1 //
2 // Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "NeonMultiplicationWorkload.hpp"
7
8 #include "NeonWorkloadUtils.hpp"
9
10 #include <aclCommon/ArmComputeUtils.hpp>
11
12 #include <armnn/utility/PolymorphicDowncast.hpp>
13
14 #include <arm_compute/runtime/NEON/functions/NEPixelWiseMultiplication.h>
15
16 namespace armnn
17 {
18
NeonMultiplicationWorkloadValidate(const TensorInfo & input0,const TensorInfo & input1,const TensorInfo & output,const ActivationDescriptor * activationDescriptor)19 arm_compute::Status NeonMultiplicationWorkloadValidate(const TensorInfo& input0,
20 const TensorInfo& input1,
21 const TensorInfo& output,
22 const ActivationDescriptor* activationDescriptor)
23 {
24 const arm_compute::TensorInfo aclInput1 = armcomputetensorutils::BuildArmComputeTensorInfo(input0);
25 const arm_compute::TensorInfo aclInput2 = armcomputetensorutils::BuildArmComputeTensorInfo(input1);
26 const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output);
27
28 auto convertPolicy = (IsQuantizedType(input0.GetDataType()) || IsQuantizedType(input1.GetDataType())) ?
29 arm_compute::ConvertPolicy::SATURATE :
30 arm_compute::ConvertPolicy::WRAP;
31
32 const arm_compute::ActivationLayerInfo activationInfo = ConvertActivationDescriptorToAclActivationLayerInfo(
33 activationDescriptor);
34
35 // At the time of writing, configure() will fail if a rounding policy other than TO_ZERO is supplied to it,
36 // when providing a scale of 1.0 for F32 tensors, even though the provided rounding policy appears to be
37 // ignored for F32 tensors.
38 return arm_compute::NEPixelWiseMultiplication::validate(&aclInput1,
39 &aclInput2,
40 &aclOutput,
41 1.0f,
42 convertPolicy,
43 arm_compute::RoundingPolicy::TO_ZERO,
44 activationInfo);
45 }
46
NeonMultiplicationWorkload(const MultiplicationQueueDescriptor & descriptor,const WorkloadInfo & info)47 NeonMultiplicationWorkload::NeonMultiplicationWorkload(const MultiplicationQueueDescriptor& descriptor,
48 const WorkloadInfo& info)
49 : NeonBaseWorkload<MultiplicationQueueDescriptor>(descriptor, info)
50 {
51 m_Data.ValidateInputsOutputs("NeonMultiplicationWorkload", 2, 1);
52
53 arm_compute::ITensor& input1 = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
54 arm_compute::ITensor& input2 = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
55 arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
56
57 auto convertPolicy = (IsQuantizedType(info.m_InputTensorInfos[0].GetDataType()) ||
58 IsQuantizedType(info.m_InputTensorInfos[1].GetDataType())) ?
59 arm_compute::ConvertPolicy::SATURATE :
60 arm_compute::ConvertPolicy::WRAP;
61
62 const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
63
64 // At the time of writing, configure() will fail if a rounding policy other than TO_ZERO is supplied to it,
65 // when providing a scale of 1.0 for F32 tensors, even though the provided rounding policy appears to be
66 // ignored for F32 tensors.
67 auto layer = std::make_unique<arm_compute::NEPixelWiseMultiplication>();
68 layer->configure(&input1,
69 &input2,
70 &output,
71 1.0f,
72 convertPolicy,
73 arm_compute::RoundingPolicy::TO_ZERO,
74 activationInfo);
75 m_PixelWiseMultiplication.reset(layer.release());
76 }
77
Execute() const78 void NeonMultiplicationWorkload::Execute() const
79 {
80 ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonMultiplicationWorkload_Execute", this->GetGuid());
81 m_PixelWiseMultiplication->run();
82 }
83
84 } //namespace armnn
85