forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
fused_rowwise_nbit_conversion_ops.cc
303 lines (288 loc) · 11.4 KB
/
fused_rowwise_nbit_conversion_ops.cc
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
302
303
#include "caffe2/operators/fused_rowwise_nbit_conversion_ops.h"
#include <fp16.h>
#include "c10/util/Registry.h"
namespace caffe2 {
using std::uint16_t;
using std::vector;
namespace internal {
void convertfp32fp16(at::Half* dst, const float* src, size_t N) {
for (size_t i = 0; i < N; i++) {
dst[i] = src[i];
}
}
} // namespace internal
REGISTER_CPU_OPERATOR(
FloatToFused4BitRowwiseQuantized,
FloatToFusedNBitRowwiseQuantizedOp<4, float, internal::convertfp32fp32>);
OPERATOR_SCHEMA(FloatToFused4BitRowwiseQuantized)
.NumInputs(1)
.NumOutputs(1)
.TensorInferenceFunction([](const OperatorDef& /* def */,
const vector<TensorShape>& in) {
vector<TensorShape> out;
TensorShape X = in[0];
// divide over 2 and round up, add 4 for the extra scale and bias
X.set_dims(
X.dims().size() - 1,
(X.dims(X.dims().size() - 1) + 1) / 2 + 2 * sizeof(at::Half));
out.push_back(std::move(X));
out[0].set_data_type(TensorProto_DataType_UINT8);
return out;
})
.SetDoc(R"DOC(
Applies 4-bit row-wise quantization by determining the range
(maximum - minimum) and offset (minimum value) of each row in the input
matrix, and then scaling each element to an 4-bit number between 0 and
15. To later de-quantize values, the scale (range / 15) and zero_point
are stored alongside the data. More precisely, each row first has quantized
values, and then 2-byte fp16 scale and 2-byte zero_offset.)
)DOC")
.Input(0, "input", "Float32 input data")
.Output(0, "output", "Fused scale, bias and quantized data");
NO_GRADIENT(FloatToFused4BitRowwiseQuantized);
REGISTER_CPU_OPERATOR(
HalfToFused4BitRowwiseQuantized,
FloatToFusedNBitRowwiseQuantizedOp<4, at::Half, internal::convertfp16fp32>);
OPERATOR_SCHEMA(HalfToFused4BitRowwiseQuantized)
.NumInputs(1)
.NumOutputs(1)
.TensorInferenceFunction([](const OperatorDef& /* def */,
const vector<TensorShape>& in) {
vector<TensorShape> out;
TensorShape X = in[0];
X.set_dims(
X.dims().size() - 1,
(X.dims(X.dims().size() - 1) + 1) / 2 + 2 * sizeof(at::Half));
out.push_back(std::move(X));
out[0].set_data_type(TensorProto_DataType_UINT8);
return out;
})
.SetDoc(R"DOC(
Applies 4-bit row-wise quantization by determining the range
(maximum - minimum) and offset (minimum value) of each row in the input
matrix, and then scaling each element to an 4-bit number between 0 and
15. To later de-quantize values, the scale (range / 15) and zero_point
are stored alongside the data. More precisely, each row first has quantized
values, and then 2-byte fp16 scale and 2-byte zero_offset.)
)DOC")
.Input(0, "input", "Float16 input data")
.Output(0, "output", "Fused scale, bias and quantized data");
NO_GRADIENT(HalfToFused4BitRowwiseQuantized);
REGISTER_CPU_OPERATOR(
Fused4BitRowwiseQuantizedToFloat,
FusedNBitRowwiseQuantizedToFloatOp<4, float, internal::convertfp32fp32>);
OPERATOR_SCHEMA(Fused4BitRowwiseQuantizedToFloat)
.NumInputs(1)
.NumOutputs(1)
.TensorInferenceFunction([](const OperatorDef& /* def */,
const vector<TensorShape>& in) {
vector<TensorShape> out;
TensorShape X = in[0];
X.set_dims(
X.dims().size() - 1,
(X.dims(X.dims().size() - 1) - 2 * sizeof(at::Half)) * 2);
out.push_back(std::move(X));
out[0].set_data_type(TensorProto_DataType_FLOAT);
return out;
})
.SetDoc(R"DOC(
De-quantizes the result of the
FloatToFused4BitRowwiseQuantized operator. The input is expected to first have
quantized values, then 2-byte fp16 scale and 1-byte zero_offset. The output is a
matrix containing only the values, but de-quantized. De-quantization is
performed by multiplying each value by its row's scale and zero_point
parameters. The de-quantized values will thus not be exactly equal to
the original, un-quantized floating point values.
