1 //
2 // Copyright © 2020 Arm Ltd and Contributors. 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 : NeonBaseWorkload<SoftmaxQueueDescriptor>(descriptor, info)
36 {
37 // Report Profiling Details
38 ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonSoftmaxWorkload_Construct",
39 descriptor.m_Parameters,
40 info,
41 this->GetGuid());
42
43 m_Data.ValidateInputsOutputs("NeonSoftmaxWorkload", 1, 1);
44
45 // The ArmCompute softmax layer uses 2D input/output tensors, so flatten the first three dimensions.
46 arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
47 arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
48
49 auto layer = std::make_unique<arm_compute::NESoftmaxLayer>(memoryManager);
50 int aclAxis = ComputeAclAxis(m_Data.m_Parameters.m_Axis, info.m_InputTensorInfos[0]);
51 layer->configure(&input, &output, m_Data.m_Parameters.m_Beta, aclAxis);
52 m_SoftmaxLayer.reset(layer.release());
53 }
54
Execute() const55 void NeonSoftmaxWorkload::Execute() const
56 {
57 ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonSoftmaxWorkload_Execute", this->GetGuid());
58 m_SoftmaxLayer->run();
59 }
60
61 } //namespace armnn
62
63