1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
|
//
// Copyright © 2022-2023 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "ClBatchMatMulWorkload.hpp"
#include "ClWorkloadUtils.hpp"
#include <aclCommon/ArmComputeTensorUtils.hpp>
#include <aclCommon/ArmComputeUtils.hpp>
#include <armnn/utility/PolymorphicDowncast.hpp>
#include <backendsCommon/WorkloadUtils.hpp>
#include <cl/ClTensorHandle.hpp>
#include <arm_compute/function_info/MatMulInfo.h>
namespace armnn
{
arm_compute::Status ClBatchMatMulValidate(const TensorInfo& inputInfoX,
const TensorInfo& inputInfoY,
const TensorInfo& outputInfo,
const BatchMatMulDescriptor& descriptor,
const ActivationDescriptor* activationDescriptor)
{
if (descriptor.m_AdjointX || descriptor.m_AdjointY )
{
throw Exception("Support for adjoint not implemented.");
}
if (descriptor.m_DataLayoutX != armnn::DataLayout::NCHW || descriptor.m_DataLayoutY != armnn::DataLayout::NCHW )
{
throw Exception("Only supported the MatMul in the last 2 dimensions");
}
arm_compute::TensorInfo aclInputInfoX = armcomputetensorutils::BuildArmComputeTensorInfo(inputInfoX);
arm_compute::TensorInfo aclInputInfoY = armcomputetensorutils::BuildArmComputeTensorInfo(inputInfoY);
const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(outputInfo);
// GeMM dispatches kernel handles dynamic inputs differently to static so this flag needs to be set
aclInputInfoX.set_are_values_constant(false);
aclInputInfoY.set_are_values_constant(false);
const arm_compute::ActivationLayerInfo activationInfo = ConvertActivationDescriptorToAclActivationLayerInfo(
activationDescriptor);
arm_compute::MatMulInfo matMulInfo;
matMulInfo.adj_lhs(descriptor.m_TransposeX);
matMulInfo.adj_rhs(descriptor.m_TransposeY);
return arm_compute::CLMatMul::validate(&aclInputInfoX, &aclInputInfoY, &aclOutputInfo, matMulInfo, activationInfo);
}
ClBatchMatMulWorkload::ClBatchMatMulWorkload(const BatchMatMulQueueDescriptor& descriptor,
const WorkloadInfo& info,
const arm_compute::CLCompileContext& clCompileContext)
: ClBaseWorkload<BatchMatMulQueueDescriptor>(descriptor, info)
{
// Report Profiling Details
ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClBatchMatMulWorkload_Construct",
descriptor.m_Parameters,
info,
this->GetGuid());
if (descriptor.m_Parameters.m_AdjointX || descriptor.m_Parameters.m_AdjointY )
{
throw Exception("Support for adjoint not implemented.");
}
if (descriptor.m_Parameters.m_DataLayoutX != armnn::DataLayout::NCHW ||
descriptor.m_Parameters.m_DataLayoutY != armnn::DataLayout::NCHW )
{
throw Exception("Only supported the MatMul in the last 2 dimensions");
}
m_Data.ValidateInputsOutputs("ClBatchMatMulWorkload", 2, 1);
arm_compute::ICLTensor& inputX = PolymorphicDowncast<ClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
arm_compute::ICLTensor& inputY = PolymorphicDowncast<ClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
auto outputHandle = PolymorphicDowncast<ClTensorHandle*>(m_Data.m_Outputs[0]);
arm_compute::ICLTensor& output = outputHandle->GetTensor();
// GeMM dispatches kernel handles dynamic inputs differently to static so this flag needs to be set
inputX.info()->set_are_values_constant(false);
inputY.info()->set_are_values_constant(false);
const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
arm_compute::MatMulInfo matMulInfo;
matMulInfo.adj_lhs(descriptor.m_Parameters.m_TransposeX);
matMulInfo.adj_rhs(descriptor.m_Parameters.m_TransposeY);
arm_compute::GpuMatMulSettings settings;
m_MatMulLayer.configure(clCompileContext, &inputX, &inputY, &output, matMulInfo, settings, activationInfo);
// Report Profiling Details
WorkloadInfo detailsInfo;
detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos;
detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos;
ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClBatchMatMulWorkload_Construct",
descriptor.m_Parameters,
detailsInfo,
GetGuid());
}
void ClBatchMatMulWorkload::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClBatchMatMulWorkload_Execute", this->GetGuid());
RunClFunction(m_MatMulLayer, CHECK_LOCATION());
}
} //namespace armnn
|
