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
|
/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
#define EIGEN_USE_THREADS
#if GOOGLE_CUDA
#define EIGEN_USE_GPU
#endif // GOOGLE_CUDA
#include "tensorflow/core/kernels/quantize_and_dequantize_op.h"
#include "tensorflow/core/framework/op.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/framework/type_traits.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/lib/core/errors.h"
namespace tensorflow {
typedef Eigen::ThreadPoolDevice CPUDevice;
typedef Eigen::GpuDevice GPUDevice;
// Simulate quantization precision loss in a float tensor by:
// 1. Quantize the tensor to fixed point numbers, which should match the target
// quantization method when it is used in inference.
// 2. Dequantize it back to floating point numbers for the following ops, most
// likely matmul.
template <typename Device, typename T>
class QuantizeAndDequantizeV2Op : public OpKernel {
public:
explicit QuantizeAndDequantizeV2Op(OpKernelConstruction* ctx)
: OpKernel(ctx) {
OP_REQUIRES_OK(ctx, ctx->GetAttr("signed_input", &signed_input_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("num_bits", &num_bits_));
OP_REQUIRES(ctx, num_bits_ > 0 && num_bits_ < (signed_input_ ? 62 : 63),
errors::InvalidArgument("num_bits is out of range: ", num_bits_,
" with signed_input_ ", signed_input_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("range_given", &range_given_));
}
void Compute(OpKernelContext* ctx) override {
const Tensor& input = ctx->input(0);
Tensor* output = nullptr;
OP_REQUIRES_OK(ctx, ctx->allocate_output(0, input.shape(), &output));
Tensor input_min_tensor;
Tensor input_max_tensor;
if (range_given_) {
input_min_tensor = ctx->input(1);
input_max_tensor = ctx->input(2);
auto min_val = input_min_tensor.scalar<T>()();
auto max_val = input_max_tensor.scalar<T>()();
OP_REQUIRES(ctx, min_val <= max_val,
errors::InvalidArgument("Invalid range: input_min ", min_val,
" > input_max ", max_val));
} else {
OP_REQUIRES_OK(ctx, ctx->allocate_temp(DataTypeToEnum<T>::value,
TensorShape(), &input_min_tensor));
OP_REQUIRES_OK(ctx, ctx->allocate_temp(DataTypeToEnum<T>::value,
TensorShape(), &input_max_tensor));
}
functor::QuantizeAndDequantizeOneScaleFunctor<Device, T> f;
f(ctx->eigen_device<Device>(), input.flat<T>(), signed_input_, num_bits_,
range_given_, &input_min_tensor, &input_max_tensor, output->flat<T>());
}
private:
bool signed_input_;
int num_bits_;
bool range_given_;
};
// Simulate quantization precision loss in a float tensor by:
// 1. Quantize the tensor to fixed point numbers, which should match the target
// quantization method when it is used in inference.
// 2. Dequantize it back to floating point numbers for the following ops, most
// likely matmul.
// Almost identical to QuantizeAndDequantizeV2Op, except that num_bits is a
// tensor.
template <typename Device, typename T>
class QuantizeAndDequantizeV3Op : public OpKernel {
public:
explicit QuantizeAndDequantizeV3Op(OpKernelConstruction* ctx)
: OpKernel(ctx) {
OP_REQUIRES_OK(ctx, ctx->GetAttr("signed_input", &signed_input_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("range_given", &range_given_));
}
void Compute(OpKernelContext* ctx) override {
const Tensor& input = ctx->input(0);
Tensor* output = nullptr;
OP_REQUIRES_OK(ctx, ctx->allocate_output(0, input.shape(), &output));
Tensor num_bits_tensor;
num_bits_tensor = ctx->input(3);
int num_bits_val = num_bits_tensor.scalar<int32>()();
OP_REQUIRES(
ctx, num_bits_val > 0 && num_bits_val < (signed_input_ ? 62 : 63),
errors::InvalidArgument("num_bits is out of range: ", num_bits_val,
" with signed_input_ ", signed_input_));
Tensor input_min_tensor;
Tensor input_max_tensor;
if (range_given_) {
input_min_tensor = ctx->input(1);
input_max_tensor = ctx->input(2);
auto min_val = input_min_tensor.scalar<T>()();
auto max_val = input_max_tensor.scalar<T>()();
OP_REQUIRES(ctx, min_val <= max_val,
errors::InvalidArgument("Invalid range: input_min ", min_val,
" > input_max ", max_val));
} else {
OP_REQUIRES_OK(ctx, ctx->allocate_temp(DataTypeToEnum<T>::value,
TensorShape(), &input_min_tensor));
OP_REQUIRES_OK(ctx, ctx->allocate_temp(DataTypeToEnum<T>::value,
TensorShape(), &input_max_tensor));
}
functor::QuantizeAndDequantizeOneScaleFunctor<Device, T> f;
f(ctx->eigen_device<Device>(), input.flat<T>(), signed_input_, num_bits_val,
range_given_, &input_min_tensor, &input_max_tensor, output->flat<T>());
}
private:
bool signed_input_;
bool range_given_;
};
// DEPRECATED: Use QuantizeAndDequantizeV2Op.