)DOC")
.Input(
0,
"scale_bias_quantized_input",
"Fused scale, bias and quantized data")
.Output(0, "float_output", "Float32 data");
NO_GRADIENT(Fused4BitRowwiseQuantizedToFloat);
REGISTER_CPU_OPERATOR(
Fused4BitRowwiseQuantizedToHalf,
FusedNBitRowwiseQuantizedToFloatOp<4, at::Half, internal::convertfp32fp16>);
OPERATOR_SCHEMA(Fused4BitRowwiseQuantizedToHalf)
.NumInputs(1)
.NumOutputs(1)
.TensorInferenceFunction([](const OperatorDef& /* def */,
const vector<TensorShape>& in) {
vector<TensorShape> out;
TensorShape X = in[0];
X.set_dims(
X.dims().size() - 1,
(X.dims(X.dims().size() - 1) - 2 * sizeof(at::Half)) * 2);
out.push_back(std::move(X));
out[0].set_data_type(TensorProto_DataType_FLOAT16);
return out;
})
.SetDoc(R"DOC(
De-quantizes the result of the
FloatToFused4BitRowwiseQuantized operator. The input is expected to first have
quantized values, then 2-byte fp16 scale and 1-byte zero_offset. The output is a
matrix containing only the values, but de-quantized. De-quantization is
performed by multiplying each value by its row's scale and zero_point
parameters. The de-quantized values will thus not be exactly equal to
the original, un-quantized floating point values.
)DOC")
.Input(
0,
"scale_bias_quantized_input",
"Fused scale, bias and quantized data")
.Output(0, "float16_output", "Float16 data");
NO_GRADIENT(Fused4BitRowwiseQuantizedToHalf);
REGISTER_CPU_OPERATOR_WITH_ENGINE(
FloatToFused4BitRowwiseQuantized,
GREEDY,
FloatToFusedNBitRowwiseQuantizedOp<
4,
float,
internal::convertfp32fp32,
true /*GREEDY*/>);
REGISTER_CPU_OPERATOR_WITH_ENGINE(
HalfToFused4BitRowwiseQuantized,
GREEDY,
FloatToFusedNBitRowwiseQuantizedOp<
4,
at::Half,
internal::convertfp16fp32,
true /*GREEDY*/>);
REGISTER_CPU_OPERATOR(
FloatToFused2BitRowwiseQuantized,
FloatToFusedNBitRowwiseQuantizedOp<2, float, internal::convertfp32fp32>);
OPERATOR_SCHEMA(FloatToFused2BitRowwiseQuantized)
.NumInputs(1)
.NumOutputs(1)
.TensorInferenceFunction([](const OperatorDef& /* def */,
const vector<TensorShape>& in) {
vector<TensorShape> out;
TensorShape X = in[0];
// divide over 4 and round up, add 4 for the extra scale and bias
X.set_dims(
X.dims().size() - 1,
(X.dims(X.dims().size() - 1) + 3) / 4 + 2 * sizeof(at::Half));
out.push_back(std::move(X));
out[0].set_data_type(TensorProto_DataType_UINT8);
return out;
})
.SetDoc(R"DOC(
Applies 2-bit row-wise quantization by determining the range
(maximum - minimum) and offset (minimum value) of each row in the input
matrix, and then scaling each element to an 2-bit number between 0 and
3. To later de-quantize values, the scale (range / 3) and zero_point
are stored alongside the data. More precisely, each row first has quantized
values, and then 2-byte fp16 scale and 2-byte zero_offset.)
