• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
2 
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6 
7     http://www.apache.org/licenses/LICENSE-2.0
8 
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15 
16 // See docs in ../ops/math_ops.cc.
17 
18 #define EIGEN_USE_THREADS
19 
20 #include "tensorflow/core/framework/op.h"
21 #include "tensorflow/core/framework/op_kernel.h"
22 #include "tensorflow/core/framework/type_traits.h"
23 #include "tensorflow/core/framework/types.h"
24 #include "tensorflow/core/kernels/cwise_ops.h"
25 #include "tensorflow/core/kernels/meta_support.h"
26 #include "tensorflow/core/kernels/quantization_utils.h"
27 #include "tensorflow/core/lib/core/errors.h"
28 
29 namespace {
30 enum {
31   QUANTIZE_MODE_MIN_COMBINED,
32   QUANTIZE_MODE_MIN_FIRST,
33   QUANTIZE_MODE_SCALED,
34 };
35 enum {
36   // Round half away from zero: if the fraction of y is exactly 0.5, then
37   // round(y) = y + 0.5 if y > 0
38   // round(y) = y - 0.5 if y < 0
39   // E.g., -5.5 gets rounded to -6, -5.4 goes to -5,
40   // 5.4 goes to 5, and 5.5 goes to 6.
41   ROUND_HALF_AWAY_FROM_ZERO,
42   // Round half to even: if the fraction of y is exactly 0.5, then round(y) is
43   // the nearest even integer to y.
44   // E.g., 23.5 gets rounded to 24, 24.5 gets rounded to 24, while -23.5 becomes
45   // -24, and -24.5 gets rounded to 24.
46   ROUND_HALF_TO_EVEN,
47 };
48 }  // namespace
49 
50 namespace tensorflow {
51 
52 typedef Eigen::ThreadPoolDevice CPUDevice;
53 
54 // Quantize a tensor from float to T, with user-specified min_range and
55 // max_range.
56 // TODO(xbing): Add a new QuantizeOp just taking scale,
57 //              rather than min_range and max_range.
58 template <typename Device, typename T>
59 class QuantizeV2Op : public OpKernel {
60  public:
QuantizeV2Op(OpKernelConstruction * ctx)61   explicit QuantizeV2Op(OpKernelConstruction* ctx) : OpKernel(ctx) {
62     half_range_ =
63         !std::is_signed<T>::value
64             ? 0.0f
65             : (static_cast<double>(std::numeric_limits<T>::max()) -
66                static_cast<double>(std::numeric_limits<T>::min()) + 1) /
67                   2.0f;
68     string mode_string;
69     OP_REQUIRES_OK(ctx, ctx->GetAttr("mode", &mode_string));
70     OP_REQUIRES(ctx,
71                 (mode_string == "MIN_COMBINED" || mode_string == "MIN_FIRST" ||
72                  mode_string == "SCALED"),
73                 errors::InvalidArgument("Mode string must be 'MIN_COMBINED',"
74                                         " 'MIN_FIRST', or 'SCALED', is '" +
75                                         mode_string + "'"));
76     if (mode_string == "MIN_COMBINED") {
77       mode_ = QUANTIZE_MODE_MIN_COMBINED;
78     } else if (mode_string == "MIN_FIRST") {
79       mode_ = QUANTIZE_MODE_MIN_FIRST;
80     } else if (mode_string == "SCALED") {
81       mode_ = QUANTIZE_MODE_SCALED;
82     }
83 
84     string round_mode_string;
85     OP_REQUIRES_OK(ctx, ctx->GetAttr("round_mode", &round_mode_string));
86     OP_REQUIRES(ctx,
87                 (round_mode_string == "HALF_AWAY_FROM_ZERO" ||
88                  round_mode_string == "HALF_TO_EVEN"),
89                 errors::InvalidArgument("Round mode string must be "
90                                         "'HALF_AWAY_FROM_ZERO' or "
91                                         "'HALF_TO_EVEN', is '" +
92                                         round_mode_string + "'"));
93     if (round_mode_string == "HALF_AWAY_FROM_ZERO") {
94       round_mode_ = ROUND_HALF_AWAY_FROM_ZERO;
95     } else if (round_mode_string == "HALF_TO_EVEN") {
