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/meta_support.h" 25 #include "tensorflow/core/kernels/quantization_utils.h" 26 #include "tensorflow/core/lib/core/errors.h" 27 28 namespace { 29 enum { 30 QUANTIZE_MODE_MIN_COMBINED, 31 QUANTIZE_MODE_MIN_FIRST, 32 QUANTIZE_MODE_SCALED, 33 }; 34 } // namespace 35 36 namespace tensorflow { 37 38 typedef Eigen::ThreadPoolDevice CPUDevice; 39 40 template <typename Device, typename T> 41 class DequantizeOp : public OpKernel { 42 public: 43 explicit DequantizeOp(OpKernelConstruction* ctx) : OpKernel(ctx) { 44 half_range_ = !std::is_signed<T>::value 45 ? 0.0f 46 : (static_cast<float>(std::numeric_limits<T>::max()) - 47 std::numeric_limits<T>::min() + 1) / 48 2.0f; 49 string mode_string; 50 OP_REQUIRES_OK(ctx, ctx->GetAttr("mode", &mode_string)); 51 OP_REQUIRES(ctx, 52 (mode_string == "MIN_COMBINED" || mode_string == "MIN_FIRST" || 53 mode_string == "SCALED"), 54 errors::InvalidArgument("Mode string must be 'MIN_COMBINED'," 55 " 'MIN_FIRST', or 'SCALED', is '" + 56 mode_string + "'")); 57 if (mode_string == "MIN_COMBINED") { 58 mode_ = QUANTIZE_MODE_MIN_COMBINED; 59 } else if (mode_string == "MIN_FIRST") { 60 mode_ = QUANTIZE_MODE_MIN_FIRST; 61 } else if (mode_string == "SCALED") { 62 mode_ = QUANTIZE_MODE_SCALED; 63 } 64 } 65 66 void Compute(OpKernelContext* ctx) override { 67 const Tensor& input = ctx->input(0); 68 const float min_range = ctx->input(1).flat<float>()(0); 69 const float max_range = ctx->input(2).flat<float>()(0); 70 71 Tensor* output = nullptr; 72 OP_REQUIRES_OK(ctx, ctx->allocate_output(0, input.shape(), &output)); 73 if (mode_ == QUANTIZE_MODE_MIN_COMBINED) { 74 const float scale_factor = 75 (max_range - min_range) / 76 (static_cast<float>(std::numeric_limits<T>::max()) - 77 std::numeric_limits<T>::min()); 78 79 float* out_ptr = output->flat<float>().data(); 80 const T* in_ptr = input.flat<T>().data(); 81 82 const int64 num_elements = input.NumElements(); 83 for (int i = 0; i < num_elements; ++i) { 84 out_ptr[i] = 85 ((static_cast<int>(in_ptr[i]) + half_range_) * scale_factor) + 86 min_range; 87 } 88 } else if (mode_ == QUANTIZE_MODE_MIN_FIRST) { 89 if (meta::IsSupportedAndEnabled() && std::is_same<T, quint8>()) { 90 auto input_ui8_array = input.flat<quint8>(); 91 meta::Dequantize(ctx, input_ui8_array.data(), input_ui8_array.size(), 92 min_range, max_range, output->flat<float>().data()); 93 } else { 94 QuantizedTensorToFloatInPlaceUsingEigen<T>( 95 ctx->template eigen_device<Device>(), input, min_range, max_range, 96 output); 97 } 98 } else if (mode_ == QUANTIZE_MODE_SCALED) { 99 // The quantization logic for mode SCALED matches that of 100 // QuantizeAndDequantizeV2 and QuantizeAndDequantizeV3. 101 static constexpr int num_bits = sizeof(T) * 8; 102 const float max_abs = std::max(std::abs(min_range), std::abs(max_range)); 103 bool is_signed = std::is_signed<T>::value; 104 // If it is signed, we try to keep 0.0 being 0 and drop one bucket. For 105 // example, if it is 8 bits, we have the range [-127, 127]. So for input 106 // range of [-x, x], the scale should be 254/(2*x). 107 // 108 // If it is unsigned and num_bits == 8, the range with 8 bits is [0, 255]. 109 // If the input range is [0, x], then the scale is x/255 instead of 254 as 110 // in the case above. 111 const int target_bits = is_signed ? (num_bits - 1) : num_bits; 112 const float target_range = 113 static_cast<float>((uint64_t{1} << target_bits) - 1); 114 const float scale_factor = max_abs / target_range; 115 float* out_ptr = output->flat<float>().data(); 116 const T* in_ptr = input.flat<T>().data(); 117 118 const int64 num_elements = input.NumElements(); 119 for (int i = 0; i < num_elements; ++i) { 120 out_ptr[i] = static_cast<int>(in_ptr[i]) * scale_factor; 121 } 122 } 123 } 124 125 private: 126 float half_range_; 127 int mode_; 128 }; 129 130 REGISTER_KERNEL_BUILDER( 131 Name("Dequantize").Device(DEVICE_CPU).TypeConstraint<quint8>("T"), 132 DequantizeOp<CPUDevice, quint8>); 133 REGISTER_KERNEL_BUILDER( 134 Name("Dequantize").Device(DEVICE_CPU).TypeConstraint<qint8>("T"), 135 DequantizeOp<CPUDevice, qint8>); 136 REGISTER_KERNEL_BUILDER( 137 Name("Dequantize").Device(DEVICE_CPU).TypeConstraint<quint16>("T"), 138 DequantizeOp<CPUDevice, quint16>); 139 REGISTER_KERNEL_BUILDER( 140 Name("Dequantize").Device(DEVICE_CPU).TypeConstraint<qint16>("T"), 141 DequantizeOp<CPUDevice, qint16>); 142 143 REGISTER_KERNEL_BUILDER( 144 Name("Dequantize").Device(DEVICE_CPU).TypeConstraint<qint32>("T"), 145 DequantizeOp<CPUDevice, qint32>); 146 147 } // namespace tensorflow 148
