()); + + static std::map map_function = + { + { "op_F32_F32", &elementwise_op } + }; +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + map_function["op_F16_F16"] = &elementwise_op; +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + + auto it = map_function.find(function_to_call); + + if(it != map_function.end()) + { + auto func = it->second; + return [func](const ITensor * input, ITensor * output, const Window & window) + { + func(input, output, window); + }; + } + return nullptr; +} +} // namespace + +NEElementwiseUnaryKernel::NEElementwiseUnaryKernel() + : _function(nullptr), _input(nullptr), _output(nullptr) +{ +} + +void NEElementwiseUnaryKernel::configure(ElementWiseUnary op, const ITensor *input, ITensor *output) +{ + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input->info(), *output->info())); + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); + + // Configure kernel window + const std::pair broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(*input->info()); + const TensorShape &out_shape = broadcast_pair.first; + const ValidRegion &valid_region = broadcast_pair.second; + + // Auto initialize output if not initialized + auto_init_if_empty(*output->info(), out_shape, 1, input->info()->data_type()); + + Window win = calculate_max_window(valid_region); + + _input = input; + _output = output; + + INEKernel::configure(win); + + switch(op) + { + case ElementWiseUnary::RSQRT: + _function = configure_func(input, output); + break; + case ElementWiseUnary::EXP: + _function = configure_func(input, output); + break; + default: + ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); + } +} + +Status NEElementwiseUnaryKernel::validate_arguments(const ITensorInfo &input, const ITensorInfo &output) +{ + ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(&input); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input, 1, DataType::F16, DataType::F32); + + // Validate in case of configured output + if(output.total_size() > 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input, &output); + } + + return Status{}; +} + +Status NEElementwiseUnaryKernel::validate(ElementWiseUnary op, const ITensorInfo *input, const ITensorInfo *output) +{ + ARM_COMPUTE_UNUSED(op); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*input, *output)); + return Status{}; +} + +void NEElementwiseUnaryKernel::run(const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); + ARM_COMPUTE_ERROR_ON(_function == nullptr); + _function(_input, _output, window); +} +} // namespace arm_compute diff --git a/src/runtime/NEON/functions/NEElementwiseUnaryLayer.cpp b/src/runtime/NEON/functions/NEElementwiseUnaryLayer.cpp new file mode 100644 index 0000000000..10142c732f --- /dev/null +++ b/src/runtime/NEON/functions/NEElementwiseUnaryLayer.cpp @@ -0,0 +1,54 @@ +/* + * Copyright (c) 2018-2019 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/runtime/NEON/functions/NEElementwiseUnaryLayer.h" + +#include "arm_compute/core/NEON/kernels/NEElementwiseUnaryKernel.h" +#include "support/ToolchainSupport.h" + +#include + +namespace arm_compute +{ +void NERsqrtLayer::configure(const ITensor *input, ITensor *output) +{ + auto k = arm_compute::support::cpp14::make_unique(); + k->configure(ElementWiseUnary::RSQRT, input, output); + _kernel = std::move(k); +} +Status NERsqrtLayer::validate(const ITensorInfo *input, const ITensorInfo *output) +{ + return NEElementwiseUnaryKernel::validate(ElementWiseUnary::RSQRT, input, output); +} + +void NEExpLayer::configure(const ITensor *input, ITensor *output) +{ + auto k = arm_compute::support::cpp14::make_unique(); + k->configure(ElementWiseUnary::EXP, input, output); + _kernel = std::move(k); +} +Status NEExpLayer::validate(const ITensorInfo *input, const ITensorInfo *output) +{ + return NEElementwiseUnaryKernel::validate(ElementWiseUnary::EXP, input, output); +} +} // namespace arm_compute diff --git a/src/runtime/NEON/functions/NEFuseBatchNormalization.cpp b/src/runtime/NEON/functions/NEFuseBatchNormalization.cpp index 7f17a15cba..dc48731c4d 100644 --- a/src/runtime/NEON/functions/NEFuseBatchNormalization.cpp +++ b/src/runtime/NEON/functions/NEFuseBatchNormalization.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -14,9 +14,9 @@ * copies or substantial portions of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INNEUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * 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 NEAIM, DAMAGES OR OTHER + * 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. diff --git a/src/runtime/NEON/functions/NEReduceMean.cpp b/src/runtime/NEON/functions/NEReduceMean.cpp index 0b022df729..c5229daae3 100644 --- a/src/runtime/NEON/functions/NEReduceMean.cpp +++ b/src/runtime/NEON/functions/NEReduceMean.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -14,9 +14,9 @@ * copies or substantial portions of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INNEUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * 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 NEAIM, DAMAGES OR OTHER + * 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. -- cgit v1.2.1