tensorflow/lite/kernels/sub.cc - platform/external/tensorflow - Git at Google

 /* Copyright 2017 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.
 ==============================================================================*/
 #include <limits>

 #include "tensorflow/lite/c/builtin_op_data.h"
 #include "tensorflow/lite/c/common.h"
 #include "tensorflow/lite/kernels/internal/optimized/cpu_check.h"
 #include "tensorflow/lite/kernels/internal/optimized/optimized_ops.h"
 #include "tensorflow/lite/kernels/internal/quantization_util.h"
 #include "tensorflow/lite/kernels/internal/reference/integer_ops/add.h"
 #include "tensorflow/lite/kernels/internal/reference/reference_ops.h"
 #include "tensorflow/lite/kernels/internal/tensor.h"
 #include "tensorflow/lite/kernels/kernel_util.h"
 #include "tensorflow/lite/kernels/op_macros.h"

 namespace tflite {
 namespace ops {
 namespace builtin {
 namespace sub {

 // This file has three implementation of Sub.
 enum KernelType {
   kReference,
   kGenericOptimized,  // Neon-free
   kNeonOptimized,
 };

 constexpr int kInputTensor1 = 0;
 constexpr int kInputTensor2 = 1;
 constexpr int kOutputTensor = 0;

 struct OpData {
   bool requires_broadcast;

   // These fields are used in both the general 8-bit -> 8bit quantized path,
   // and the special 16-bit -> 16bit quantized path
   int input1_shift;
   int input2_shift;
   int32 output_activation_min;
   int32 output_activation_max;

   // These fields are used only in the general 8-bit -> 8bit quantized path
   int32 input1_multiplier;
   int32 input2_multiplier;
   int32 output_multiplier;
   int output_shift;
   int left_shift;
   int32 input1_offset;
   int32 input2_offset;
   int32 output_offset;

   // This parameter is used to indicate whether
   // parameter scale is power of two.
   // It is used in 16-bit -> 16-bit quantization.
   bool pot_scale_16bit;
 };

 void* Init(TfLiteContext* context, const char* buffer, size_t length) {
   auto* data = new OpData;
   data->requires_broadcast = false;
   return data;
 }

 void Free(TfLiteContext* context, void* buffer) {
   delete reinterpret_cast<OpData*>(buffer);
 }

 TfLiteStatus PrepareGeneralSubOp(TfLiteContext* context,
                                  const TfLiteTensor* input_1,
                                  const TfLiteTensor* input_2,
                                  TfLiteTensor* output, TfLiteSubParams* params,
                                  OpData* op_params, int op_sign) {
   TF_LITE_ENSURE(context, output->type == kTfLiteUInt8 ||
                               output->type == kTfLiteInt8 ||
                               output->type == kTfLiteInt16);
   const auto& input1_quantization_params = input_1->params;
   const auto& input2_quantization_params = input_2->params;
   const auto& output_quantization_params = output->params;
   int32_t integer_type_min = 0;
   int32_t integer_type_max = 0;
   if (output->type == kTfLiteUInt8) {
     integer_type_min = std::numeric_limits<uint8_t>::min();
     integer_type_max = std::numeric_limits<uint8_t>::max();
   } else if (output->type == kTfLiteInt16) {
     integer_type_min = std::numeric_limits<int16_t>::min();
     integer_type_max = std::numeric_limits<int16_t>::max();
   } else {
     // output->type == kTfLiteInt8
     integer_type_min = std::numeric_limits<int8_t>::min();
     integer_type_max = std::numeric_limits<int8_t>::max();
   }

   TF_LITE_ENSURE(context,
                  input1_quantization_params.zero_point >= integer_type_min);
   TF_LITE_ENSURE(context,
                  input1_quantization_params.zero_point <= integer_type_max);
   TF_LITE_ENSURE(context,
                  input2_quantization_params.zero_point >= integer_type_min);
   TF_LITE_ENSURE(context,
                  input2_quantization_params.zero_point <= integer_type_max);
   TF_LITE_ENSURE(context,
                  output_quantization_params.zero_point >= integer_type_min);
   TF_LITE_ENSURE(context,
                  output_quantization_params.zero_point <= integer_type_max);

   op_params->input1_offset = -input1_quantization_params.zero_point;
   op_params->input2_offset = -input2_quantization_params.zero_point;
   op_params->output_offset = output_quantization_params.zero_point;

