1 //
2 // Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "ClConvolution3dWorkload.hpp"
7
8 #include "ClWorkloadUtils.hpp"
9
10 #include <cl/ClLayerSupport.hpp>
11 #include <cl/ClTensorHandle.hpp>
12 #include <cl/ClLayerSupport.hpp>
13 #include <aclCommon/ArmComputeUtils.hpp>
14 #include <aclCommon/ArmComputeTensorUtils.hpp>
15 #include <armnn/backends/TensorHandle.hpp>
16
17 #include <arm_compute/runtime/CL/functions/CLConv3D.h>
18
19 namespace armnn
20 {
21 using namespace armcomputetensorutils;
22
ClConvolution3dWorkloadValidate(const TensorInfo & input,const TensorInfo & output,const Convolution3dDescriptor & descriptor,const TensorInfo & weights,const Optional<TensorInfo> & biases,bool isFastMathEnabled,const ActivationDescriptor * activationDescriptor)23 arm_compute::Status ClConvolution3dWorkloadValidate(const TensorInfo& input,
24 const TensorInfo& output,
25 const Convolution3dDescriptor& descriptor,
26 const TensorInfo& weights,
27 const Optional<TensorInfo>& biases,
28 bool isFastMathEnabled,
29 const ActivationDescriptor* activationDescriptor)
30 {
31 const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
32 const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout);
33
34 arm_compute::TensorInfo aclBiasesInfo;
35 arm_compute::TensorInfo* optionalAclBiasesInfo = nullptr;
36 if (descriptor.m_BiasEnabled)
37 {
38 ARMNN_ASSERT(biases.has_value());
39 aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
40 optionalAclBiasesInfo = &aclBiasesInfo;
41 }
42
43 const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
44
45 const arm_compute::Conv3dInfo aclConv3DInfo = ComputeConv3DInfo(descriptor,
46 isFastMathEnabled,
47 activationDescriptor);
48
49 return arm_compute::CLConv3D::validate(&aclInputInfo,
50 &aclWeightsInfo,
51 optionalAclBiasesInfo,
52 &aclOutputInfo,
53 aclConv3DInfo);
54 }
55
ClConvolution3dWorkload(const Convolution3dQueueDescriptor & descriptor,const WorkloadInfo & info,std::shared_ptr<arm_compute::MemoryManagerOnDemand> & memoryManager,const arm_compute::CLCompileContext & clCompileContext,const bool isFastMathEnabled)56 ClConvolution3dWorkload::ClConvolution3dWorkload(const Convolution3dQueueDescriptor& descriptor,
57 const WorkloadInfo& info,
58 std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager,
59 const arm_compute::CLCompileContext& clCompileContext,
60 const bool isFastMathEnabled)
61 : ClBaseWorkload<Convolution3dQueueDescriptor>(descriptor, info)
62 , m_ConvolutionLayer()
63 {
64 IgnoreUnused(memoryManager);
65
66 uint32_t numInputs = m_Data.m_Parameters.m_BiasEnabled ? 3: 2;
67 m_Data.ValidateInputsOutputs("ClConvolution3dWorkload", numInputs, 1);
68
69 arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
70 arm_compute::ICLTensor& weights = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
71 arm_compute::ICLTensor* biasesPtr = nullptr;
72 if (m_Data.m_Parameters.m_BiasEnabled)
73 {
74 biasesPtr = &static_cast<IClTensorHandle*>(m_Data.m_Inputs[2])->GetTensor();
75 }
76 arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
77
78 arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
79 input.info()->set_data_layout(aclDataLayout);
80 weights.info()->set_data_layout(aclDataLayout);
81 output.info()->set_data_layout(aclDataLayout);
82
83 const arm_compute::Conv3dInfo aclConv3DInfo = ComputeConv3DInfo(descriptor,
84 isFastMathEnabled);
85
86 {
87 ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClConvolution3dWorkload_configure");
88 m_ConvolutionLayer.configure(clCompileContext,
89 &input,
90 &weights,
91 biasesPtr,
92 &output,
93 aclConv3DInfo);
94 }
95 // Add details for profiling output
96 WorkloadInfo detailsInfo;
97
98 detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos;
99 detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos;
100
101 // Report Profiling Details
102 ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClConvolution3dWorkload_Construct",
103 descriptor.m_Parameters,
104 detailsInfo,
105 this->GetGuid());
106
107 // Force Compute Library to perform the necessary copying and reshaping, after which
108 // delete all the input tensors that will no longer be needed
109 m_ConvolutionLayer.prepare();
110 }
111
Execute() const112 void ClConvolution3dWorkload::Execute() const
113 {
114 ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClConvolution3dWorkload_Execute", this->GetGuid());
115 RunClFunction(m_ConvolutionLayer, CHECK_LOCATION());
116 }
117
118 } //namespace armnn
119