diff options
author | Michalis Spyrou <michalis.spyrou@arm.com> | 2018-02-23 15:01:52 +0000 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:48:33 +0000 |
commit | 33a6990ee6ba7bf85b88822d9723060262d00785 (patch) | |
tree | 104e1b614b79a891e78fb7796a91c1981b4d6dd1 /src/runtime/NEON | |
parent | 6a8eb0c9c8d3f71ed4e718464c6d22ca0ffbbdd8 (diff) | |
download | ComputeLibrary-33a6990ee6ba7bf85b88822d9723060262d00785.tar.gz |
COMPMID-540 Replace NEDeconvolutionLayerUpsampleKernel with NEScaleKernel
Change-Id: Ic29557cca24447ef40fa2cfca84f208b4d43f8de
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/122180
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele DiGiorgio <michele.digiorgio@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Diffstat (limited to 'src/runtime/NEON')
-rw-r--r-- | src/runtime/NEON/functions/NEDeconvolutionLayer.cpp | 37 |
1 files changed, 7 insertions, 30 deletions
diff --git a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp index c1ba5dd36e..693d7a4f70 100644 --- a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp @@ -34,6 +34,7 @@ using namespace arm_compute::misc::shape_calculator; NEDeconvolutionLayer::NEDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT : _memory_group(std::move(memory_manager)), _conv_f(), + _upsample_f(), _scaled_output(), _input(nullptr), _info(), @@ -79,44 +80,20 @@ void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, con // Allocate auxiliary tensors _scaled_output.allocator()->allocate(); + + // configure upsample function + _upsample_f.configure(input, &_scaled_output, info, inner_border_right, inner_border_top); } void NEDeconvolutionLayer::run() { _memory_group.acquire(); - // Initialize _scaled_output buffer - const int width_in = _input->info()->dimension(0); - const int height_in = _input->info()->dimension(1); - const int width_scaled = _scaled_output.info()->dimension(0); - const int height_scaled = _scaled_output.info()->dimension(1); - const int num_2d_slices = _input->info()->tensor_shape().total_size() / (width_in * height_in); - const int stride_x = _info.stride().first; - const int stride_y = _info.stride().second; - - std::fill_n(_scaled_output.buffer(), _scaled_output.info()->total_size(), 0); - - // scaled_output is the input for the forward convolution. We copy the input elements to scaled_output - // and insert rows and columns with zeroes depending on the stride values. - for(int slice = 0; slice < num_2d_slices; ++slice) - { - const int start_x = _info.pad().first; - const int start_y = _inner_border.second + _info.pad().second; - const int end_y = height_scaled - _info.pad().second; - const int end_x = width_scaled - _inner_border.first - _info.pad().first; - - for(int yi = start_y, in_y = 0; yi < end_y; yi += stride_y, in_y++) - { - for(int xi = start_x, in_x = 0; xi < end_x; xi += stride_x, in_x++) - { - const auto in = *(reinterpret_cast<float *>(_input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(in_x, in_y, slice)))); - *(reinterpret_cast<float *>(_scaled_output.buffer() + _scaled_output.info()->offset_element_in_bytes(Coordinates(xi, yi, slice)))) = in; - } - } - } + // Run upsample kernel + _upsample_f.run(); // Run convolution layer _conv_f.run(); _memory_group.release(); -} +}
\ No newline at end of file |