aboutsummaryrefslogtreecommitdiffhomepage
path: root/tensorflow/core/kernels/quantized_activation_ops.cc
blob: 2896c3d45a7023d5cab842e1aee237f545b86a48 (plain)
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
/* 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.
==============================================================================*/

// Implements a quantized version of the Relu6 operation.
#define EIGEN_USE_THREADS

#include "tensorflow/core/framework/numeric_op.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/kernels/meta_support.h"
#include "tensorflow/core/kernels/quantization_utils.h"
#include "tensorflow/core/lib/core/errors.h"

namespace tensorflow {

template <typename T>
class QuantizedReluOp : public OpKernel {
 public:
  explicit QuantizedReluOp(OpKernelConstruction* context) : OpKernel(context) {}

  void Compute(OpKernelContext* context) override {
    const Tensor& input = context->input(0);
    const float min_input = context->input(1).flat<float>()(0);
    const float max_input = context->input(2).flat<float>()(0);
    Tensor* output = nullptr;
    OP_REQUIRES_OK(context,
                   context->allocate_output(0, input.shape(), &output));
    const T min_as_quantized = FloatToQuantized<T>(0.0f, min_input, max_input);

    if (meta::IsSupportedAndEnabled() && std::is_same<T, quint8>()) {
      auto input_ui8_array = input.flat<quint8>();
      meta::Clamp(context, input_ui8_array.data(), input_ui8_array.size(),
                  min_as_quantized, 255, output->flat<quint8>().data());
    } else {
      output->flat<T>().device(context->eigen_cpu_device()) =
          input.flat<T>().cwiseMax(min_as_quantized).template cast<T>();
    }

    Tensor* output_min = nullptr;
    OP_REQUIRES_OK(context, context->allocate_output(1, {}, &output_min));
    output_min->flat<float>()(0) = min_input;
    Tensor* output_max = nullptr;
    OP_REQUIRES_OK(context, context->allocate_output(2, {}, &output_max));
    output_max->flat<float>()(0) = max_input;
  }
};

template <typename T>
class QuantizedRelu6Op : public OpKernel {
 public:
  explicit QuantizedRelu6Op(OpKernelConstruction* context)
      : OpKernel(context) {}

  void Compute(OpKernelContext* context) override {
    const Tensor& input = context->input(0);
    const float min_input = context->input(1).flat<float>()(0);
    const float max_input = context->input(2).flat<float>()(0);
    Tensor* output = nullptr;
    OP_REQUIRES_OK(context,
                   context->allocate_output(0, input.shape(), &output));
    const T min_as_quantized = FloatToQuantized<T>(0.0f, min_input, max_input);
    const T max_as_quantized = FloatToQuantized<T>(6.0f, min_input, max_input);

    if (meta::IsSupportedAndEnabled() && std::is_same<T, quint8>()) {
      auto input_ui8_array = input.flat<quint8>();
      meta::Clamp(context, input_ui8_array.data(), input_ui8_array.size(),
                  min_as_quantized, max_as_quantized,
                  output->flat<quint8>().data());
    } else {
      output->flat<T>().device(context->eigen_cpu_device()) =
          input.flat<T>()
              .cwiseMax(min_as_quantized)
              .cwiseMin(max_as_quantized)
              .template cast<T>();
    }

    Tensor* output_min = nullptr;
    OP_REQUIRES_OK(context, context->allocate_output(1, {}, &output_min));
    output_min->flat<float>()(0) = min_input;
    Tensor* output_max = nullptr;
    OP_REQUIRES_OK(context, context->allocate_output(2, {}, &output_max));
    output_max->flat<float>()(0) = max_input;
  }
};

REGISTER_KERNEL_BUILDER(Name("QuantizedRelu")
                            .Device(DEVICE_CPU)
                            .TypeConstraint<qint32>("Tinput")
                            .TypeConstraint<qint32>("out_type"),
                        QuantizedReluOp<qint32>);
REGISTER_KERNEL_BUILDER(Name("QuantizedRelu")
                            .Device(DEVICE_CPU)
                            .TypeConstraint<quint8>("Tinput")
                            .TypeConstraint<quint8>("out_type"),
                        QuantizedReluOp<quint8>);

REGISTER_KERNEL_BUILDER(Name("QuantizedRelu6")
                            .Device(DEVICE_CPU)
                            .TypeConstraint<qint32>("Tinput")
                            .TypeConstraint<qint32>("out_type"),
                        QuantizedRelu6Op<qint32>);
REGISTER_KERNEL_BUILDER(Name("QuantizedRelu6")
                            .Device(DEVICE_CPU)
                            .TypeConstraint<quint8>("Tinput")
                            .TypeConstraint<quint8>("out_type"),
                        QuantizedRelu6Op<quint8>);
}  // namespace tensorflow