1 //
2 // Copyright © 2020 Arm Ltd. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "NeonSoftmaxWorkload.hpp"
7 #include "NeonWorkloadUtils.hpp"
8
9 #include <armnn/utility/PolymorphicDowncast.hpp>
10
11 #include <aclCommon/ArmComputeUtils.hpp>
12 #include <aclCommon/ArmComputeTensorUtils.hpp>
13
14 #include <arm_compute/runtime/NEON/functions/NESoftmaxLayer.h>
15
16 namespace armnn
17 {
18
NeonSoftmaxWorkloadValidate(const TensorInfo & input,const TensorInfo & output,const SoftmaxDescriptor & descriptor)19 arm_compute::Status NeonSoftmaxWorkloadValidate(const TensorInfo& input,
20 const TensorInfo& output,
21 const SoftmaxDescriptor& descriptor)
22 {
23 const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input);
24 const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
25
26 int aclAxis = ComputeAclAxis(descriptor.m_Axis, input);
27 return arm_compute::NESoftmaxLayer::validate(&aclInputInfo,
28 &aclOutputInfo,
29 descriptor.m_Beta,
30 aclAxis);
31 }
32
NeonSoftmaxWorkload(const SoftmaxQueueDescriptor & descriptor,const WorkloadInfo & info,std::shared_ptr<arm_compute::MemoryManagerOnDemand> & memoryManager)33 NeonSoftmaxWorkload::NeonSoftmaxWorkload(const SoftmaxQueueDescriptor& descriptor,
34 const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
35 : BaseWorkload<SoftmaxQueueDescriptor>(descriptor, info)
36 {
37 m_Data.ValidateInputsOutputs("NeonSoftmaxWorkload", 1, 1);
38
39 // The ArmCompute softmax layer uses 2D input/output tensors, so flatten the first three dimensions.
40 arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
41 arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
42
43 auto layer = std::make_unique<arm_compute::NESoftmaxLayer>(memoryManager);
44 int aclAxis = ComputeAclAxis(m_Data.m_Parameters.m_Axis, info.m_InputTensorInfos[0]);
45 layer->configure(&input, &output, m_Data.m_Parameters.m_Beta, aclAxis);
46 m_SoftmaxLayer.reset(layer.release());
47 }
48
Execute() const49 void NeonSoftmaxWorkload::Execute() const
50 {
51 ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonSoftmaxWorkload_Execute");
52 m_SoftmaxLayer->run();
53 }
54
55 } //namespace armnn
56
57