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
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
|
//
// Copyright © 2024 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "GpuFsaConvolution2d.hpp"
#include "UtilsGpuFsa.hpp"
#include <aclCommon/ArmComputeTensorUtils.hpp>
#include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadContext.h>
#include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h>
#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuConv2d.h>
#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h>
#include <vector>
using namespace arm_compute::experimental::dynamic_fusion;
using namespace armnn::armcomputetensorutils;
namespace armnn
{
arm_compute::Status GpuFsaConvolution2dValidate(const TensorInfo& input,
const Convolution2dDescriptor& descriptor,
const TensorInfo& weights,
const Optional<TensorInfo>& biases)
{
// Create a new workload sketch, for validation purposes
auto compileCtx = arm_compute::CLKernelLibrary::get().get_compile_context();
auto workloadContext = GpuWorkloadContext(&compileCtx);
GpuWorkloadSketch sketch{ &workloadContext };
// Build and create tensor infos using the sketch
const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout);
aclWeightsInfo.set_are_values_constant(weights.IsConstant());
auto inputInfo = workloadContext.create_tensor_info(aclInputInfo);
auto weightInfo = workloadContext.create_tensor_info(aclWeightsInfo);
// Only create the bias tensor info if enabled, otherwise pass nullptr to validate_op
arm_compute::TensorInfo aclBiasInfo;
arm_compute::ITensorInfo* biasSketchInfoPtr = nullptr;
if (descriptor.m_BiasEnabled)
{
if(!biases.has_value())
{
throw InvalidArgumentException("GpuFsaConvolution2d::ValidateOp: No biases set when biases are enabled");
}
aclBiasInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
aclBiasInfo.set_are_values_constant(biases.value().IsConstant());
biasSketchInfoPtr = workloadContext.create_tensor_info(aclBiasInfo);
}
Conv2dAttributes conv2dAttributes = CreateConv2dAttributes(descriptor);
// Validate operator, check status and update reasonIfUnsupported
arm_compute::Status aclStatus = GpuConv2d::validate_op(sketch,
inputInfo,
weightInfo,
biasSketchInfoPtr,
conv2dAttributes);
return aclStatus;
}
void GpuFsaConvolution2dCreateOp(GpuFsaPreCompiledBlob* blob,
const TensorInfo& input,
const Convolution2dDescriptor& descriptor,
const TensorInfo& weights,
const Optional<TensorInfo>& biases)
{
/*
* Creating an Op for the GpuFsa backend requires us to create and maintain quite a bit of data, which is then stored
* in a GpuFsaPreCompiledBlob for execution later. Specifically we need:
* GpuWorkloadContext, this contains the TensorInfos and is unique to the Graph being executed
* Sketch, this is similar to a subgraph and can contain one or more operations. Multiple ops can be "fused" together
* using a single sketch.
* The inputTensorinfos / outputTensorInfos, these are pointers to the TensorInfos used when creating the sketch.
* They refer to the TensorInfos stored within the GpuWorkloadContext and are needed when executing the sketch
* as the TensorInfos used when creating the Tensors must match those used to create the Sketch. Otherwise the runtime
* doesn't know which Tensors to use.
*/
GpuWorkloadSketch* sketch = blob->sketch.get();
GpuWorkloadContext* workloadContext = blob->workloadContext.get();
std::vector<arm_compute::ITensorInfo*> inputTensorInfos = {};
std::vector<arm_compute::ITensorInfo*> outputTensorInfos = {};
// Build and create tensor infos using the sketch
const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout);
aclWeightsInfo.set_are_values_constant(weights.IsConstant());
inputTensorInfos.emplace_back(workloadContext->create_tensor_info(aclInputInfo));
inputTensorInfos.emplace_back(workloadContext->create_tensor_info(aclWeightsInfo));
// Only create the bias tensor info if enabled, otherwise pass nullptr to validate_op / create_op
arm_compute::TensorInfo aclBiasInfo;
arm_compute::ITensorInfo* biasSketchInfoPtr = nullptr;
if (descriptor.m_BiasEnabled)
{
if(!biases.has_value())
{
throw InvalidArgumentException("GpuFsaConvolution2d::CreateOp: No biases set when biases are enabled");
}
aclBiasInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
aclBiasInfo.set_are_values_constant(biases.value().IsConstant());
inputTensorInfos.emplace_back(workloadContext->create_tensor_info(aclBiasInfo));
biasSketchInfoPtr = inputTensorInfos[2];
}
Conv2dAttributes conv2dAttributes = CreateConv2dAttributes(descriptor);
// Validate operator, check status and update reasonIfUnsupported
arm_compute::Status aclStatus = GpuConv2d::validate_op(*sketch,
inputTensorInfos[0],
inputTensorInfos[1],
biasSketchInfoPtr,
conv2dAttributes);
const bool supported = (aclStatus.error_code() == arm_compute::ErrorCode::OK);
if (!supported)
{
throw BackendCapabilityException("\"GpuFsa\" backend failed during Convolution2D operation validation");
}
// Create the Op within the Sketch using the TensorInfos we have stored
arm_compute::ITensorInfo* convOutInfo = GpuConv2d::create_op(*sketch,
inputTensorInfos[0],
inputTensorInfos[1],
biasSketchInfoPtr,
conv2dAttributes);
// Create the Output
outputTensorInfos.emplace_back(workloadContext->create_tensor_info());
GpuOutput::create_op(*sketch, convOutInfo, outputTensorInfos[0]);
// Store the TensorInfos within the blob as unique_ptrs to be used later
blob->inputTensorInfos = std::make_unique<std::vector<arm_compute::ITensorInfo*>>(inputTensorInfos);
blob->outputTensorInfos = std::make_unique<std::vector<arm_compute::ITensorInfo*>>(outputTensorInfos);
}
} // namespace armnn
|
