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;
63 "ArmNN NeonDepthwiseConvolutionWorkload has empty bias value."};
69 "ArmNN NeonDepthwiseConvolutionWorkload does not support non constant bias."};
71 aclBiasesInfo = BuildArmComputeTensorInfo(biases.
value(), descriptor.
m_DataLayout);
73 optionalAclBiasesInfo = &aclBiasesInfo;
76 const arm_compute::PadStrideInfo aclPadStrideInfo = BuildArmComputePadStrideInfo(descriptor);
77 const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(descriptor.
m_DilationX,
81 activationDescriptor);
83 return arm_compute::NEDepthwiseConvolutionLayer::validate(&aclInputInfo,
85 optionalAclBiasesInfo,
98 arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(
m_Data.
m_Inputs[0])->GetTensor();
99 arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(
m_Data.
m_Outputs[0])->GetTensor();
100 arm_compute::ITensor& weights = PolymorphicDowncast<IAclTensorHandle*>(
m_Data.
m_Inputs[1])->GetTensor();
101 weights.info()->set_are_values_constant(
info.m_InputTensorInfos[1].IsConstant());
102 arm_compute::ITensor* biasesPtr =
nullptr;
105 biasesPtr = &PolymorphicDowncast<IAclTensorHandle *>(
m_Data.
m_Inputs[2])->GetTensor();
106 biasesPtr->info()->set_are_values_constant(
info.m_InputTensorInfos[2].IsConstant());
111 arm_compute::TensorShape weightsShape = weights.info()->tensor_shape();
112 arm_compute::TensorShape inputShape = input.info()->tensor_shape();
116 unsigned int depthMultiplier =
126 input.info()->set_data_layout(aclDataLayout);
127 weights.info()->set_data_layout(aclDataLayout);
128 output.info()->set_data_layout(aclDataLayout);
130 const arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(
m_Data.
m_Parameters);
134 m_pDepthwiseConvolutionLayer = std::make_unique<arm_compute::NEDepthwiseConvolutionLayer>();
135 static_cast<arm_compute::NEDepthwiseConvolutionLayer*
>(
136 m_pDepthwiseConvolutionLayer.get())->configure(&input,
164 m_pDepthwiseConvolutionLayer->prepare();
172 m_pDepthwiseConvolutionLayer->run();