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);
41 unsigned int aclDepthMultiplier;
46 const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weightsPermuted, descriptor.
m_DataLayout);
48 arm_compute::TensorInfo aclBiasesInfo;
49 arm_compute::TensorInfo* optionalAclBiasesInfo =
nullptr;
55 aclBiasesInfo = BuildArmComputeTensorInfo(biases.
value(), descriptor.
m_DataLayout);
56 optionalAclBiasesInfo = &aclBiasesInfo;
59 arm_compute::PadStrideInfo aclPadStrideInfo = BuildArmComputePadStrideInfo(descriptor);
60 const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(
64 activationDescriptor);
66 return arm_compute::NEDepthwiseConvolutionLayer::validate(&aclInputInfo,
68 optionalAclBiasesInfo,
85 unsigned int depthMultiplier;
86 std::unique_ptr<unsigned char[]> permuteBuffer(
new unsigned char[weightInfo.GetNumBytes()]);
93 m_KernelTensor = std::make_unique<arm_compute::Tensor>();
98 m_BiasTensor = std::make_unique<arm_compute::Tensor>();
102 const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(
110 arm_compute::ITensor& input = inputTensorHandle->
GetTensor();
111 arm_compute::ITensor& output = outputTensorHandle->
GetTensor();
114 input.info()->set_data_layout(aclDataLayout);
115 output.info()->set_data_layout(aclDataLayout);
117 arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(
m_Data.
m_Parameters);
121 m_pDepthwiseConvolutionLayer = std::make_unique<arm_compute::NEDepthwiseConvolutionLayer>();
122 static_cast<arm_compute::NEDepthwiseConvolutionLayer*
>(
123 m_pDepthwiseConvolutionLayer.get())->configure(&input,
124 m_KernelTensor.get(),
159 m_pDepthwiseConvolutionLayer->prepare();
168 m_pDepthwiseConvolutionLayer->run();
171 void NeonDepthwiseConvolutionWorkload::FreeUnusedTensors()
173 FreeTensorIfUnused(m_KernelTensor);
174 FreeTensorIfUnused(m_BiasTensor);
virtual arm_compute::ITensor & GetTensor()=0
bool m_BiasEnabled
Enable/disable bias.
DataLayout m_DataLayout
The data layout to be used (NCHW, NHWC).
arm_compute::Status NeonDepthwiseConvolutionWorkloadValidate(const TensorInfo &input, const TensorInfo &output, const DepthwiseConvolution2dDescriptor &descriptor, const TensorInfo &weights, const Optional< TensorInfo > &biases, const ActivationDescriptor *activationDescriptor)
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 TensorInfo & GetTensorInfo() const
std::vector< TensorInfo > m_InputTensorInfos
std::tuple< ConstTensor, unsigned int > Convert1HWOTensorToAcl(const ConstTensorHandle *weightTensor, const TensorInfo &inputInfo, const DataLayout dataLayout, void *permuteBuffer)
Weights for depthwise have a datalayout of [1,H,W,O] = [1,H,W,I*M] This function coverts a ConstCpuTe...
uint32_t m_DilationX
Dilation factor value for width dimension.
DepthwiseConvolution2dQueueDescriptor m_Data
bool has_value() const noexcept
A tensor defined by a TensorInfo (shape and data type) and an immutable backing store.
const ConstTensorHandle * m_Bias
#define ARMNN_ASSERT(COND)
std::tuple< TensorInfo, unsigned int > Convert1HWOTensorInfoToAcl(const TensorInfo &weightInfo, const TensorInfo &inputInfo, const DataLayout dataLayout)
Weights for depthwise have a datalayout of [1,H,W,O] = [1,H,W,I*M] This function coverts a TensorInfo...
std::vector< TensorInfo > m_OutputTensorInfos
An ActivationDescriptor for the ActivationLayer.
const TensorInfo & GetInfo() const
const ConstTensorHandle * m_Weight
profiling::ProfilingGuid GetGuid() const final
virtual void Execute() const override
std::vector< ITensorHandle * > m_Outputs
#define ARMNN_REPORT_PROFILING_WORKLOAD_DESC(name, desc, infos, guid)
void InitializeArmComputeTensorData(arm_compute::Tensor &tensor, const ConstTensorHandle *handle)
Contains information about TensorInfos of a layer.
std::vector< ITensorHandle * > m_Inputs
NeonDepthwiseConvolutionWorkload(const DepthwiseConvolution2dQueueDescriptor &descriptor, const WorkloadInfo &info)
#define ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID(name, guid)
A DepthwiseConvolution2dDescriptor for the DepthwiseConvolution2dLayer.
Depthwise Convolution 2D layer workload data.
arm_compute::ActivationLayerInfo ConvertActivationDescriptorToAclActivationLayerInfo(const ActivationDescriptor &actDesc)