diff options
author | Giorgio Arena <giorgio.arena@arm.com> | 2017-08-23 16:36:24 +0100 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:35:24 +0000 |
commit | 9fe414430c3c989b1cdc79d41e031495aed2cb7c (patch) | |
tree | 5f893eeb81fe75824918273f49171c85f108fe4e /src/runtime/CL/functions | |
parent | 580f4c7f6c2cbe320c606eb6ba73b3f4a20791ef (diff) | |
download | ComputeLibrary-9fe414430c3c989b1cdc79d41e031495aed2cb7c.tar.gz |
COMPMID-452 CL Generic Depthwise Convolution implementation.
Change-Id: I115e48fe6ce5e281f3791aa5d80fdc754cdd2b5e
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/85082
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Diffstat (limited to 'src/runtime/CL/functions')
-rw-r--r-- | src/runtime/CL/functions/CLDepthwiseConvolution.cpp | 85 |
1 files changed, 80 insertions, 5 deletions
diff --git a/src/runtime/CL/functions/CLDepthwiseConvolution.cpp b/src/runtime/CL/functions/CLDepthwiseConvolution.cpp index 7dac885ed0..22c037fc2a 100644 --- a/src/runtime/CL/functions/CLDepthwiseConvolution.cpp +++ b/src/runtime/CL/functions/CLDepthwiseConvolution.cpp @@ -30,23 +30,98 @@ using namespace arm_compute; -CLDepthwiseConvolution::CLDepthwiseConvolution() +CLDepthwiseConvolution3x3::CLDepthwiseConvolution3x3() : _kernel(), _border_handler() { } -void CLDepthwiseConvolution::configure(ICLTensor *input, ICLTensor *output, const ICLTensor *weights, const PadStrideInfo &conv_info) +void CLDepthwiseConvolution3x3::configure(ICLTensor *input, ICLTensor *output, const ICLTensor *weights, const PadStrideInfo &conv_info) { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F32); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); _kernel.configure(input, output, weights, conv_info); _border_handler.configure(input, _kernel.border_size(), BorderMode::CONSTANT, PixelValue(0)); } -void CLDepthwiseConvolution::run() +void CLDepthwiseConvolution3x3::run() { CLScheduler::get().enqueue(_border_handler); CLScheduler::get().enqueue(_kernel); -}
\ No newline at end of file +} + +CLDepthwiseConvolution::CLDepthwiseConvolution() + : _im2col_kernel(), _weights_reshape_kernel(), _v2mm_kernel(), _vector_to_tensor_kernel(), _v2mm_input_fill_border(), _v2mm_weights_fill_border(), _input_reshaped(), _weights_reshaped(), + _v2mm_output() +{ +} + +void CLDepthwiseConvolution::configure(ICLTensor *input, ICLTensor *output, const ICLTensor *weights, const PadStrideInfo &conv_info) +{ + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); + ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != weights->info()->dimension(2)); + + const size_t weights_w = weights->info()->dimension(0); + const size_t weights_h = weights->info()->dimension(1); + const size_t weights_z = weights->info()->dimension(2); + + unsigned int conv_w = 0; + unsigned int conv_h = 0; + std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights_w, weights_h, conv_info); + + // Set up intermediate tensors + const size_t patch_size = weights_w * weights_h; + const size_t conv_size = conv_w * conv_h; + + TensorShape shape_im2col = input->info()->tensor_shape(); + shape_im2col.set(0, patch_size); + shape_im2col.set(1, conv_size); + shape_im2col.set(2, weights_z); + + const TensorShape shape_weights_reshape(patch_size, weights_z); + TensorShape shape_v2mm_out = output->info()->tensor_shape(); + shape_v2mm_out.set(0, conv_size * weights_z); + shape_v2mm_out.set(1, 1); + shape_v2mm_out.set(2, 1); + + const TensorInfo info_im2col(shape_im2col, 1, input->info()->data_type(), input->info()->fixed_point_position()); + const TensorInfo info_weights_reshape(shape_weights_reshape, 1, weights->info()->data_type(), weights->info()->fixed_point_position()); + const TensorInfo info_v2mm_out(shape_v2mm_out, 1, input->info()->data_type(), input->info()->fixed_point_position()); + + _input_reshaped.allocator()->init(info_im2col); + _weights_reshaped.allocator()->init(info_weights_reshape); + _v2mm_output.allocator()->init(info_v2mm_out); + + // Configure kernels + _im2col_kernel.configure(input, &_input_reshaped, Size2D(weights_w, weights_h), conv_info); + _weights_reshape_kernel.configure(weights, &_weights_reshaped); + _v2mm_kernel.configure(&_input_reshaped, &_weights_reshaped, &_v2mm_output); + _vector_to_tensor_kernel.configure(&_v2mm_output, output, conv_w, conv_h); + + BorderSize border_size = _v2mm_kernel.border_size(); + _v2mm_input_fill_border.configure(&_input_reshaped, border_size, BorderMode::CONSTANT, PixelValue(0)); + + border_size.bottom = 0; + _v2mm_weights_fill_border.configure(&_weights_reshaped, border_size, BorderMode::CONSTANT, PixelValue(0)); + + // Allocate intermediate tensors + _input_reshaped.allocator()->allocate(); + _weights_reshaped.allocator()->allocate(); + _v2mm_output.allocator()->allocate(); +} + +void CLDepthwiseConvolution::run() +{ + CLScheduler::get().enqueue(_im2col_kernel); + + CLScheduler::get().enqueue(_weights_reshape_kernel); + + CLScheduler::get().enqueue(_v2mm_input_fill_border); + CLScheduler::get().enqueue(_v2mm_weights_fill_border); + CLScheduler::get().enqueue(_v2mm_kernel); + + CLScheduler::get().enqueue(_vector_to_tensor_kernel); +} |