   // The shift is set to 15 in case of 16-bit and 20 in case of 8-bit,
   // accordingly. In case of 16-bit we have 65535 << 15 which is less than 1 <<
   // 31, therefore the addition will still fit in a 32 bit accumulator.
   op_params->left_shift = output->type == kTfLiteInt16 ? 15 : 20;
   const double twice_max_input_scale =
       2 * std::max(input1_quantization_params.scale,
                    input2_quantization_params.scale);
   const double real_input1_multiplier =
       input1_quantization_params.scale / twice_max_input_scale;
   const double real_input2_multiplier =
       input2_quantization_params.scale / twice_max_input_scale;
   const double real_output_multiplier =
       twice_max_input_scale /
       ((1 << op_params->left_shift) * output_quantization_params.scale);

   tflite::QuantizeMultiplierSmallerThanOneExp(real_input1_multiplier,
                                               &op_params->input1_multiplier,
                                               &op_params->input1_shift);
   tflite::QuantizeMultiplierSmallerThanOneExp(real_input2_multiplier,
                                               &op_params->input2_multiplier,
                                               &op_params->input2_shift);
   op_params->input2_multiplier *= op_sign;
   tflite::QuantizeMultiplierSmallerThanOneExp(real_output_multiplier,
                                               &op_params->output_multiplier,
                                               &op_params->output_shift);

   TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized(
       context, params->activation, output, &op_params->output_activation_min,
       &op_params->output_activation_max));

   return kTfLiteOk;
 }

 TfLiteStatus PrepareInt16SubOpPOT(TfLiteContext* context,
                                   const TfLiteTensor* input1,
                                   const TfLiteTensor* input2,
                                   TfLiteTensor* output, TfLiteSubParams* params,
                                   OpData* data) {
   // 16bit -> 16bit special quantized path, supporting only a rather
   // narrow case of quantization parameters: zero_points must all be 0
   // ("symmetric quantization") and scales must be power-of-two (which
   // we abbreviate as "POT" below). The intended use case for this path
   // is in LSTM cells, where, due to the constraints of implementing
   // some of the math in these LSTM cells in fixed-point arithmetic,
   // we need to have such symmetric, power-of-two quantization
   // (Fixed-point formats are inherently symmetric, power-of-two).
   TF_LITE_ENSURE_EQ(context, input1->params.zero_point, 0);
   TF_LITE_ENSURE_EQ(context, input2->params.zero_point, 0);
   TF_LITE_ENSURE_EQ(context, output->params.zero_point, 0);

   int input1_scale_log2_rounded;
   bool input1_scale_is_pot =
       CheckedLog2(input1->params.scale, &input1_scale_log2_rounded);
   TF_LITE_ENSURE(context, input1_scale_is_pot);

   int input2_scale_log2_rounded;
   bool input2_scale_is_pot =
       CheckedLog2(input2->params.scale, &input2_scale_log2_rounded);
   TF_LITE_ENSURE(context, input2_scale_is_pot);

   int output_scale_log2_rounded;
   bool output_scale_is_pot =
       CheckedLog2(output->params.scale, &output_scale_log2_rounded);
   TF_LITE_ENSURE(context, output_scale_is_pot);

   data->input1_shift = input1_scale_log2_rounded - output_scale_log2_rounded;
   data->input2_shift = input2_scale_log2_rounded - output_scale_log2_rounded;

   // Shifting of one input is supported. The graph quantization should ensure
   // that the other input matches the output.
   TF_LITE_ENSURE(context, data->input1_shift == 0 || data->input2_shift == 0);
   TF_LITE_ENSURE(context, data->input1_shift <= 0);
   TF_LITE_ENSURE(context, data->input2_shift <= 0);

   TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized(
       context, params->activation, output, &data->output_activation_min,
       &data->output_activation_max));
   return kTfLiteOk;
 }

 TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
   OpData* data = reinterpret_cast<OpData*>(node->user_data);
   auto* params = reinterpret_cast<TfLiteSubParams*>(node->builtin_data);

   TF_LITE_ENSURE_EQ(context, NumInputs(node), 2);
   TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1);

   const TfLiteTensor* input1 = GetInput(context, node, kInputTensor1);
   const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2);
   TfLiteTensor* output = GetOutput(context, node, kOutputTensor);

   TF_LITE_ENSURE_EQ(context, input1->type, input2->type);
   output->type = input2->type;

   data->requires_broadcast = !HaveSameShapes(input1, input2);

   TfLiteIntArray* output_size = nullptr;
   if (data->requires_broadcast) {
     TF_LITE_ENSURE_OK(context, CalculateShapeForBroadcast(
                                    context, input1, input2, &output_size));
   } else {
     output_size = TfLiteIntArrayCopy(input1->dims);
   }

