12 #include <arm_compute/runtime/NEON/functions/NEConvolutionLayer.h> 20 using namespace armcomputetensorutils;
28 const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.
m_DataLayout);
29 const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.
m_DataLayout);
30 const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.
m_DataLayout);
32 const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(descriptor.
m_DilationX,
35 arm_compute::TensorInfo aclBiasesInfo;
36 arm_compute::TensorInfo *optionalAclBiasesInfo =
nullptr;
42 aclBiasesInfo = BuildArmComputeTensorInfo(biases.
value(), descriptor.
m_DataLayout);
43 optionalAclBiasesInfo = &aclBiasesInfo;
46 arm_compute::PadStrideInfo layerInfo = BuildArmComputePadStrideInfo(descriptor);
48 return arm_compute::NEConvolutionLayer::validate(&aclInputInfo,
50 optionalAclBiasesInfo,
53 arm_compute::WeightsInfo(),
59 std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
62 using arm_compute::NEDirectConvolutionLayer;
68 arm_compute::ITensor& input = boost::polymorphic_downcast<IAclTensorHandle*>(
m_Data.
m_Inputs[0])->GetTensor();
69 arm_compute::ITensor& output = boost::polymorphic_downcast<IAclTensorHandle*>(
m_Data.
m_Outputs[0])->GetTensor();
72 input.info()->set_data_layout(aclDataLayout);
73 output.info()->set_data_layout(aclDataLayout);
75 m_KernelTensor = std::make_unique<arm_compute::Tensor>();
80 m_BiasTensor = std::make_unique<arm_compute::Tensor>();
84 arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(
m_Data.
m_Parameters);
89 auto convolutionLayer = std::make_unique<arm_compute::NEConvolutionLayer>(memoryManager);
90 convolutionLayer->configure(&input,
95 arm_compute::WeightsInfo(),
98 m_ConvolutionLayer.reset(convolutionLayer.release());
100 BOOST_ASSERT(m_ConvolutionLayer);
109 m_ConvolutionLayer->prepare();
116 m_ConvolutionLayer->run();
119 void NeonConvolution2dWorkload::FreeUnusedTensors()
121 FreeTensorIfUnused(m_KernelTensor);
122 FreeTensorIfUnused(m_BiasTensor);
DataLayout m_DataLayout
The data layout to be used (NCHW, NHWC).
arm_compute::Status NeonConvolution2dWorkloadValidate(const TensorInfo &input, const TensorInfo &output, const Convolution2dDescriptor &descriptor, const TensorInfo &weights, const Optional< TensorInfo > &biases)
const TensorInfo & GetTensorInfo() const
#define ARMNN_SCOPED_PROFILING_EVENT_NEON(name)
LayerDescriptor m_Parameters
const ConstCpuTensorHandle * m_Weight
NeonConvolution2dWorkload(const Convolution2dQueueDescriptor &descriptor, const WorkloadInfo &info, std::shared_ptr< arm_compute::MemoryManagerOnDemand > &memoryManager)
const Convolution2dQueueDescriptor m_Data
void Execute() const override
const ConstCpuTensorHandle * m_Bias
bool m_BiasEnabled
Enable/disable bias.
void ValidateInputsOutputs(const std::string &descName, unsigned int numExpectedIn, unsigned int numExpectedOut) const
uint32_t m_DilationY
Dilation along y axis.
std::vector< ITensorHandle * > m_Outputs
std::vector< ITensorHandle * > m_Inputs
A Convolution2dDescriptor for the Convolution2dLayer.
uint32_t m_DilationX
Dilation along x axis.
bool has_value() const noexcept
void InitializeArmComputeTensorData(arm_compute::Tensor &tensor, const ConstCpuTensorHandle *handle)