From 04a8f8c4994f1c32b3f16a832c0e6f2599364c02 Mon Sep 17 00:00:00 2001 From: Giorgio Arena Date: Thu, 23 Nov 2017 11:45:24 +0000 Subject: COMPMID-692 Consistent names for the interfaces Change-Id: I4b1f3f0da9ff5342c7de7083736fe91871d14e5b Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110351 Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com Reviewed-by: Georgios Pinitas Reviewed-by: Anthony Barbier --- .../CL/functions/CLDepthwiseConvolutionLayer.cpp | 138 +++++++++++++++++++++ 1 file changed, 138 insertions(+) create mode 100644 src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp (limited to 'src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp') diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp new file mode 100644 index 0000000000..02273fe08b --- /dev/null +++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp @@ -0,0 +1,138 @@ +/* + * Copyright (c) 2017 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/CL/functions/CLDepthwiseConvolutionLayer.h" + +#include "arm_compute/core/CL/ICLTensor.h" +#include "arm_compute/core/PixelValue.h" +#include "arm_compute/runtime/CL/CLScheduler.h" +#include "support/ToolchainSupport.h" + +using namespace arm_compute; + +CLDepthwiseConvolutionLayer3x3::CLDepthwiseConvolutionLayer3x3() + : _kernel(), _border_handler() +{ +} + +void CLDepthwiseConvolutionLayer3x3::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info) +{ + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F32); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); + + _kernel.set_target(CLScheduler::get().target()); + _kernel.configure(input, weights, biases, output, conv_info); + + // Configure border handler + PixelValue &&zero_value(0.f); + if(is_data_type_quantized_asymmetric(input->info()->data_type())) + { + zero_value = PixelValue(static_cast(input->info()->quantization_info().offset)); + } + _border_handler.configure(input, _kernel.border_size(), BorderMode::CONSTANT, zero_value); +} + +void CLDepthwiseConvolutionLayer3x3::run() +{ + CLScheduler::get().enqueue(_border_handler); + CLScheduler::get().enqueue(_kernel); +} + +CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayer() + : _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 CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *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); + + const bool has_bias = (biases != nullptr); + const GPUTarget gpu_target = CLScheduler::get().target(); + + 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.set_target(gpu_target); + _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.set_target(gpu_target); + _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 CLDepthwiseConvolutionLayer::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); +} -- cgit v1.2.1