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
|
//
// Copyright © 2023 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "GpuFsaConvolution2dValidate.hpp"
#include <armnn/Types.hpp>
#include <armnn/utility/IgnoreUnused.hpp>
#include <aclCommon/ArmComputeTensorUtils.hpp>
#include <arm_compute/core/ITensorInfo.h>
#include <arm_compute/core/TensorInfo.h>
#include <arm_compute/core/TensorShape.h>
#include <arm_compute/core/CL/CLKernelLibrary.h>
#include <arm_compute/core/CL/CLCompileContext.h>
#include <arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h>
#include <arm_compute/dynamic_fusion/sketch/OperatorAttributes.h>
#include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadContext.h>
#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuConv2d.h>
#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h>
#include <vector>
#include <iostream>
namespace armnn
{
using namespace armcomputetensorutils;
inline arm_compute::Status ValidateAndCreateOp(const TensorInfo& input,
const Convolution2dDescriptor& descriptor,
const TensorInfo& weights,
const Optional<TensorInfo>& biases,
GpuWorkloadSketch& sketch,
const bool createOp = false)
{
// 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 = sketch.create_tensor_info(aclInputInfo);
auto weightInfo = sketch.create_tensor_info(aclWeightsInfo);
// Only create the bias tensor info if enabled, otherwise pass nullptr to validate_op
arm_compute::TensorInfo aclBiasInfo;
arm_compute::TensorInfo biasSketchInfo;
arm_compute::TensorInfo* biasSketchInfoPtr = nullptr;
if (descriptor.m_BiasEnabled)
{
ARMNN_ASSERT(biases.has_value());
aclBiasInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
aclBiasInfo.set_are_values_constant(biases.value().IsConstant());
biasSketchInfo = sketch.create_tensor_info(aclBiasInfo);
biasSketchInfoPtr = &biasSketchInfo;
}
// Set Conv2d attributes using descriptor
const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(descriptor.m_DilationX,
descriptor.m_DilationY);
const arm_compute::Padding2D aclPadInfo = BuildArmComputePaddingInfo(descriptor);
const arm_compute::Size2D aclStrideInfo = BuildArmComputeSize2D(descriptor.m_StrideX, descriptor.m_StrideY);
Conv2dAttributes conv2DAttributes{};
conv2DAttributes.dilation(aclDilationInfo);
conv2DAttributes.pad(aclPadInfo);
conv2DAttributes.stride(aclStrideInfo);
// Validate operator, check status and update reasonIfUnsupported
arm_compute::Status aclStatus = GpuConv2d::validate_op(sketch,
&inputInfo,
&weightInfo,
biasSketchInfoPtr,
conv2DAttributes);
if (createOp)
{
const bool supported = (aclStatus.error_code() == arm_compute::ErrorCode::OK);
if (!supported)
{
throw BackendCapabilityException("\"GpuFsa\" backend failed during operation validation when attempting "
"to fuse a GpuConv2d operator into the existing workload sketch.");
}
arm_compute::ITensorInfo* convOutInfo = GpuConv2d::create_op(sketch,
&inputInfo,
&weightInfo,
biasSketchInfoPtr,
conv2DAttributes);
// Temporary fix until fusing attempt is make for GpuFsa backend and Output layer workload is created.
auto outputInfo = sketch.create_tensor_info();
GpuOutput::create_op(sketch, convOutInfo, &outputInfo);
}
return aclStatus;
}
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 gpuCtx = GpuWorkloadContext(&compileCtx);
GpuWorkloadSketch sketch{ &gpuCtx };
return ValidateAndCreateOp(input, descriptor, weights, biases, sketch);
}
void GpuFsaConvolution2dCreateOp(const TensorInfo& input,
const Convolution2dDescriptor& descriptor,
const TensorInfo& weights,
const Optional<TensorInfo>& biases,
GpuWorkloadSketch& sketch)
{
ValidateAndCreateOp(input, descriptor, weights, biases, sketch, true);
}
} // namespace armnn
|
