Home | History | Annotate | Download | only in kernels 
      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 // Implements a quantized version of the Relu6 operation.
     17 #define EIGEN_USE_THREADS
     18 
     19 #include "tensorflow/core/framework/numeric_op.h"
     20 #include "tensorflow/core/framework/op_kernel.h"
     21 #include "tensorflow/core/framework/tensor.h"
     22 #include "tensorflow/core/kernels/meta_support.h"
     23 #include "tensorflow/core/kernels/quantization_utils.h"
     24 #include "tensorflow/core/lib/core/errors.h"
     25 
     26 namespace tensorflow {
     27 
     28 template <typename T>
     29 class QuantizedReluOp : public OpKernel {
     30  public:
     31   explicit QuantizedReluOp(OpKernelConstruction* context) : OpKernel(context) {}
     32 
     33   void Compute(OpKernelContext* context) override {
     34     const Tensor& input = context->input(0);
     35     const float min_input = context->input(1).flat<float>()(0);
     36     const float max_input = context->input(2).flat<float>()(0);
     37     Tensor* output = nullptr;
     38     OP_REQUIRES_OK(context,
     39                    context->allocate_output(0, input.shape(), &output));
     40     const T min_as_quantized = FloatToQuantized<T>(0.0f, min_input, max_input);
     41 
     42     if (meta::IsSupportedAndEnabled() && std::is_same<T, quint8>()) {
     43       auto input_ui8_array = input.flat<quint8>();
     44       meta::Clamp(context, input_ui8_array.data(), input_ui8_array.size(),
     45                   min_as_quantized, 255, output->flat<quint8>().data());
     46     } else {
     47       output->flat<T>().device(context->eigen_cpu_device()) =
     48           input.flat<T>().cwiseMax(min_as_quantized).template cast<T>();
     49     }
     50 
     51     Tensor* output_min = nullptr;
     52     OP_REQUIRES_OK(context, context->allocate_output(1, {}, &output_min));
     53     output_min->flat<float>()(0) = min_input;
     54     Tensor* output_max = nullptr;
     55     OP_REQUIRES_OK(context, context->allocate_output(2, {}, &output_max));
     56     output_max->flat<float>()(0) = max_input;
     57   }
     58 };
     59 
     60 template <typename T>
     61 class QuantizedRelu6Op : public OpKernel {
     62  public:
     63   explicit QuantizedRelu6Op(OpKernelConstruction* context)
     64       : OpKernel(context) {}
     65 
     66   void Compute(OpKernelContext* context) override {
     67     const Tensor& input = context->input(0);
     68     const float min_input = context->input(1).flat<float>()(0);
     69     const float max_input = context->input(2).flat<float>()(0);
     70     Tensor* output = nullptr;
     71     OP_REQUIRES_OK(context,
     72                    context->allocate_output(0, input.shape(), &output));
     73     const T min_as_quantized = FloatToQuantized<T>(0.0f, min_input, max_input);
     74     const T max_as_quantized = FloatToQuantized<T>(6.0f, min_input, max_input);
     75 
     76     if (meta::IsSupportedAndEnabled() && std::is_same<T, quint8>()) {
     77       auto input_ui8_array = input.flat<quint8>();
     78       meta::Clamp(context, input_ui8_array.data(), input_ui8_array.size(),
     79                   min_as_quantized, max_as_quantized,
     80                   output->flat<quint8>().data());
     81     } else {
     82       output->flat<T>().device(context->eigen_cpu_device()) =
     83           input.flat<T>()
     84               .cwiseMax(min_as_quantized)
     85               .cwiseMin(max_as_quantized)
     86               .template cast<T>();
     87     }
     88 
     89     Tensor* output_min = nullptr;
     90     OP_REQUIRES_OK(context, context->allocate_output(1, {}, &output_min));
     91     output_min->flat<float>()(0) = min_input;
     92     Tensor* output_max = nullptr;
     93     OP_REQUIRES_OK(context, context->allocate_output(2, {}, &output_max));
     94     output_max->flat<float>()(0) = max_input;
     95   }
     96 };
     97 
     98 REGISTER_KERNEL_BUILDER(Name("QuantizedRelu")
     99                             .Device(DEVICE_CPU)
    100                             .TypeConstraint<qint32>("Tinput")
    101                             .TypeConstraint<qint32>("out_type"),
    102                         QuantizedReluOp<qint32>);
    103 REGISTER_KERNEL_BUILDER(Name("QuantizedRelu")
    104                             .Device(DEVICE_CPU)
    105                             .TypeConstraint<quint8>("Tinput")
    106                             .TypeConstraint<quint8>("out_type"),
    107                         QuantizedReluOp<quint8>);
    108 
    109 REGISTER_KERNEL_BUILDER(Name("QuantizedRelu6")
    110                             .Device(DEVICE_CPU)
    111                             .TypeConstraint<qint32>("Tinput")
    112                             .TypeConstraint<qint32>("out_type"),
    113                         QuantizedRelu6Op<qint32>);
    114 REGISTER_KERNEL_BUILDER(Name("QuantizedRelu6")
    115                             .Device(DEVICE_CPU)
    116                             .TypeConstraint<quint8>("Tinput")
    117                             .TypeConstraint<quint8>("out_type"),
    118                         QuantizedRelu6Op<quint8>);
    119 }  // namespace tensorflow
    120