96       OP_REQUIRES(ctx, mode_string == "SCALED",
97                   errors::InvalidArgument("Round mode 'HALF_TO_EVEN' "
98                                           "only supported for mode 'SCALED', "
99                                           "but mode is '" +
100                                           mode_string + "'."));
101       round_mode_ = ROUND_HALF_TO_EVEN;
102     }
103   }
104 
Compute(OpKernelContext * ctx)105   void Compute(OpKernelContext* ctx) override {
106     const Tensor& input = ctx->input(0);
107     const float input_min_range = ctx->input(1).flat<float>()(0);
108     const float input_max_range = ctx->input(2).flat<float>()(0);
109 
110     float min_range;
111     float max_range;
112     OP_REQUIRES(ctx, !(input_max_range < input_min_range),
113                 errors::InvalidArgument(
114                     "input_max_range must be larger than input_min_range."));
115 
116     // When the minimum and maximum ranges are too close together, nudge them
117     // apart by a small value so that they are slightly different. This helps
118     // us avoid creating ill-formed buffers where all quantized values map to
119     // the same float number. These kinds of buffers cause problems for
120     // downstream ops when they need to do calculations on them.
121     // We pick the value by making sure that zero is not more than 100x the
122     // overall range from the maximum, so that the value can be easily
123     // represented when we promote the quantized value to a higher
124     // intermediate bit depth, since that's a common requirement.
125     min_range = std::min(0.0f, input_min_range);
126     const float epsilon = std::max(1.0f, std::max(fabsf(input_min_range),
127                                                   fabsf(input_max_range))) /
128                           100.0f;
129     max_range = std::max(input_max_range, min_range + epsilon);
130     max_range = std::max(0.0f, max_range);
131 
132     Tensor* output = nullptr;
133     OP_REQUIRES_OK(ctx, ctx->allocate_output(0, input.shape(), &output));
134     typename TTypes<T>::Vec o = output->template flat<T>();
135     if (mode_ == QUANTIZE_MODE_MIN_COMBINED) {
136       const float scale_factor =
137           (static_cast<double>(std::numeric_limits<T>::max()) -
138            static_cast<double>(std::numeric_limits<T>::min())) /
139           (max_range - min_range);
140 
141       // Quantize:
142       // Make input in range of [min_range, max_range], then
143       // subtract min_range to be in range of [0, max_range - min_range]
144       // Divide by (max_range - min_range) to get to [0, 1.0]
145       // Multiply by range of T, after that shift left 1/2 range of T if
146       // T is signed.
147       // Note that the number is rounded before the cast. Rounding follows the
148       // semantic of std::round, which implements "round-half-away-zero",
149       // e.g., -5.5 gets rounded to -6, -5.4 goes to -5, 5.4 goes to 5,
150       // and 5.5 goes to 6.
151       bool is_signed = std::is_signed<T>::value;
152       if (is_signed) {
153         // The slow path.
154         // TODO(xbing,yonghui): Speedup this path as well.
155         o.device(ctx->template eigen_device<Device>()) =
156             ((input.flat<float>().cwiseMin(max_range).cwiseMax(min_range) -
157               min_range) *
158                  scale_factor -
159              half_range_)
160                 .round()
161                 .template cast<T>();
162       } else {
163         // The fast path that avoids unaryExpr
164         // According to the micro-benchmark, adding device here doesn't help.
