File size: 9,352 Bytes
a0e12a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
{
  "domain": "ai.onnx",
  "name": "BatchNormalization",
  "sinceVersion": 15,
  "description": "Applies inference-mode batch normalization: `Y = (X - input_mean) / sqrt(input_var + epsilon) * scale + B`. This package supports `training_mode=0`, rank-2-or-higher inputs, and a common float16 or float32 dtype for every tensor. ONNX training mode is intentionally not implemented because this inference-only release does not expose its required running-mean and running-variance outputs.",
  "inputs": [
    {
      "role": "X",
      "dtype": "T",
      "description": "Input data tensor with shape `(N, C, D1, ..., Dn)`, normalized independently per channel using the supplied estimated statistics."
    },
    { "role": "scale", "dtype": "T", "rank": 1, "description": "Per-channel scale tensor with shape `(C)`." },
    { "role": "B", "dtype": "T", "rank": 1, "description": "Per-channel bias tensor with shape `(C)`." },
    {
      "role": "input_mean",
      "dtype": "T",
      "rank": 1,
      "description": "Precomputed estimated mean tensor with shape `(C)` used for inference."
    },
    {
      "role": "input_var",
      "dtype": "T",
      "rank": 1,
      "description": "Precomputed estimated variance tensor with shape `(C)` used for inference."
    }
  ],
  "outputs": [
    {
      "role": "Y",
      "dtype": "T",
      "rank": "ranks.X",
      "description": "Batch-normalized output tensor with the same shape as `X`.",
      "shape": "shapes.X"
    }
  ],
  "attributes": { "epsilon": 0.00001, "momentum": 0.9, "training_mode": 0 },
  "attributeDescriptions": {
    "epsilon": "Small value added to the variance before taking the square root to avoid division by zero.",
    "momentum": "Standard ONNX running-statistics momentum. This inference-only package accepts the default `0.9`; non-default values are reserved for the unsupported training-state update.",
    "training_mode": "Execution mode. This inference-only package supports the default value 0; value 1 is rejected because the ONNX training outputs are not exposed."
  },
  "attributeConstraints": {
    "momentum": { "values": [0.9], "comparison": "float32" },
    "training_mode": { "values": [0] }
  },
  "typeConstraints": { "T": ["float32", "float16"] },
  "args": {
    "x": { "kind": "tensor", "semantic": "X", "role": "input" },
    "scale": { "kind": "tensor", "semantic": "scale", "role": "input" },
    "b": { "kind": "tensor", "semantic": "B", "role": "input" },
    "inputMean": { "kind": "tensor", "semantic": "input_mean", "role": "input" },
    "inputVar": { "kind": "tensor", "semantic": "input_var", "role": "input" },
    "y": { "kind": "tensor", "semantic": "Y", "role": "output" }
  },
  "tunables": { "WORKGROUP_SIZE": 256 },
  "derive": {
    "normalizationParamsOk": "ranks.scale == 1 and ranks.B == 1 and ranks.input_mean == 1 and ranks.input_var == 1 and dim(shapes.scale, 0) == dim(shapes.X, 1) and dim(shapes.B, 0) == dim(shapes.X, 1) and dim(shapes.input_mean, 0) == dim(shapes.X, 1) and dim(shapes.input_var, 0) == dim(shapes.X, 1)",
    "inferenceContractOk": "f16Ok(dtypes.T) and ranks.X >= 2 and ranks.Y == ranks.X and sameShape(shapes.Y, shapes.X) and normalizationParamsOk"
  },
  "bindingSets": {
    "ncInferenceVec4": [
      {
        "name": "x",
        "arg": "x",
        "semantic": "X",
        "buffer": { "type": "read-only-storage" },
        "elementType": "vec4<f32>"
      },
      {
        "name": "scale",
        "arg": "scale",
        "semantic": "scale",
        "buffer": { "type": "read-only-storage" },
        "elementType": "vec4<f32>"
      },
      {
        "name": "bias",
        "arg": "b",
        "semantic": "B",
        "buffer": { "type": "read-only-storage" },
        "elementType": "vec4<f32>"
      },
      {
        "name": "input_mean",
        "arg": "inputMean",
        "semantic": "input_mean",
        "buffer": { "type": "read-only-storage" },
        "elementType": "vec4<f32>"
      },
      {
        "name": "input_var",
        "arg": "inputVar",
        "semantic": "input_var",
        "buffer": { "type": "read-only-storage" },
        "elementType": "vec4<f32>"
      },
      { "name": "y", "arg": "y", "semantic": "Y", "buffer": { "type": "storage" }, "elementType": "vec4<f32>" },
      {
        "name": "params",
        "semantic": "kernel.params",
        "buffer": { "type": "uniform" },
        "struct": {
          "name": "Params",
          "fields": [
            { "name": "count4", "type": "u32", "value": "numel(shapes.Y) / 4" },
            { "name": "channels4", "type": "u32", "value": "dim(shapes.X, 1) / 4" },
