forked from TensorStack-AI/TensorStack
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathOrtExtensions.cs
More file actions
301 lines (257 loc) · 12.2 KB
/
OrtExtensions.cs
File metadata and controls
301 lines (257 loc) · 12.2 KB
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
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
// Copyright (c) TensorStack. All rights reserved.
// Licensed under the Apache 2.0 License.
using Microsoft.ML.OnnxRuntime;
using System;
using TensorStack.Core.Inference;
using TensorStack.Common.Tensor;
using TensorStack.Common;
namespace TensorStack.Core
{
/// <summary>
/// Helper extensions for conversion from OrtValue to Tensor, TensorSpan
/// </summary>
public static class OrtExtensions
{
/// <summary>
/// Creates an allocated output buffer on the device.
/// </summary>
/// <param name="metadata">The metadata.</param>
/// <param name="dimensions">The dimensions.</param>
/// <returns>OrtValue.</returns>
public static OrtValue CreateOutputBuffer(this NamedMetadata metadata, ReadOnlySpan<int> dimensions)
{
return OrtValue.CreateAllocatedTensorValue(OrtAllocator.DefaultInstance, metadata.Value.ElementDataType, dimensions.ToLong());
}
/// <summary>
/// Span access to the OrtValue data
/// </summary>
/// <typeparam name="T"></typeparam>
/// <param name="ortValue">The ort value.</param>
/// <returns>ReadOnlySpan<T>.</returns>
public static ReadOnlySpan<T> AsSpan<T>(this OrtValue ortValue) where T : unmanaged
{
return ortValue.GetTensorDataAsSpan<T>();
}
/// <summary>
/// ReadOnlySpan access to the OrtValue data
/// </summary>
/// <typeparam name="T"></typeparam>
/// <param name="ortValue">The ort value.</param>
/// <returns>ReadOnlySpan<T>.</returns>
public static ReadOnlySpan<T> AsReadOnlySpan<T>(this OrtValue ortValue) where T : unmanaged
{
return ortValue.GetTensorDataAsSpan<T>();
}
/// <summary>
/// Create a view of the OrtValue as TensorSpan
/// </summary>
/// <param name="ortValue">The ort value.</param>
/// <returns>TensorSpan<System.Single>.</returns>
public static TensorSpan<float> AsTensorSpan(this OrtValue ortValue)
{
var dimensions = ortValue.GetDimensions();
var typeInfo = ortValue.GetTensorTypeAndShape();
return typeInfo.ElementDataType switch
{
// NOTE: Float16 & BFloat16 will cause a copy off device
// These types dont have math functions so makes sense to copy to float for convienece
Microsoft.ML.OnnxRuntime.Tensors.TensorElementType.Float16 => ortValue.ToTensor().AsTensorSpan(),
Microsoft.ML.OnnxRuntime.Tensors.TensorElementType.BFloat16 => ortValue.ToTensor().AsTensorSpan(),
_ => new TensorSpan<float>(ortValue.GetTensorMutableDataAsSpan<float>(), dimensions)
};
}
/// <summary>
/// Create a view of the OrtValue as TensorSpan
/// </summary>
/// <typeparam name="T"></typeparam>
/// <param name="ortValue">The ort value.</param>
/// <returns>TensorSpan<T>.</returns>
public static TensorSpan<T> AsTensorSpan<T>(this OrtValue ortValue) where T : unmanaged
{
return new TensorSpan<T>(ortValue.GetTensorMutableDataAsSpan<T>(), ortValue.GetDimensions());
}
/// <summary>
/// Copy TensorSpan data to OrtValue.
/// </summary>
/// <param name="tensor">The tensor.</param>
/// <param name="metadata">The metadata.</param>
/// <returns>OrtValue.</returns>
public static OrtValue ToOrtValue<T>(this TensorSpan<T> tensor, NamedMetadata metadata) where T : unmanaged
{
return OrtValue.CreateTensorValueFromMemory<T>(OrtMemoryInfo.DefaultInstance, tensor.Span.ToArray(), tensor.Dimensions.ToLong());
}
/// <summary>
/// Copy TensorSpan data to OrtValue.
