1 //
2 // Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "NeonConstantWorkload.hpp"
7
8 #include <arm_compute/core/Types.h>
9 #include <BFloat16.hpp>
10 #include <Half.hpp>
11 #include <aclCommon/ArmComputeTensorUtils.hpp>
12 #include <armnn/utility/PolymorphicDowncast.hpp>
13 #include <neon/NeonTensorHandle.hpp>
14 #include <armnn/backends/TensorHandle.hpp>
15 #include "NeonBaseWorkload.hpp"
16
17 namespace armnn
18 {
19
NeonConstantWorkloadValidate(const TensorInfo & output)20 arm_compute::Status NeonConstantWorkloadValidate(const TensorInfo& output)
21 {
22 const arm_compute::TensorInfo neonOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
23
24 std::array<arm_compute::DataType,9> supportedTypes = {
25 arm_compute::DataType::BFLOAT16,
26 arm_compute::DataType::F16,
27 arm_compute::DataType::F32,
28 arm_compute::DataType::QASYMM8,
29 arm_compute::DataType::QASYMM8_SIGNED,
30 arm_compute::DataType::QSYMM16,
31 arm_compute::DataType::QSYMM8,
32 arm_compute::DataType::QSYMM8_PER_CHANNEL,
33 arm_compute::DataType::S32
34 };
35 auto it = std::find(begin(supportedTypes), end(supportedTypes), neonOutputInfo.data_type());
36
37 if (it != end(supportedTypes))
38 {
39 return arm_compute::Status{};
40 }
41 else
42 {
43 return arm_compute::Status{arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported DataType"};
44 }
45 }
46
NeonConstantWorkload(const ConstantQueueDescriptor & descriptor,const WorkloadInfo & info)47 NeonConstantWorkload::NeonConstantWorkload(const ConstantQueueDescriptor& descriptor,
48 const WorkloadInfo& info)
49 : NeonBaseWorkload<ConstantQueueDescriptor>(descriptor, info)
50 , m_RanOnce(false)
51 {
52 }
53
Execute() const54 void NeonConstantWorkload::Execute() const
55 {
56 ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonConstantWorkload_Execute", this->GetGuid());
57
58 using namespace armcomputetensorutils;
59
60 // The intermediate tensor held by the corresponding layer output handler can be initialised with the
61 // given data on the first inference, then reused for subsequent inferences.
62 // The initialisation cannot happen at workload construction time since the ACL kernel for the next layer
63 // may not have been configured at the time.
64 if (!m_RanOnce)
65 {
66 const ConstantQueueDescriptor& data = this->m_Data;
67
68 ARMNN_ASSERT(data.m_LayerOutput != nullptr);
69 arm_compute::ITensor& output =
70 PolymorphicDowncast<NeonTensorHandle*>(data.m_Outputs[0])->GetTensor();
71 arm_compute::DataType computeDataType =
72 PolymorphicDowncast<NeonTensorHandle*>(data.m_Outputs[0])->GetDataType();
73
74 switch (computeDataType)
75 {
76 case arm_compute::DataType::BFLOAT16:
77 {
78 CopyArmComputeITensorData(data.m_LayerOutput->GetConstTensor<BFloat16>(), output);
79 break;
80 }
81 case arm_compute::DataType::F16:
82 {
83 CopyArmComputeITensorData(data.m_LayerOutput->GetConstTensor<Half>(), output);
84 break;
85 }
86 case arm_compute::DataType::F32:
87 {
88 CopyArmComputeITensorData(data.m_LayerOutput->GetConstTensor<float>(), output);
89 break;
90 }
91 case arm_compute::DataType::QASYMM8:
92 {
93 CopyArmComputeITensorData(data.m_LayerOutput->GetConstTensor<uint8_t>(), output);
94 break;
95 }
96 case arm_compute::DataType::QASYMM8_SIGNED:
97 {
98 CopyArmComputeITensorData(data.m_LayerOutput->GetConstTensor<int8_t>(), output);
99 break;
100 }
101 case arm_compute::DataType::QSYMM16:
102 {
103 CopyArmComputeITensorData(data.m_LayerOutput->GetConstTensor<int16_t>(), output);
104 break;
105 }
106 case arm_compute::DataType::QSYMM8:
107 case arm_compute::DataType::QSYMM8_PER_CHANNEL:
108 {
109 CopyArmComputeITensorData(data.m_LayerOutput->GetConstTensor<int8_t>(), output);
110 break;
111 }
112 case arm_compute::DataType::S32:
113 {
114 CopyArmComputeITensorData(data.m_LayerOutput->GetConstTensor<int32_t>(), output);
115 break;
116 }
117 default:
118 {
119 ARMNN_ASSERT_MSG(false, "Unknown data type");
120 break;
121 }
122 }
123
124 m_RanOnce = true;
125 }
126 }
127
128 } //namespace armnn
129