)DOC")
.Input(0, "input", "Float32 input data")
.Output(0, "output", "Fused scale, bias and quantized data");
NO_GRADIENT(FloatToFused2BitRowwiseQuantized);
REGISTER_CPU_OPERATOR(
HalfToFused2BitRowwiseQuantized,
FloatToFusedNBitRowwiseQuantizedOp<2, at::Half, internal::convertfp16fp32>);
OPERATOR_SCHEMA(HalfToFused2BitRowwiseQuantized)
.NumInputs(1)
.NumOutputs(1)
.TensorInferenceFunction([](const OperatorDef& /* def */,
const vector<TensorShape>& in) {
vector<TensorShape> out;
TensorShape X = in[0];
X.set_dims(
X.dims().size() - 1,
(X.dims(X.dims().size() - 1) + 3) / 4 + 2 * sizeof(at::Half));
out.push_back(std::move(X));
out[0].set_data_type(TensorProto_DataType_UINT8);
return out;
})
.SetDoc(R"DOC(
Applies 2-bit row-wise quantization by determining the range
(maximum - minimum) and offset (minimum value) of each row in the input
matrix, and then scaling each element to an 2-bit number between 0 and
3. To later de-quantize values, the scale (range / 3) and zero_point
are stored alongside the data. More precisely, each row first has quantized
values, and then 2-byte fp16 scale and 2-byte zero_offset.)
)DOC")
.Input(0, "input", "Float16 input data")
.Output(0, "output", "Fused scale, bias and quantized data");
NO_GRADIENT(HalfToFused2BitRowwiseQuantized);
REGISTER_CPU_OPERATOR(
Fused2BitRowwiseQuantizedToFloat,
FusedNBitRowwiseQuantizedToFloatOp<2, float, internal::convertfp32fp32>);
OPERATOR_SCHEMA(Fused2BitRowwiseQuantizedToFloat)
.NumInputs(1)
.NumOutputs(1)
.TensorInferenceFunction([](const OperatorDef& /* def */,
const vector<TensorShape>& in) {
vector<TensorShape> out;
TensorShape X = in[0];
X.set_dims(
X.dims().size() - 1,
(X.dims(X.dims().size() - 1) - 2 * sizeof(at::Half)) * 4);
out.push_back(std::move(X));
out[0].set_data_type(TensorProto_DataType_FLOAT);
return out;
})
.SetDoc(R"DOC(
De-quantizes the result of the
FloatToFused2BitRowwiseQuantized operator. The input is expected to first have
quantized values, then 2-byte fp16 scale and 1-byte zero_offset. The output is a
matrix containing only the values, but de-quantized. De-quantization is
performed by multiplying each value by its row's scale and zero_point
parameters. The de-quantized values will thus not be exactly equal to
the original, un-quantized floating point values.
)DOC")
.Input(
0,
"scale_bias_quantized_input",
"Fused scale, bias and quantized data")
.Output(0, "float_output", "Float32 data");
NO_GRADIENT(Fused2BitRowwiseQuantizedToFloat);
REGISTER_CPU_OPERATOR(
Fused2BitRowwiseQuantizedToHalf,
FusedNBitRowwiseQuantizedToFloatOp<2, at::Half, internal::convertfp32fp16>);
OPERATOR_SCHEMA(Fused2BitRowwiseQuantizedToHalf)
.NumInputs(1)
.NumOutputs(1)
.TensorInferenceFunction([](const OperatorDef& /* def */,
const vector<TensorShape>& in) {
vector<TensorShape> out;
TensorShape X = in[0];
X.set_dims(
X.dims().size() - 1,
(X.dims(X.dims().size() - 1) - 2 * sizeof(at::Half)) * 4);
out.push_back(std::move(X));
out[0].set_data_type(TensorProto_DataType_FLOAT16);
return out;
})
.SetDoc(R"DOC(
De-quantizes the result of the
FloatToFused2BitRowwiseQuantized operator. The input is expected to first have
quantized values, then 2-byte fp16 scale and 1-byte zero_offset. The output is a
matrix containing only the values, but de-quantized. De-quantization is
performed by multiplying each value by its row's scale and zero_point
parameters. The de-quantized values will thus not be exactly equal to
the original, un-quantized floating point values.
)DOC")
.Input(
0,
"scale_bias_quantized_input",
"Fused scale, bias and quantized data")
.Output(0, "float16_output", "Float16 data");
NO_GRADIENT(Fused2BitRowwiseQuantizedToHalf);
REGISTER_CPU_OPERATOR_WITH_ENGINE(
FloatToFused2BitRowwiseQuantized,
GREEDY,
FloatToFusedNBitRowwiseQuantizedOp<
2,
float,
internal::convertfp32fp32,
true /*GREEDY*/>);
REGISTER_CPU_OPERATOR_WITH_ENGINE(
HalfToFused2BitRowwiseQuantized,
GREEDY,
FloatToFusedNBitRowwiseQuantizedOp<
2,
at::Half,
internal::convertfp16fp32,
true /*GREEDY*/>);
} // namespace caffe2