17 #include <arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h> 22 using namespace armcomputetensorutils;
30 const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.
m_DataLayout);
31 const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.
m_DataLayout);
34 const unsigned int aclDepthMultiplier = weights.
GetShape()[0];
41 const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weightsPermuted, descriptor.
m_DataLayout);
43 arm_compute::TensorInfo aclBiasesInfo;
44 arm_compute::TensorInfo *optionalAclBiasesInfo =
nullptr;
50 aclBiasesInfo = BuildArmComputeTensorInfo(biases.
value(), descriptor.
m_DataLayout);
51 optionalAclBiasesInfo = &aclBiasesInfo;
54 const arm_compute::PadStrideInfo aclPadStrideInfo = BuildArmComputePadStrideInfo(descriptor);
55 const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(
59 return arm_compute::CLDepthwiseConvolutionLayer::validate(&aclInputInfo,
61 optionalAclBiasesInfo,
65 arm_compute::ActivationLayerInfo(),
90 m_BiasTensor = std::make_unique<arm_compute::CLTensor>();
94 const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(
99 std::string name = std::string(
"ClDepthwiseConvolutionWorkload");
106 input.info()->set_data_layout(aclDataLayout);
107 output.info()->set_data_layout(aclDataLayout);
113 const unsigned int depthMultiplier = weightInfo.
GetShape()[0];
115 arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(
m_Data.
m_Parameters);
120 m_KernelTensor.get(),
125 arm_compute::ActivationLayerInfo(),
bool m_BiasEnabled
Enable/disable bias.
const TensorShape & GetShape() const
armnn::ConstTensor ConvertWeightTensorFromArmnnToAcl(const ConstCpuTensorHandle *weightTensor, DataLayout dataLayout, void *permuteBuffer)
void InitializeArmComputeClTensorData(arm_compute::CLTensor &clTensor, const ConstCpuTensorHandle *handle)
TensorInfo ConvertWeightTensorInfoFromArmnnToAcl(const TensorInfo &weightInfo, DataLayout dataLayout)
DataLayout m_DataLayout
The data layout to be used (NCHW, NHWC).
#define ARMNN_SCOPED_PROFILING_EVENT_CL(name)
void RunClFunction(arm_compute::IFunction &function, const CheckLocation &location)
const ConstCpuTensorHandle * m_Bias
const DepthwiseConvolution2dQueueDescriptor m_Data
unsigned int GetNumBytes() const
std::unique_ptr< arm_compute::IFunction > m_DepthwiseConvolutionLayer
void ValidateInputsOutputs(const std::string &descName, unsigned int numExpectedIn, unsigned int numExpectedOut) const
std::unique_ptr< arm_compute::CLTensor > m_BiasTensor
Copyright (c) 2020 ARM Limited.
ClDepthwiseConvolutionWorkload(const DepthwiseConvolution2dQueueDescriptor &descriptor, const WorkloadInfo &info)
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.
void Execute() const override
bool has_value() const noexcept
A tensor defined by a TensorInfo (shape and data type) and an immutable backing store.
std::unique_ptr< arm_compute::CLTensor > m_KernelTensor
std::vector< ITensorHandle * > m_Outputs
arm_compute::Status ClDepthwiseConvolutionWorkloadValidate(const TensorInfo &input, const TensorInfo &output, const DepthwiseConvolution2dDescriptor &descriptor, const TensorInfo &weights, const Optional< TensorInfo > &biases)
Contains information about inputs and outputs to a layer.
std::vector< ITensorHandle * > m_Inputs
const TensorInfo & GetTensorInfo() const
A DepthwiseConvolution2dDescriptor for the DepthwiseConvolution2dLayer.