   // 8bit -> 8bit general quantized path, with general rescalings
   // as well as, 16bit -> 16bit with general rescalings
   bool pot_scale_16bit = false;

   bool input1_scale_is_pot = false;
   bool input2_scale_is_pot = false;
   bool output_scale_is_pot = false;

   int input1_scale_log2_rounded;
   int input2_scale_log2_rounded;
   int output_scale_log2_rounded;

   if (input1->type == kTfLiteInt16 && input2->type == kTfLiteInt16 &&
       output->type == kTfLiteInt16) {
     // Check that param scale is POT
     input1_scale_is_pot =
         CheckedLog2(input1->params.scale, &input1_scale_log2_rounded);

     input2_scale_is_pot =
         CheckedLog2(input2->params.scale, &input2_scale_log2_rounded);

     output_scale_is_pot =
         CheckedLog2(output->params.scale, &output_scale_log2_rounded);

     pot_scale_16bit = input1_scale_log2_rounded && input2_scale_log2_rounded &&
                       output_scale_log2_rounded;
   }

   data->pot_scale_16bit = pot_scale_16bit;

   if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8 ||
       pot_scale_16bit) {
     TF_LITE_ENSURE_OK(context, PrepareGeneralSubOp(context, input1, input2,
                                                    output, params, data, -1));
   } else if (output->type == kTfLiteInt16) {
     // LSTM-special case with scale parameter of POT
     TF_LITE_ENSURE_OK(context, PrepareInt16SubOpPOT(context, input1, input2,
                                                     output, params, data));
   }

   return context->ResizeTensor(context, output, output_size);
 }

 template <KernelType kernel_type>
 void EvalSub(TfLiteContext* context, TfLiteNode* node, TfLiteSubParams* params,
              const OpData* data, const TfLiteTensor* input1,
              const TfLiteTensor* input2, TfLiteTensor* output) {
 #define TF_LITE_SUB(type, opname, data_type)                             \
   data_type output_activation_min, output_activation_max;                \
   CalculateActivationRange(params->activation, &output_activation_min,   \
                            &output_activation_max);                      \
   tflite::ArithmeticParams op_params;                                    \
   SetActivationParams(output_activation_min, output_activation_max,      \
                       &op_params);                                       \
   type::opname(op_params, GetTensorShape(input1),                        \
                GetTensorData<data_type>(input1), GetTensorShape(input2), \
                GetTensorData<data_type>(input2), GetTensorShape(output), \
                GetTensorData<data_type>(output))
   if (output->type == kTfLiteInt32) {
     if (kernel_type == kReference) {
       if (data->requires_broadcast) {
         TF_LITE_SUB(reference_ops, BroadcastSub4DSlow, int32_t);
       } else {
         TF_LITE_SUB(reference_ops, SubWithActivation, int32_t);
       }
     } else {
       if (data->requires_broadcast) {
         TF_LITE_SUB(optimized_ops, BroadcastSub4DSlow, int32_t);
       } else {
         TF_LITE_SUB(optimized_ops, SubWithActivation, int32_t);
       }
     }
   } else if (output->type == kTfLiteFloat32) {
     if (kernel_type == kReference) {
       if (data->requires_broadcast) {
         TF_LITE_SUB(reference_ops, BroadcastSub4DSlow, float);
       } else {
         TF_LITE_SUB(reference_ops, SubWithActivation, float);
       }
     } else {
       if (data->requires_broadcast) {
         TF_LITE_SUB(optimized_ops, BroadcastSub4DSlow, float);
       } else {
         TF_LITE_SUB(optimized_ops, SubWithActivation, float);
       }
     }
   }
 #undef TF_LITE_SUB
 }

 template <KernelType kernel_type>
 void EvalQuantized(TfLiteContext* context, TfLiteNode* node,
                    TfLiteSubParams* params, const OpData* data,
                    const TfLiteTensor* input1, const TfLiteTensor* input2,
                    TfLiteTensor* output) {
   tflite::ArithmeticParams op_params;
   op_params.left_shift = data->left_shift;
   op_params.input1_offset = data->input1_offset;
   op_params.input1_multiplier = data->input1_multiplier;
   op_params.input1_shift = data->input1_shift;
   op_params.input2_offset = data->input2_offset;
   op_params.input2_multiplier = data->input2_multiplier;
   op_params.input2_shift = data->input2_shift;
   op_params.output_offset = data->output_offset;
   op_params.output_multiplier = data->output_multiplier;
   op_params.output_shift = data->output_shift;
   SetActivationParams(data->output_activation_min, data->output_activation_max,
                       &op_params);

   const bool need_broadcast = optimized_ops::ProcessBroadcastShapes(
       GetTensorShape(input1), GetTensorShape(input2), &op_params);

