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
|
#include "caffe2/operators/lengths_reducer_rowwise_8bit_ops.h"
#include "c10/util/Registry.h"
namespace caffe2 {
REGISTER_CPU_OPERATOR(
Rowwise8BitQuantizedToFloat,
Rowwise8BitQuantizedToFloatOp<CPUContext>);
REGISTER_CPU_OPERATOR(
FloatToRowwiseQuantized8Bits,
FloatToRowwiseQuantized8BitsOp<CPUContext>);
REGISTER_CPU_OPERATOR(
SparseLengthsSum8BitsRowwise,
SparseLengths8BitsRowwiseOp<CPUContext>);
REGISTER_CPU_OPERATOR(
SparseLengthsWeightedSum8BitsRowwise,
SparseLengths8BitsRowwiseOp<CPUContext, 1>);
REGISTER_CPU_OPERATOR(
SparseLengthsMean8BitsRowwise,
SparseLengths8BitsRowwiseOp<CPUContext, 0, 1>);
REGISTER_CPU_OPERATOR(
SparseLengthsWeightedMean8BitsRowwise,
SparseLengths8BitsRowwiseOp<CPUContext, 1, 1>);
OPERATOR_SCHEMA(SparseLengthsSum8BitsRowwise)
.NumInputs(4)
.NumOutputs(1)
.ValueLengthInputFillers(
SparseLengths8BitsRowwiseOp<CPUContext>::DATA,
SparseLengths8BitsRowwiseOp<CPUContext>::LENGTHS)
.SetDoc(R"DOC(
Variation of SparseLengthsSum operator, where DATA is
stored using 8bits. DATA was quantized with 8Bit row-wise
quantization (see doc to FloatToRowwiseQuantized8Bits operator). To
restore DATA from 8Bit, we use additional input that stores scales
and biases.
)DOC")
.Input(
0,
"DATA",
"uint8 tensor obtained with "
"operator FloatToRowwiseQuantized8Bits")
.Input(
1,
"INDICES",
"Integer vector containing indices of the first "
"dimension of DATA for the slices that are being aggregated")
.Input(
2,
"LENGTHS",
"Vector with the same sum of elements as the first dimension of DATA")
.Input(
3,
"scale_bias",
"Matrix of floats, each row r_i of which stores a pair "
"s_i, b_i -- scale and bias for i-th row")
.Output(0, "output", "output");
OPERATOR_SCHEMA(SparseLengthsWeightedSum8BitsRowwise)
.NumInputs(5)
.NumOutputs(1)
.ValueLengthInputFillers(
// NOLINTNEXTLINE(modernize-use-bool-literals)
SparseLengths8BitsRowwiseOp<CPUContext, 1>::DATA,
// NOLINTNEXTLINE(modernize-use-bool-literals)
SparseLengths8BitsRowwiseOp<CPUContext, 1>::LENGTHS)
.SetDoc(R"DOC(
Variation of SparseLengthsWeightedSum operator, where
DATA is stored using 8bits. DATA was quantized with 8Bit row-wise
quantization (see doc to FloatToRowwiseQuantized8Bits operator). To
restore DATA from 8Bit, we use additional input that stores scales
and biases.
)DOC")
.Input(
0,
"DATA",
"uint8 tensor obtained with "
"operator FloatToRowwiseQuantized8Bits")
.Input(
1,
"SCALARS",
"Scalar multipliers for the input slices. Must "
"be a vector with the length matching the length of INDICES")
.Input(
2,
"INDICES",
"Integer vector containing indices of the first "
"dimension of DATA for the slices that are being aggregated")
.Input(
3,
"LENGTHS",
"Vector with the same sum of elements as the first dimension of DATA")
.Input(
4,
"scale_bias",
"Matrix of floats, each row r_i of which stores a pair "
"s_i, b_i -- scale and bias for i-th row")
.Output(0, "output", "output");
OPERATOR_SCHEMA(SparseLengthsMean8BitsRowwise)
.NumInputs(4)
.NumOutputs(1)
.ValueLengthInputFillers(
// NOLINTNEXTLINE(modernize-use-bool-literals)
SparseLengths8BitsRowwiseOp<CPUContext, 0, 1>::DATA,
// NOLINTNEXTLINE(modernize-use-bool-literals)
SparseLengths8BitsRowwiseOp<CPUContext, 0, 1>::LENGTHS)
.SetDoc(R"DOC(
Variation of SparseLengthsMean operator, where DATA is
stored using 8bits. DATA was quantized with 8Bit row-wise
quantization (see doc to FloatToRowwiseQuantized8Bits operator). To
restore DATA from 8Bit, we use additional input that stores scales
and biases.