/// </summary>
/// <param name="tensor">The tensor.</param>
/// <param name="metadata">The metadata.</param>
/// <returns>OrtValue.</returns>
public static OrtValue ToOrtValue(this TensorSpan<string> tensor, NamedMetadata metadata)
{
return OrtValue.CreateFromStringTensor(new Microsoft.ML.OnnxRuntime.Tensors.DenseTensor<string>(tensor.Span.ToArray(), tensor.Dimensions));
}
/// <summary>
/// Copy TensorSpan data to OrtValue.
/// </summary>
/// <param name="tensor">The tensor.</param>
/// <param name="metadata">The metadata.</param>
/// <returns>OrtValue.</returns>
public static OrtValue ToOrtValue(this TensorSpan<float> tensor, NamedMetadata metadata)
{
var dimensions = tensor.Dimensions.ToLong();
return metadata.Value.ElementDataType switch
{
Microsoft.ML.OnnxRuntime.Tensors.TensorElementType.Int64 => OrtValue.CreateTensorValueFromMemory(OrtMemoryInfo.DefaultInstance, tensor.Span.ToLongMemory(), dimensions),
Microsoft.ML.OnnxRuntime.Tensors.TensorElementType.Float16 => OrtValue.CreateTensorValueFromMemory(OrtMemoryInfo.DefaultInstance, tensor.Span.ToFloat16Memory(), dimensions),
Microsoft.ML.OnnxRuntime.Tensors.TensorElementType.BFloat16 => OrtValue.CreateTensorValueFromMemory(OrtMemoryInfo.DefaultInstance, tensor.Span.ToBFloat16Memory(), dimensions),
_ => OrtValue.CreateTensorValueFromMemory(OrtMemoryInfo.DefaultInstance, tensor.Span.ToFloatMemory(), dimensions)
};
}
/// <summary>
/// Copy OrtValue data to float Tensor.
/// </summary>
/// <typeparam name="T"></typeparam>
/// <param name="ortValue">The ort value.</param>
/// <returns>Tensor<T>.</returns>
public static Tensor<T> ToTensor<T>(this OrtValue ortValue) where T : unmanaged
{
return new Tensor<T>(ortValue.GetTensorDataAsSpan<T>().ToArray(), ortValue.GetDimensions());
}
/// <summary>
/// Copy OrtValue data to float Tensor.
/// </summary>
/// <param name="ortValue">The ort value.</param>
/// <returns>Tensor<System.Single>.</returns>
public static Tensor<float> ToTensor(this OrtValue ortValue)
{
var dimensions = ortValue.GetDimensions();
var typeInfo = ortValue.GetTensorTypeAndShape();
return typeInfo.ElementDataType switch
{
Microsoft.ML.OnnxRuntime.Tensors.TensorElementType.Float16 => new Tensor<float>(ortValue.GetTensorMutableDataAsSpan<Float16>().ToFloatMemory(), dimensions),
Microsoft.ML.OnnxRuntime.Tensors.TensorElementType.BFloat16 => new Tensor<float>(ortValue.GetTensorMutableDataAsSpan<BFloat16>().ToFloatMemory(), dimensions),
_ => new Tensor<float>(ortValue.GetTensorDataAsSpan<float>().ToArray(), dimensions)
};
}
/// <summary>
/// Copy OrtValue data to array.
/// </summary>
/// <typeparam name="T"></typeparam>
/// <param name="ortValue">The ort value.</param>
/// <returns>T[].</returns>
public static T[] ToArray<T>(this OrtValue ortValue) where T : unmanaged
{
return ortValue.AsReadOnlySpan<T>().ToArray();
}
/// <summary>
/// Copy OrtValue data to flot array.
/// </summary>
/// <param name="ortValue">The ort value.</param>
/// <returns>System.Single[].</returns>
public static float[] ToArray(this OrtValue ortValue)
{
var typeInfo = ortValue.GetTensorTypeAndShape();
return typeInfo.ElementDataType switch
{
Microsoft.ML.OnnxRuntime.Tensors.TensorElementType.Float16 => ortValue.GetTensorMutableDataAsSpan<Float16>().ToFloatMemory().ToArray(),
Microsoft.ML.OnnxRuntime.Tensors.TensorElementType.BFloat16 => ortValue.GetTensorMutableDataAsSpan<BFloat16>().ToFloatMemory().ToArray(),
_ => ortValue.GetTensorDataAsSpan<float>().ToArray()
};
}
/// <summary>
/// Converts Optimization to GraphOptimizationLevel.
