From b7b31538eb9137e4d3e8de6d381dcbe9fc58df94 Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Thu, 23 Nov 2017 12:10:21 +0000 Subject: COMPMID-464 Implement Depthwise separable convolution on NEON Change-Id: Iccd686be18381e96bcf09b14c7017c6dda0f38d8 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/109824 Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com Reviewed-by: Pablo Tello --- .../NEON/functions/NEDepthwiseConvolution.cpp | 64 ++++++++++++++++++++++ 1 file changed, 64 insertions(+) (limited to 'src/runtime/NEON/functions/NEDepthwiseConvolution.cpp') diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolution.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolution.cpp index fd8d419fa1..e12bc07464 100644 --- a/src/runtime/NEON/functions/NEDepthwiseConvolution.cpp +++ b/src/runtime/NEON/functions/NEDepthwiseConvolution.cpp @@ -59,4 +59,68 @@ void NEDepthwiseConvolution3x3::run() { NEScheduler::get().schedule(&_bias_kernel, Window::DimX); } +} + +NEDepthwiseConvolution::NEDepthwiseConvolution() + : _im2col_kernel(), _weights_reshape_kernel(), _v2mm_kernel(), _vector_to_tensor_kernel(), _input_reshaped(), _weights_reshaped(), _v2mm_output() +{ +} + +void NEDepthwiseConvolution::configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info) +{ + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 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); + + bool has_bias = (biases != nullptr); + + 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 + ((has_bias) ? 1 : 0); + const size_t conv_size = conv_w * conv_h; + + // Im2Col configuration + 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 TensorInfo info_im2col(shape_im2col, 1, input->info()->data_type(), input->info()->fixed_point_position()); + _input_reshaped.allocator()->init(info_im2col); + _im2col_kernel.configure(input, &_input_reshaped, Size2D(weights_w, weights_h), conv_info, has_bias); + + // Weights reshape configuration + const TensorShape shape_weights_reshape(patch_size, weights_z); + const TensorInfo info_weights_reshape(shape_weights_reshape, 1, weights->info()->data_type(), weights->info()->fixed_point_position()); + _weights_reshaped.allocator()->init(info_weights_reshape); + _weights_reshape_kernel.configure(weights, &_weights_reshaped, biases); + + // GEMV configuration + TensorShape shape_v2mm_out = input->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_v2mm_out(shape_v2mm_out, 1, input->info()->data_type(), input->info()->fixed_point_position()); + _v2mm_output.allocator()->init(info_v2mm_out); + _v2mm_kernel.configure(&_input_reshaped, &_weights_reshaped, &_v2mm_output); + _vector_to_tensor_kernel.configure(&_v2mm_output, output, conv_w, conv_h); + + // Allocate intermediate tensors + _input_reshaped.allocator()->allocate(); + _weights_reshaped.allocator()->allocate(); + _v2mm_output.allocator()->allocate(); +} + +void NEDepthwiseConvolution::run() +{ + NEScheduler::get().schedule(&_im2col_kernel, Window::DimX); + NEScheduler::get().schedule(&_weights_reshape_kernel, Window::DimX); + NEScheduler::get().schedule(&_v2mm_kernel, Window::DimX); + NEScheduler::get().schedule(&_vector_to_tensor_kernel, Window::DimX); } \ No newline at end of file -- cgit v1.2.1