20 #include <arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h> 27 using namespace armcomputetensorutils;
36 const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.
m_DataLayout);
37 const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.
m_DataLayout);
40 const unsigned int aclDepthMultiplier = weights.
GetShape()[0];
47 const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weightsPermuted, descriptor.
m_DataLayout);
49 arm_compute::TensorInfo aclBiasesInfo;
50 arm_compute::TensorInfo *optionalAclBiasesInfo =
nullptr;
56 aclBiasesInfo = BuildArmComputeTensorInfo(biases.
value(), descriptor.
m_DataLayout);
57 optionalAclBiasesInfo = &aclBiasesInfo;
60 arm_compute::PadStrideInfo aclPadStrideInfo = BuildArmComputePadStrideInfo(descriptor);
61 const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(
65 activationDescriptor);
67 return arm_compute::NEDepthwiseConvolutionLayer::validate(&aclInputInfo,
69 optionalAclBiasesInfo,
95 m_KernelTensor = std::make_unique<arm_compute::Tensor>();
100 m_BiasTensor = std::make_unique<arm_compute::Tensor>();
104 const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(
112 arm_compute::ITensor& input = inputTensorHandle->
GetTensor();
113 arm_compute::ITensor& output = outputTensorHandle->
GetTensor();
116 input.info()->set_data_layout(aclDataLayout);
117 output.info()->set_data_layout(aclDataLayout);
120 const unsigned int depthMultiplier = weightInfo.GetShape()[0];
122 arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(
m_Data.
m_Parameters);
126 m_pDepthwiseConvolutionLayer = std::make_unique<arm_compute::NEDepthwiseConvolutionLayer>();
127 static_cast<arm_compute::NEDepthwiseConvolutionLayer*
>(
128 m_pDepthwiseConvolutionLayer.get())->configure(&input,
129 m_KernelTensor.get(),
147 m_pDepthwiseConvolutionLayer->prepare();
156 m_pDepthwiseConvolutionLayer->run();
159 void NeonDepthwiseConvolutionWorkload::FreeUnusedTensors()
161 FreeTensorIfUnused(m_KernelTensor);
162 FreeTensorIfUnused(m_BiasTensor);
virtual arm_compute::ITensor & GetTensor()=0
bool m_BiasEnabled
Enable/disable bias.
const TensorShape & GetShape() const
armnn::ConstTensor ConvertWeightTensorFromArmnnToAcl(const ConstCpuTensorHandle *weightTensor, DataLayout dataLayout, void *permuteBuffer)
TensorInfo ConvertWeightTensorInfoFromArmnnToAcl(const TensorInfo &weightInfo, DataLayout dataLayout)
DataLayout m_DataLayout
The data layout to be used (NCHW, NHWC).
const ConstCpuTensorHandle * m_Bias
arm_compute::Status NeonDepthwiseConvolutionWorkloadValidate(const TensorInfo &input, const TensorInfo &output, const DepthwiseConvolution2dDescriptor &descriptor, const TensorInfo &weights, const Optional< TensorInfo > &biases, const ActivationDescriptor *activationDescriptor)
const DepthwiseConvolution2dQueueDescriptor m_Data
#define ARMNN_SCOPED_PROFILING_EVENT_NEON(name)
unsigned int GetNumBytes() const
arm_compute::ActivationLayerInfo ConvertAdditionalInfoToAclActivationLayerInfo(const QueueDescriptor &queueDescriptor)
void ValidateInputsOutputs(const std::string &descName, unsigned int numExpectedIn, unsigned int numExpectedOut) const
Copyright (c) 2021 ARM Limited and Contributors.
LayerDescriptor m_Parameters
uint32_t m_DilationY
Dilation factor value for height dimension.
const ConstCpuTensorHandle * m_Weight
uint32_t m_DilationX
Dilation factor value for width dimension.
bool has_value() const noexcept
A tensor defined by a TensorInfo (shape and data type) and an immutable backing store.
#define ARMNN_ASSERT(COND)
An ActivationDescriptor for the ActivationLayer.
virtual void Execute() const override
void InitializeArmComputeTensorData(arm_compute::Tensor &tensor, const ConstCpuTensorHandle *handle)
std::vector< ITensorHandle * > m_Outputs
Contains information about inputs and outputs to a layer.
std::vector< ITensorHandle * > m_Inputs
NeonDepthwiseConvolutionWorkload(const DepthwiseConvolution2dQueueDescriptor &descriptor, const WorkloadInfo &info)
const TensorInfo & GetTensorInfo() const
A DepthwiseConvolution2dDescriptor for the DepthwiseConvolution2dLayer.
arm_compute::ActivationLayerInfo ConvertActivationDescriptorToAclActivationLayerInfo(const ActivationDescriptor &actDesc)