/// </summary>
/// <param name="configuration">The configuration.</param>
/// <returns>GraphOptimizationLevel.</returns>
public static GraphOptimizationLevel ToGraphOptimizationLevel(this Optimization configuration)
{
return configuration switch
{
Optimization.None => GraphOptimizationLevel.ORT_DISABLE_ALL,
Optimization.Basic => GraphOptimizationLevel.ORT_ENABLE_BASIC,
Optimization.Extended => GraphOptimizationLevel.ORT_ENABLE_EXTENDED,
Optimization.All => GraphOptimizationLevel.ORT_ENABLE_ALL,
_ => GraphOptimizationLevel.ORT_DISABLE_ALL,
};
}
/// <summary>
/// Gets the dimensions.
/// </summary>
/// <param name="ortValue">The ort value.</param>
/// <returns>System.Int32[].</returns>
private static int[] GetDimensions(this OrtValue ortValue)
{
return ortValue.GetTensorTypeAndShape().Shape.ToInt();
}
/// <summary>
/// Copy float Span to long Memory
/// </summary>
/// <param name="inputMemory">The input memory.</param>
/// <returns>Memory<System.Int64>.</returns>
private static Memory<long> ToLongMemory(this Span<float> inputMemory)
{
return Array.ConvertAll(inputMemory.ToArray(), Convert.ToInt64).AsMemory();
}
/// <summary>
/// Copy float Span to float Memory .
/// </summary>
/// <param name="inputMemory">The input memory.</param>
/// <returns>Memory<System.Single>.</returns>
private static Memory<float> ToFloatMemory(this Span<float> inputMemory)
{
return inputMemory.ToArray().AsMemory();
}
/// <summary>
/// Copy float Span to Float16 Memory.
/// </summary>
/// <param name="inputMemory">The input memory.</param>
/// <returns>Memory<Float16>.</returns>
private static Memory<Float16> ToFloat16Memory(this Span<float> inputMemory)
{
var elementCount = inputMemory.Length;
var floatArray = GC.AllocateUninitializedArray<Float16>(elementCount);
for (int i = 0; i < elementCount; i++)
floatArray[i] = (Float16)inputMemory[i];
return floatArray.AsMemory();
}
/// <summary>
/// Copy to float Span tp BFloat16 Memory.
/// </summary>
/// <param name="inputMemory">The input memory.</param>
/// <returns>Memory<BFloat16>.</returns>
private static Memory<BFloat16> ToBFloat16Memory(this Span<float> inputMemory)
{
var elementCount = inputMemory.Length;
var floatArray = GC.AllocateUninitializedArray<BFloat16>(elementCount);
for (int i = 0; i < elementCount; i++)
floatArray[i] = (BFloat16)inputMemory[i];
return floatArray.AsMemory();
}
/// <summary>
/// Copt to Float16 Span to float Memory.
/// </summary>
/// <param name="inputMemory">The input memory.</param>
/// <returns>Memory<System.Single>.</returns>
private static Memory<float> ToFloatMemory(this Span<Float16> inputMemory)
{
var elementCount = inputMemory.Length;
var floatArray = GC.AllocateUninitializedArray<float>(elementCount);
for (int i = 0; i < elementCount; i++)
floatArray[i] = (float)inputMemory[i];
return floatArray.AsMemory();
}
/// <summary>
/// Copy to BFloat16 Span to float Memory
/// </summary>
/// <param name="inputMemory">The input memory.</param>
/// <returns>Memory<System.Single>.</returns>
private static Memory<float> ToFloatMemory(this Span<BFloat16> inputMemory)
{
var elementCount = inputMemory.Length;
var floatArray = GC.AllocateUninitializedArray<float>(elementCount);
for (int i = 0; i < elementCount; i++)
floatArray[i] = (float)inputMemory[i];
return floatArray.AsMemory();
}
}
}