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);
48 unsigned int aclDepthMultiplier;
53 arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weightsPermuted, descriptor.
m_DataLayout);
54 aclWeightsInfo.set_are_values_constant(weights.
IsConstant());
56 arm_compute::TensorInfo aclBiasesInfo;
57 arm_compute::TensorInfo* optionalAclBiasesInfo =
nullptr;
65 "ArmNN NeonDepthwiseConv2dWorkload does not support non constant bias."};
67 aclBiasesInfo = BuildArmComputeTensorInfo(biases.
value(), descriptor.
m_DataLayout);
69 optionalAclBiasesInfo = &aclBiasesInfo;
72 arm_compute::PadStrideInfo aclPadStrideInfo = BuildArmComputePadStrideInfo(descriptor);
73 const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(
77 activationDescriptor);
79 return arm_compute::NEDepthwiseConvolutionLayer::validate(&aclInputInfo,
81 optionalAclBiasesInfo,
94 arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(
m_Data.
m_Inputs[0])->GetTensor();
95 arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(
m_Data.
m_Outputs[0])->GetTensor();
96 arm_compute::ITensor& weights = PolymorphicDowncast<IAclTensorHandle*>(
m_Data.
m_Inputs[1])->GetTensor();
97 arm_compute::ITensor* biasesPtr =
nullptr;
100 biasesPtr = &PolymorphicDowncast<IAclTensorHandle *>(
m_Data.
m_Inputs[2])->GetTensor();
103 arm_compute::ITensorInfo* weightsInfo = weights.info();
104 arm_compute::ITensorInfo* inputInfo = input.info();
105 auto weightsShape = weightsInfo->tensor_shape();
106 auto inputShape = inputInfo->tensor_shape();
110 unsigned int depthMultiplier =
113 const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(
120 input.info()->set_data_layout(aclDataLayout);
121 weights.info()->set_data_layout(aclDataLayout);
122 output.info()->set_data_layout(aclDataLayout);
124 arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(
m_Data.
m_Parameters);
128 m_pDepthwiseConvolutionLayer = std::make_unique<arm_compute::NEDepthwiseConvolutionLayer>();
129 static_cast<arm_compute::NEDepthwiseConvolutionLayer*
>(
130 m_pDepthwiseConvolutionLayer.get())->configure(&input,
158 m_pDepthwiseConvolutionLayer->prepare();
166 m_pDepthwiseConvolutionLayer->run();
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)
arm::pipe::ProfilingGuid GetGuid() const final
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
uint32_t m_DilationX
Dilation factor value for width dimension.
DepthwiseConvolution2dQueueDescriptor m_Data
bool has_value() const noexcept
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 ConstTensorHandle * m_Weight
virtual void Execute() const override
Optional< TensorInfo > m_BiasTensorInfo
std::vector< ITensorHandle * > m_Outputs
#define ARMNN_REPORT_PROFILING_WORKLOAD_DESC(name, desc, infos, guid)
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)
Optional< TensorInfo > m_WeightsTensorInfo
A DepthwiseConvolution2dDescriptor for the DepthwiseConvolution2dLayer.
Depthwise Convolution 2D layer workload data.
arm_compute::ActivationLayerInfo ConvertActivationDescriptorToAclActivationLayerInfo(const ActivationDescriptor &actDesc)