template <typename Device, typename T>
class QuantizeAndDequantizeOp : public OpKernel {
public:
explicit QuantizeAndDequantizeOp(OpKernelConstruction* ctx) : OpKernel(ctx) {
OP_REQUIRES_OK(ctx, ctx->GetAttr("signed_input", &signed_input_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("num_bits", &num_bits_));
OP_REQUIRES(ctx, num_bits_ > 0 && num_bits_ < (signed_input_ ? 62 : 63),
errors::InvalidArgument("num_bits is out of range: ", num_bits_,
" with signed_input_ ", signed_input_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("range_given", &range_given_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("input_min", &input_min_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("input_max", &input_max_));
if (range_given_) {
OP_REQUIRES(
ctx, input_min_ <= input_max_,
errors::InvalidArgument("Invalid range: input_min ", input_min_,
" > input_max ", input_max_));
}
}
void Compute(OpKernelContext* ctx) override {
const Tensor& input = ctx->input(0);
Tensor* output = nullptr;
OP_REQUIRES_OK(ctx, ctx->allocate_output(0, input.shape(), &output));
// One global scale.
Tensor input_min_tensor(DataTypeToEnum<T>::value, TensorShape());
Tensor input_max_tensor(DataTypeToEnum<T>::value, TensorShape());
// Initialize the tensors with the values in the Attrs.
input_min_tensor.template scalar<T>()() = static_cast<T>(input_min_);
input_max_tensor.template scalar<T>()() = static_cast<T>(input_max_);
functor::QuantizeAndDequantizeOneScaleFunctor<Device, T> functor;
functor(ctx->eigen_device<Device>(), input.flat<T>(), signed_input_,
num_bits_, range_given_, &input_min_tensor, &input_max_tensor,
output->flat<T>());
}
private:
bool signed_input_;
int num_bits_;
bool range_given_;
float input_min_;
float input_max_;
};
// Specialization for CPUDevice.
namespace functor {
template <typename T>
struct QuantizeAndDequantizeOneScaleFunctor<CPUDevice, T> {
void operator()(const CPUDevice& d, typename TTypes<T>::ConstVec input,
const bool signed_input, const int num_bits,
const bool range_given, Tensor* input_min_tensor,
Tensor* input_max_tensor, typename TTypes<T>::Vec out) {
QuantizeAndDequantizeOneScaleImpl<CPUDevice, T>::Compute(
d, input, signed_input, num_bits, range_given, input_min_tensor,
input_max_tensor, out);
}
};
} // namespace functor
#define REGISTER_CPU_KERNEL(T) \
REGISTER_KERNEL_BUILDER(Name("QuantizeAndDequantizeV2") \
.Device(DEVICE_CPU) \
.TypeConstraint<T>("T"), \
QuantizeAndDequantizeV2Op<CPUDevice, T>); \
REGISTER_KERNEL_BUILDER(Name("QuantizeAndDequantizeV3") \
.Device(DEVICE_CPU) \
.TypeConstraint<T>("T"), \
QuantizeAndDequantizeV3Op<CPUDevice, T>); \
REGISTER_KERNEL_BUILDER( \
Name("QuantizeAndDequantize").Device(DEVICE_CPU).TypeConstraint<T>("T"), \
QuantizeAndDequantizeOp<CPUDevice, T>);
TF_CALL_float(REGISTER_CPU_KERNEL);
TF_CALL_double(REGISTER_CPU_KERNEL);
#undef REGISTER_CPU_KERNEL
#if GOOGLE_CUDA
#define REGISTER_GPU_KERNEL(T) \
REGISTER_KERNEL_BUILDER(Name("QuantizeAndDequantizeV2") \
.Device(DEVICE_GPU) \
.HostMemory("input_max") \
.HostMemory("input_min") \
.TypeConstraint<T>("T"), \
QuantizeAndDequantizeV2Op<GPUDevice, T>); \
REGISTER_KERNEL_BUILDER(Name("QuantizeAndDequantizeV3") \
.Device(DEVICE_GPU) \
.HostMemory("input_max") \
.HostMemory("input_min") \
.HostMemory("num_bits") \
.TypeConstraint<T>("T"), \
QuantizeAndDequantizeV3Op<GPUDevice, T>); \
REGISTER_KERNEL_BUILDER( \
Name("QuantizeAndDequantize").Device(DEVICE_GPU).TypeConstraint<T>("T"), \
QuantizeAndDequantizeOp<GPUDevice, T>);
TF_CALL_float(REGISTER_GPU_KERNEL);
TF_CALL_double(REGISTER_GPU_KERNEL);
#undef REGISTER_GPU_KERNEL
#endif
} // namespace tensorflow
|
