14 #include <arm_compute/runtime/NEON/functions/NEConvolutionLayer.h> 22 using namespace armcomputetensorutils;
29 bool isFastMathEnabled,
32 const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.
m_DataLayout);
33 const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.
m_DataLayout);
34 const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.
m_DataLayout);
36 const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(descriptor.
m_DilationX,
39 arm_compute::TensorInfo aclBiasesInfo;
40 arm_compute::TensorInfo *optionalAclBiasesInfo =
nullptr;
46 aclBiasesInfo = BuildArmComputeTensorInfo(biases.
value(), descriptor.
m_DataLayout);
47 optionalAclBiasesInfo = &aclBiasesInfo;
50 arm_compute::PadStrideInfo layerInfo = BuildArmComputePadStrideInfo(descriptor);
53 activationDescriptor);
55 return arm_compute::NEConvolutionLayer::validate(&aclInputInfo,
57 optionalAclBiasesInfo,
60 arm_compute::WeightsInfo(),
69 std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager,
70 const bool isFastMathEnabled)
73 using arm_compute::NEConvolutionLayer;
79 arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(
m_Data.
m_Inputs[0])->GetTensor();
80 arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(
m_Data.
m_Outputs[0])->GetTensor();
83 input.info()->set_data_layout(aclDataLayout);
84 output.info()->set_data_layout(aclDataLayout);
86 m_KernelTensor = std::make_unique<arm_compute::Tensor>();
91 m_BiasTensor = std::make_unique<arm_compute::Tensor>();
95 arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(
m_Data.
m_Parameters);
102 auto convolutionLayer = std::make_unique<arm_compute::NEConvolutionLayer>(memoryManager);
103 convolutionLayer->configure(&input,
104 m_KernelTensor.get(),
108 arm_compute::WeightsInfo(),
113 m_ConvolutionMethod =
114 convolutionLayer->get_convolution_method(input.info(),
115 m_KernelTensor->info(),
118 arm_compute::WeightsInfo(),
123 m_ConvolutionLayer.reset(convolutionLayer.release());
134 m_ConvolutionLayer->prepare();
141 m_ConvolutionLayer->run();
146 return m_ConvolutionMethod;
149 void NeonConvolution2dWorkload::FreeUnusedTensors()
151 FreeTensorIfUnused(m_KernelTensor);
152 FreeTensorIfUnused(m_BiasTensor);
bool m_BiasEnabled
Enable/disable bias.
const ConstCpuTensorHandle * m_Bias
DataLayout m_DataLayout
The data layout to be used (NCHW, NHWC).
void Execute() const override
NeonConvolution2dWorkload(const Convolution2dQueueDescriptor &descriptor, const WorkloadInfo &info, std::shared_ptr< arm_compute::MemoryManagerOnDemand > &memoryManager, const bool isFastMathENabled=false)
A Convolution2dDescriptor for the Convolution2dLayer.
const Convolution2dQueueDescriptor m_Data
#define ARMNN_SCOPED_PROFILING_EVENT_NEON(name)
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.
uint32_t m_DilationY
Dilation along y axis.
LayerDescriptor m_Parameters
const ConstCpuTensorHandle * m_Weight
bool has_value() const noexcept
#define ARMNN_ASSERT(COND)
arm_compute::Status NeonConvolution2dWorkloadValidate(const TensorInfo &input, const TensorInfo &output, const Convolution2dDescriptor &descriptor, const TensorInfo &weights, const Optional< TensorInfo > &biases, bool isFastMathEnabled, const ActivationDescriptor *activationDescriptor)
An ActivationDescriptor for the ActivationLayer.
arm_compute::ConvolutionMethod GetConvolutionMethod() const
void InitializeArmComputeTensorData(arm_compute::Tensor &tensor, const ConstCpuTensorHandle *handle)
uint32_t m_DilationX
Dilation along x axis.
std::vector< ITensorHandle * > m_Outputs
Contains information about inputs and outputs to a layer.
std::vector< ITensorHandle * > m_Inputs
const TensorInfo & GetTensorInfo() const
arm_compute::ActivationLayerInfo ConvertActivationDescriptorToAclActivationLayerInfo(const ActivationDescriptor &actDesc)