diff options
Diffstat (limited to 'runtimes/neurun/backend/cpu/kernel/DepthwiseConvolutionLayer.cc')
-rw-r--r-- | runtimes/neurun/backend/cpu/kernel/DepthwiseConvolutionLayer.cc | 138 |
1 files changed, 138 insertions, 0 deletions
diff --git a/runtimes/neurun/backend/cpu/kernel/DepthwiseConvolutionLayer.cc b/runtimes/neurun/backend/cpu/kernel/DepthwiseConvolutionLayer.cc new file mode 100644 index 000000000..1c750e0e1 --- /dev/null +++ b/runtimes/neurun/backend/cpu/kernel/DepthwiseConvolutionLayer.cc @@ -0,0 +1,138 @@ +/* + * Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "DepthwiseConvolutionLayer.h" + +#include <cker/operation/DepthwiseConv.h> + +namespace neurun +{ +namespace backend +{ +namespace cpu +{ +namespace kernel +{ + +DepthwiseConvolutionLayer::DepthwiseConvolutionLayer() + : _inputData(), _kernelData(), _outputData(), _biasData(), _inputShape(), _kernelShape(), + _outputShape(), _biasShape(), _paddingLeft(0), _paddingTop(0), _paddingRight(0), + _paddingBottom(0), _strideWidth(0), _strideHeight(0), _multiplier(0), + _activation(model::Activation::NONE), _inputType(OperandType::FLOAT32) +{ + // DO NOTHING +} + +void DepthwiseConvolutionLayer::convFloat32() +{ + float output_activation_min, output_activation_max; + CalculateActivationRangeFloat(_activation, &output_activation_min, &output_activation_max); + + nnfw::cker::DepthwiseConvParams op_params; + op_params.stride_width = _strideWidth; + op_params.stride_height = _strideHeight; + op_params.dilation_width_factor = 1; + op_params.dilation_height_factor = 1; + op_params.padding_values.width = _paddingLeft; + op_params.padding_values.height = _paddingTop; + op_params.depth_multiplier = _multiplier; + op_params.float_activation_min = output_activation_min; + op_params.float_activation_max = output_activation_max; + + nnfw::cker::DepthwiseConv(op_params, convertShapeToCkerShape(_inputShape), _inputData.f, + convertShapeToCkerShape(_kernelShape), _kernelData.f, + convertShapeToCkerShape(_biasShape), _biasData.f, + convertShapeToCkerShape(_outputShape), _outputData.f); +} + +void DepthwiseConvolutionLayer::convQuant8() +{ + int32_t output_activation_min = 0; + int32_t output_activation_max = 0; + CalculateActivationRangeUint8(_activation, _outputShape, &output_activation_min, + &output_activation_max); + + float real_multiplier = 0.0; + int32_t output_multiplier = 0; + int32_t output_shift = 0; + GetQuantizedConvolutionMultiplier(_inputShape, _kernelShape, _biasShape, _outputShape, + &real_multiplier); + QuantizeMultiplier(real_multiplier, &output_multiplier, &output_shift); + + nnfw::cker::DepthwiseConvParams op_params; + op_params.stride_width = _strideWidth; + op_params.stride_height = _strideHeight; + op_params.dilation_width_factor = 1; + op_params.dilation_height_factor = 1; + op_params.padding_values.width = _paddingLeft; + op_params.padding_values.height = _paddingTop; + op_params.depth_multiplier = _multiplier; + op_params.input_offset = -_inputShape.offset; + op_params.weights_offset = -_kernelShape.offset; + op_params.output_offset = _outputShape.offset; + op_params.output_multiplier = output_multiplier; + op_params.output_shift = output_shift; + op_params.quantized_activation_min = output_activation_min; + op_params.quantized_activation_max = output_activation_max; + + nnfw::cker::DepthwiseConv(op_params, convertShapeToCkerShape(_inputShape), _inputData.u8, + convertShapeToCkerShape(_kernelShape), _kernelData.u8, + convertShapeToCkerShape(_biasShape), _biasData.i32, + convertShapeToCkerShape(_outputShape), _outputData.u8); +} + +void DepthwiseConvolutionLayer::configure( + uint8_t *inputData, const Shape inputShape, uint8_t *kernelData, const Shape kernelShape, + uint8_t *biasData, const Shape biasShape, const uint32_t paddingLeft, + const uint32_t paddingRight, const uint32_t paddingTop, const uint32_t paddingBottom, + const uint32_t strideWidth, const uint32_t strideHeight, const uint32_t multiplier, + const model::Activation activation, uint8_t *outputData, const Shape outputShape) +{ + _inputData.u8 = inputData; + _inputShape = inputShape; + _inputType = inputShape.type; + _kernelData.u8 = kernelData; + _kernelShape = kernelShape; + _biasData.u8 = biasData; + _biasShape = biasShape; + _paddingLeft = paddingLeft; + _paddingRight = paddingRight; + _paddingTop = paddingTop; + _paddingBottom = paddingBottom; + _strideWidth = strideWidth; + _strideHeight = strideHeight; + _multiplier = multiplier; + _activation = activation; + _outputData.u8 = outputData; + _outputShape = outputShape; +} + +void DepthwiseConvolutionLayer::run() +{ + if (_inputType == OperandType::FLOAT32) + { + convFloat32(); + } + else if (_inputType == OperandType::QUANT8_ASYMM) + { + convQuant8(); + } +} + +} // namespace kernel +} // namespace cpu +} // namespace backend +} // namespace neurun |