            { "name": "epsilon", "type": "f32", "value": "attrs.epsilon" }
          ]
        }
      }
    ],
    "spatialInferenceVec4": [
      {
        "name": "x",
        "arg": "x",
        "semantic": "X",
        "buffer": { "type": "read-only-storage" },
        "elementType": "vec4<f32>"
      },
      {
        "name": "scale",
        "arg": "scale",
        "semantic": "scale",
        "buffer": { "type": "read-only-storage" },
        "elementType": "$T"
      },
      { "name": "bias", "arg": "b", "semantic": "B", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
      {
        "name": "input_mean",
        "arg": "inputMean",
        "semantic": "input_mean",
        "buffer": { "type": "read-only-storage" },
        "elementType": "$T"
      },
      {
        "name": "input_var",
        "arg": "inputVar",
        "semantic": "input_var",
        "buffer": { "type": "read-only-storage" },
        "elementType": "$T"
      },
      { "name": "y", "arg": "y", "semantic": "Y", "buffer": { "type": "storage" }, "elementType": "vec4<f32>" },
      {
        "name": "params",
        "semantic": "kernel.params",
        "buffer": { "type": "uniform" },
        "struct": {
          "name": "Params",
          "fields": [
            { "name": "count4", "type": "u32", "value": "numel(shapes.Y) / 4" },
            { "name": "spatial4", "type": "u32", "value": "inner(shapes.X, 1) / 4" },
            { "name": "channels", "type": "u32", "value": "dim(shapes.X, 1)" },
            { "name": "epsilon", "type": "f32", "value": "attrs.epsilon" }
          ]
        }
      }
    ],
    "inferenceScalar": [
      { "name": "x", "arg": "x", "semantic": "X", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
      {
        "name": "scale",
        "arg": "scale",
        "semantic": "scale",
        "buffer": { "type": "read-only-storage" },
        "elementType": "$T"
      },
      { "name": "bias", "arg": "b", "semantic": "B", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
      {
        "name": "input_mean",
        "arg": "inputMean",
        "semantic": "input_mean",
        "buffer": { "type": "read-only-storage" },
        "elementType": "$T"
      },
      {
        "name": "input_var",
        "arg": "inputVar",
        "semantic": "input_var",
        "buffer": { "type": "read-only-storage" },
        "elementType": "$T"
      },
      { "name": "y", "arg": "y", "semantic": "Y", "buffer": { "type": "storage" }, "elementType": "$T" },
      {
        "name": "params",
        "semantic": "kernel.params",
        "buffer": { "type": "uniform" },
        "struct": {
          "name": "Params",
          "fields": [
            { "name": "channels", "type": "u32", "value": "dim(shapes.X, 1)" },
            { "name": "height", "type": "u32", "value": "1 if ranks.X == 2 else dim(shapes.X, 2)" },
            { "name": "width", "type": "u32", "value": "1 if ranks.X == 2 else inner(shapes.X, 2)" },
            { "name": "epsilon", "type": "f32", "value": "attrs.epsilon" },
            { "name": "count", "type": "u32", "value": "numel(shapes.Y)" }
          ]
        }
      }
    ]
  },
  "variants": [
    {
      "id": "inference_scalar",
      "when": ["inferenceContractOk"],
      "constants": { "scalar": "dtypes.T", "usesF16": "dtypes.T == \"f16\"" },
      "passes": [
        {
          "id": "main",
          "name": "BatchNormalization.InferenceScalar",
          "shader": "batch-normalization-nchw.wgsl.jinja",
          "bindings": "inferenceScalar",
          "dispatch": { "threads": "numel(shapes.Y)", "workgroupSize": "tunables.WORKGROUP_SIZE" }
        }
      ]
    },
    {
      "id": "nc_inference_vec4",
      "priority": 110,
      "when": ["inferenceContractOk", "dtypes.T == \"f32\"", "ranks.X == 2", "dim(shapes.X, 1) % 4 == 0"],
      "passes": [
        {
          "id": "main",
          "name": "BatchNormalization.NcInferenceVec4",
          "shader": "batch-normalization-nc-vec4.wgsl.jinja",
          "bindings": "ncInferenceVec4",
          "dispatch": { "threads": "numel(shapes.Y) / 4", "workgroupSize": "tunables.WORKGROUP_SIZE" }
        }
      ]
    },
    {
      "id": "nchw_inference_vec4",
      "priority": 100,
      "when": ["inferenceContractOk", "dtypes.T == \"f32\"", "ranks.X >= 3", "inner(shapes.X, 1) % 4 == 0"],
      "passes": [
        {
          "id": "main",
          "name": "BatchNormalization.InferenceVec4",
          "shader": "batch-normalization-nchw-vec4.wgsl.jinja",
          "bindings": "spatialInferenceVec4",
          "dispatch": { "threads": "numel(shapes.Y) / 4", "workgroupSize": "tunables.WORKGROUP_SIZE" }
        }
      ]
    }
  ]
}