)DOC")
.Input(
0,
"DATA",
"uint8 tensor obtained with "
"operator FloatToRowwiseQuantized8Bits")
.Input(
1,
"INDICES",
"Integer vector containing indices of the first "
"dimension of DATA for the slices that are being aggregated")
.Input(
2,
"LENGTHS",
"Vector with the same sum of elements as the first dimension of DATA")
.Input(
3,
"scale_bias",
"Matrix of floats, each row r_i of which stores a pair "
"s_i, b_i -- scale and bias for i-th row")
.Output(0, "output", "output");
OPERATOR_SCHEMA(SparseLengthsWeightedMean8BitsRowwise)
.NumInputs(5)
.NumOutputs(1)
.ValueLengthInputFillers(
// NOLINTNEXTLINE(modernize-use-bool-literals)
SparseLengths8BitsRowwiseOp<CPUContext, 1, 1>::DATA,
// NOLINTNEXTLINE(modernize-use-bool-literals)
SparseLengths8BitsRowwiseOp<CPUContext, 1, 1>::LENGTHS)
.SetDoc(R"DOC(
Variation of SparseLengthsWeightedMean operator, where
DATA is stored using 8bits. DATA was quantized with 8Bit row-wise
quantization (see doc to FloatToRowwiseQuantized8Bits operator). To
restore DATA from 8Bit, we use additional input that stores scales
and biases.
)DOC")
.Input(
0,
"DATA",
"uint8 tensor obtained with "
"operator FloatToRowwiseQuantized8Bits")
.Input(
1,
"SCALARS",
"Scalar multipliers for the input slices. Must "
"be a vector with the length matching the length of INDICES")
.Input(
2,
"INDICES",
"Integer vector containing indices of the first "
"dimension of DATA for the slices that are being aggregated")
.Input(
3,
"LENGTHS",
"Vector with the same sum of elements as the first dimension of DATA")
.Input(
4,
"scale_bias",
"Matrix of floats, each row r_i of which stores a pair "
"s_i, b_i -- scale and bias for i-th row")
.Output(0, "output", "output");
OPERATOR_SCHEMA(FloatToRowwiseQuantized8Bits)
.NumInputs(1)
.NumOutputs(2)
.ValueLengthInputFillers(
SparseLengths8BitsRowwiseOp<CPUContext>::DATA,
SparseLengths8BitsRowwiseOp<CPUContext>::LENGTHS)
.SetDoc(R"DOC(
This operator applies 8Bit row-wise quantization to
input tensor and returns quantized tensor. Row wise quantization of
input tensor is the following process. We take tensor of size
(m_1, m_2,...,m_n), n >= 2, reshape it into matrix of size
(m_1, m_2 x... x m_n) and apply row-wise quantization. After this,
we compute scale_i= (min_i - max_i) / 255 and bias_i = min_i for
i-th row r_i of reshaped matrix, where min_i and max_i -- minimum
and maximum elements of i-th row, and quantize each element r_{ij} as
0 <= round(r_ij - bias_i) / scale_i) < 256. Instead of input tensor
we obtain uint8 tensor and auxiliary information as scale and bias to
restore input tensor (with losses).
)DOC")
.Input(0, "input", "input")
.Output(0, "quantized_input", "quantized_input")
.Output(
1,
"scale_bias",
"Matrix of floats, each row r_i of which stores a pair "
"s_i, b_i");
OPERATOR_SCHEMA(Rowwise8BitQuantizedToFloat)
.NumInputs(2)
.NumOutputs(1)
.ValueLengthInputFillers(
SparseLengths8BitsRowwiseOp<CPUContext>::DATA,
SparseLengths8BitsRowwiseOp<CPUContext>::LENGTHS)
.SetDoc(R"DOC(
Given uint8 tensor, quantized using 8bit row-wise
quantization, and auxiliary scales and biases, this operator
restores float tensor in the following way. We take input 8bits tensor
of size (m_1, m_2, ..., m_n), n >= 2, reshape it into matrix of size
(m_1, m_2 x... x m_n). We compute element r_{ij} of output matrix as
r_{ij} * s_i + b_i and after this we reshape this output matrix into
output tensor of size (m_1, m_2, ..., m_n).
)DOC")
.Input(0, "quantized_input", "quantized_input")
.Input(
1,
"scale_bias",
"Matrix of floats, each row r_i of which stores a pair "
"s_i, b_i -- scale and bias for i-th row")
.Output(1, "output", "output");
NO_GRADIENT(Rowwise8BitQuantizedToFloat);
NO_GRADIENT(FloatToRowwiseQuantized8Bits);
NO_GRADIENT(SparseLengthsSum8BitsRowwise);
NO_GRADIENT(SparseLengthsWeightedSum8BitsRowwise);
NO_GRADIENT(SparseLengthsMean8BitsRowwise);
NO_GRADIENT(SparseLengthsWeightedMean8BitsRowwise);
}
|
