From ebcebf1dee7f8314976b1e0cabd62b4cf893d765 Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Wed, 21 Oct 2020 00:04:14 +0100 Subject: COMPMID-3638: Move NEON kernels Signed-off-by: Michalis Spyrou Change-Id: Ieed3e4bc8be7fef80c90c5094599b477a56fc473 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4285 Comments-Addressed: Arm Jenkins Reviewed-by: Georgios Pinitas Tested-by: Arm Jenkins --- src/runtime/NEON/functions/NEFastCorners.cpp | 35 ++++++++++++++++++---------- 1 file changed, 23 insertions(+), 12 deletions(-) (limited to 'src/runtime/NEON/functions/NEFastCorners.cpp') diff --git a/src/runtime/NEON/functions/NEFastCorners.cpp b/src/runtime/NEON/functions/NEFastCorners.cpp index 303c593f84..1bde3cc508 100644 --- a/src/runtime/NEON/functions/NEFastCorners.cpp +++ b/src/runtime/NEON/functions/NEFastCorners.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016-2019 Arm Limited. + * Copyright (c) 2016-2020 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -25,15 +25,21 @@ #include "arm_compute/core/Error.h" #include "arm_compute/core/ITensor.h" -#include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h" #include "arm_compute/core/PixelValue.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Validate.h" #include "arm_compute/runtime/Array.h" #include "arm_compute/runtime/NEON/NEScheduler.h" #include "arm_compute/runtime/TensorAllocator.h" +#include "src/core/NEON/kernels/NEFastCornersKernel.h" +#include "src/core/NEON/kernels/NEFillArrayKernel.h" +#include "src/core/NEON/kernels/NEFillBorderKernel.h" +#include "src/core/NEON/kernels/NENonMaximaSuppression3x3Kernel.h" +#include "support/MemorySupport.h" -using namespace arm_compute; +namespace arm_compute +{ +NEFastCorners::~NEFastCorners() = default; NEFastCorners::NEFastCorners(std::shared_ptr memory_manager) : _memory_group(std::move(memory_manager)), @@ -62,24 +68,28 @@ void NEFastCorners::configure(IImage *input, float threshold, bool nonmax_suppre _output.allocator()->init(tensor_info); _memory_group.manage(&_output); + _fast_corners_kernel = arm_compute::support::cpp14::make_unique(); + _border_handler = arm_compute::support::cpp14::make_unique(); + _fill_kernel = arm_compute::support::cpp14::make_unique(); // If border is UNDEFINED _fast_corners_kernel will operate in xwindow (3, // width - 3) and ywindow (3, height -3) so the output image will leave the // pixels on the borders unchanged. This is reflected in the valid region // of the output. The non maxima suppression is only run on the valid // pixels. - _fast_corners_kernel.configure(input, &_output, threshold, nonmax_suppression, BorderMode::UNDEFINED == border_mode); - _border_handler.configure(input, _fast_corners_kernel.border_size(), border_mode, constant_border_value); + _fast_corners_kernel->configure(input, &_output, threshold, nonmax_suppression, BorderMode::UNDEFINED == border_mode); + _border_handler->configure(input, _fast_corners_kernel->border_size(), border_mode, constant_border_value); if(!_non_max) { - _fill_kernel.configure(&_output, 1 /* we keep all texels >0 */, corners); + _fill_kernel->configure(&_output, 1 /* we keep all texels >0 */, corners); } else { _suppressed.allocator()->init(tensor_info); _memory_group.manage(&_suppressed); - _nonmax_kernel.configure(&_output, &_suppressed, BorderMode::UNDEFINED == border_mode); - _fill_kernel.configure(&_suppressed, 1 /* we keep all texels >0 */, corners); + _nonmax_kernel = arm_compute::support::cpp14::make_unique(); + _nonmax_kernel->configure(&_output, &_suppressed, BorderMode::UNDEFINED == border_mode); + _fill_kernel->configure(&_suppressed, 1 /* we keep all texels >0 */, corners); // Allocate intermediate tensors _suppressed.allocator()->allocate(); @@ -91,16 +101,17 @@ void NEFastCorners::configure(IImage *input, float threshold, bool nonmax_suppre void NEFastCorners::run() { - NEScheduler::get().schedule(&_border_handler, Window::DimZ); + NEScheduler::get().schedule(_border_handler.get(), Window::DimZ); MemoryGroupResourceScope scope_mg(_memory_group); - NEScheduler::get().schedule(&_fast_corners_kernel, Window::DimY); + NEScheduler::get().schedule(_fast_corners_kernel.get(), Window::DimY); if(_non_max) { - NEScheduler::get().schedule(&_nonmax_kernel, Window::DimY); + NEScheduler::get().schedule(_nonmax_kernel.get(), Window::DimY); } - NEScheduler::get().schedule(&_fill_kernel, Window::DimY); + NEScheduler::get().schedule(_fill_kernel.get(), Window::DimY); } +} // namespace arm_compute -- cgit v1.2.1