165         o = ((input.flat<float>().cwiseMin(max_range).cwiseMax(min_range) -
166               min_range) *
167                  scale_factor +
168              0.5f)
169                 .template cast<T>();
170       }
171     } else if (mode_ == QUANTIZE_MODE_MIN_FIRST) {
172       if (meta::IsSupportedAndEnabled() && std::is_same<T, quint8>()) {
173         TTypes<const float>::Vec input_array = input.flat<float>();
174 
175         meta::Quantize(ctx, input_array.data(), input_array.size(), min_range,
176                        max_range, output->flat<quint8>().data());
177       } else {
178         FloatTensorToQuantizedInPlaceUsingEigen<T>(
179             ctx->template eigen_device<Device>(), input, min_range, max_range,
180             output);
181       }
182     } else if (mode_ == QUANTIZE_MODE_SCALED) {
183       const int min_output_value = std::numeric_limits<T>::min();
184       const int max_output_value = std::numeric_limits<T>::max();
185       const float scale_factor_from_min_side =
186           (min_output_value * min_range > 0)
187               ? min_output_value / min_range
188               : std::numeric_limits<float>::max();
189       const float scale_factor_from_max_side =
190           (max_output_value * max_range > 0)
191               ? max_output_value / max_range
192               : std::numeric_limits<float>::max();
193       const float scale_factor =
194           std::min(scale_factor_from_min_side, scale_factor_from_max_side);
195       min_range = min_output_value / scale_factor;
196       max_range = max_output_value / scale_factor;
197       if (round_mode_ == ROUND_HALF_TO_EVEN) {
198         // scalar_round_op_google implements "round-half-to-even".
199         o.device(ctx->template eigen_device<Device>()) =
200             (input.flat<float>().cwiseMin(max_range).cwiseMax(min_range) *
201              scale_factor)
202                 .unaryExpr(Eigen::internal::scalar_round_op_google<float>())
203                 .template cast<T>();
204       } else if (round_mode_ == ROUND_HALF_AWAY_FROM_ZERO) {
205         // scalar_round_op implements "round-half-away-from-zero".
206         o.device(ctx->template eigen_device<Device>()) =
207             (input.flat<float>().cwiseMin(max_range).cwiseMax(min_range) *
208              scale_factor)
209                 .unaryExpr(Eigen::internal::scalar_round_op<float>())
210                 .template cast<T>();
211       }
212     }
213 
214     Tensor* output_min_tensor = nullptr;
215     OP_REQUIRES_OK(ctx, ctx->allocate_output(1, {}, &output_min_tensor));
216     output_min_tensor->flat<float>()(0) = min_range;
217 
218     Tensor* output_max_tensor = nullptr;
219     OP_REQUIRES_OK(ctx, ctx->allocate_output(2, {}, &output_max_tensor));
220     output_max_tensor->flat<float>()(0) = max_range;
221   }
222 
223  private:
224   float half_range_;
225   int mode_;
226   int round_mode_;
227 };
228 
229 REGISTER_KERNEL_BUILDER(
230     Name("QuantizeV2").Device(DEVICE_CPU).TypeConstraint<quint8>("T"),
231     QuantizeV2Op<CPUDevice, quint8>);
232 REGISTER_KERNEL_BUILDER(
233     Name("QuantizeV2").Device(DEVICE_CPU).TypeConstraint<qint8>("T"),
234     QuantizeV2Op<CPUDevice, qint8>);
235 REGISTER_KERNEL_BUILDER(
236     Name("QuantizeV2").Device(DEVICE_CPU).TypeConstraint<quint16>("T"),
237     QuantizeV2Op<CPUDevice, quint16>);
238 REGISTER_KERNEL_BUILDER(
239     Name("QuantizeV2").Device(DEVICE_CPU).TypeConstraint<qint16>("T"),
240     QuantizeV2Op<CPUDevice, qint16>);
241 REGISTER_KERNEL_BUILDER(
242     Name("QuantizeV2").Device(DEVICE_CPU).TypeConstraint<qint32>("T"),
243     QuantizeV2Op<CPUDevice, qint32>);
244 }  // namespace tensorflow
245