1 //
2 // Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "ClMultiplicationWorkload.hpp"
7
8 #include <aclCommon/ArmComputeUtils.hpp>
9 #include <armnn/backends/TensorHandle.hpp>
10
11 #include <cl/ClTensorHandle.hpp>
12
13 #include "ClWorkloadUtils.hpp"
14
15 namespace armnn
16 {
17
ClMultiplicationWorkloadValidate(const TensorInfo & input0,const TensorInfo & input1,const TensorInfo & output,const ActivationDescriptor * activationDescriptor)18 arm_compute::Status ClMultiplicationWorkloadValidate(const TensorInfo& input0,
19 const TensorInfo& input1,
20 const TensorInfo& output,
21 const ActivationDescriptor* activationDescriptor)
22 {
23 const arm_compute::TensorInfo aclInput1 = armcomputetensorutils::BuildArmComputeTensorInfo(input0);
24 const arm_compute::TensorInfo aclInput2 = armcomputetensorutils::BuildArmComputeTensorInfo(input1);
25 const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output);
26
27 auto convertPolicy = (IsQuantizedType(input0.GetDataType()) || IsQuantizedType(input1.GetDataType())) ?
28 arm_compute::ConvertPolicy::SATURATE :
29 arm_compute::ConvertPolicy::WRAP;
30
31 const arm_compute::ActivationLayerInfo activationInfo = ConvertActivationDescriptorToAclActivationLayerInfo(
32 activationDescriptor);
33
34 // At the time of writing, configure() will fail if a rounding policy other than TO_ZERO is supplied to it,
35 // when providing a scale of 1.0 for F32 tensors, even though the provided rounding policy appears to be
36 // ignored for F32 tensors.
37 return arm_compute::CLPixelWiseMultiplication::validate(&aclInput1,
38 &aclInput2,
39 &aclOutput,
40 1.0f,
41 convertPolicy,
42 arm_compute::RoundingPolicy::TO_ZERO,
43 activationInfo);
44 }
45
46
ClMultiplicationWorkload(const MultiplicationQueueDescriptor & descriptor,const WorkloadInfo & info,const arm_compute::CLCompileContext & clCompileContext)47 ClMultiplicationWorkload::ClMultiplicationWorkload(const MultiplicationQueueDescriptor& descriptor,
48 const WorkloadInfo& info,
49 const arm_compute::CLCompileContext& clCompileContext)
50 : ClBaseWorkload<MultiplicationQueueDescriptor>(descriptor, info)
51 {
52 m_Data.ValidateInputsOutputs("ClMultiplicationWorkload", 2, 1);
53
54 arm_compute::ICLTensor& input0 = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
55 arm_compute::ICLTensor& input1 = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
56 arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
57
58 auto convertPolicy = (IsQuantizedType(info.m_InputTensorInfos[0].GetDataType()) ||
59 IsQuantizedType(info.m_InputTensorInfos[1].GetDataType())) ?
60 arm_compute::ConvertPolicy::SATURATE :
61 arm_compute::ConvertPolicy::WRAP;
62
63 const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
64
65 {
66 ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClMultiplicationWorkload_configure");
67 // Construct
68 m_PixelWiseMultiplication.configure(clCompileContext,
69 &input0,
70 &input1,
71 &output,
72 1.0f,
73 convertPolicy,
74 arm_compute::RoundingPolicy::TO_NEAREST_EVEN,
75 activationInfo);
76 }
77 }
78
Execute() const79 void ClMultiplicationWorkload::Execute() const
80 {
81 ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClMultiplicationWorkload_Execute", this->GetGuid());
82 RunClFunction(m_PixelWiseMultiplication, CHECK_LOCATION());
83 }
84
85 } //namespace armnn
86