 #define TF_LITE_SUB(type, opname, data_type)                             \
   type::opname(op_params, GetTensorShape(input1),                        \
                GetTensorData<data_type>(input1), GetTensorShape(input2), \
                GetTensorData<data_type>(input2), GetTensorShape(output), \
                GetTensorData<data_type>(output))
   // NOTE: We are using the add kernels. This is possible as the second values
   // multiplier is negated before being passed down.
   if (output->type == kTfLiteInt8) {
     if (need_broadcast) {
       TF_LITE_SUB(reference_integer_ops, BroadcastAdd4DSlow, int8_t);
     } else {
       TF_LITE_SUB(reference_integer_ops, Add, int8_t);
     }
   } else if (data->pot_scale_16bit) {
     if (need_broadcast) {
       TF_LITE_SUB(reference_ops, BroadcastAdd4DSlow, int16_t);
     } else {
       reference_ops::Add(op_params, GetTensorShape(input1),
                          GetTensorData<int16_t>(input1), GetTensorShape(input2),
                          GetTensorData<int16_t>(input2), GetTensorShape(output),
                          GetTensorData<int16_t>(output), false);
     }
   } else if (output->type == kTfLiteUInt8) {
     if (kernel_type == kReference) {
       if (need_broadcast) {
         TF_LITE_SUB(reference_ops, BroadcastAdd4DSlow, uint8_t);
       } else {
         TF_LITE_SUB(reference_ops, Add, uint8_t);
       }
     } else {
       if (op_params.broadcast_category ==
           BroadcastableOpCategory::kGenericBroadcast) {
         TF_LITE_SUB(optimized_ops, BroadcastAdd4DSlow, uint8_t);
       } else if (need_broadcast) {
         TF_LITE_SUB(optimized_ops, BroadcastAddFivefold, uint8_t);
       } else {
         TF_LITE_SUB(optimized_ops, Add, uint8_t);
       }
     }
   } else {
     if (kernel_type == kReference) {
       if (need_broadcast) {
         TF_LITE_SUB(reference_ops, BroadcastSub4DSlow, int16_t);
       } else {
         TF_LITE_SUB(reference_ops, Sub16, int16_t);
       }
     } else {
       if (need_broadcast) {
         TF_LITE_SUB(optimized_ops, BroadcastSub4DSlow, int16_t);
       } else {
         TF_LITE_SUB(optimized_ops, Sub16, int16_t);
       }
     }
   }
 #undef TF_LITE_SUB
 }

 template <KernelType kernel_type>
 TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
   auto* params = reinterpret_cast<TfLiteSubParams*>(node->builtin_data);
   OpData* data = reinterpret_cast<OpData*>(node->user_data);

   const TfLiteTensor* input1 = GetInput(context, node, kInputTensor1);
   const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2);
   TfLiteTensor* output = GetOutput(context, node, kOutputTensor);

   if (output->type == kTfLiteFloat32 || output->type == kTfLiteInt32) {
     EvalSub<kernel_type>(context, node, params, data, input1, input2, output);
   } else if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8 ||
              output->type == kTfLiteInt16) {
     EvalQuantized<kernel_type>(context, node, params, data, input1, input2,
                                output);
   } else {
     context->ReportError(
         context,
         "output type %d is not supported, requires float|uint8|int32 types.",
         output->type);
     return kTfLiteError;
   }

   return kTfLiteOk;
 }

 }  // namespace sub

 TfLiteRegistration* Register_SUB_REF() {
   static TfLiteRegistration r = {sub::Init, sub::Free, sub::Prepare,
                                  sub::Eval<sub::kReference>};
   return &r;
 }

 TfLiteRegistration* Register_SUB_GENERIC_OPT() {
   static TfLiteRegistration r = {sub::Init, sub::Free, sub::Prepare,
                                  sub::Eval<sub::kGenericOptimized>};
   return &r;
 }

 TfLiteRegistration* Register_SUB_NEON_OPT() {
   static TfLiteRegistration r = {sub::Init, sub::Free, sub::Prepare,
                                  sub::Eval<sub::kNeonOptimized>};
   return &r;
 }

 TfLiteRegistration* Register_SUB() {
 #ifdef USE_NEON
   return Register_SUB_NEON_OPT();
 #else
   return Register_SUB_GENERIC_OPT();
 #endif
 }

 }  // namespace builtin
 }  // namespace ops
 }  // namespace tflite
	/* Copyright 2017 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.
	==============================================================================*/
	#include <limits>

	#include "tensorflow/lite/c/builtin_op_data.h"
	#include "tensorflow/lite/c/common.h"
	#include "tensorflow/lite/kernels/internal/optimized/cpu_check.h"
	#include "tensorflow/lite/kernels/internal/optimized/optimized_ops.h"
	#include "tensorflow/lite/kernels/internal/quantization_util.h"
	#include "tensorflow/lite/kernels/internal/reference/integer_ops/add.h"
	#include "tensorflow/lite/kernels/internal/reference/reference_ops.h"
	#include "tensorflow/lite/kernels/internal/tensor.h"
	#include "tensorflow/lite/kernels/kernel_util.h"
	#include "tensorflow/lite/kernels/op_macros.h"

	namespace tflite {
	namespace ops {
	namespace builtin {
	namespace sub {

	// This file has three implementation of Sub.
	enum KernelType {
	kReference,
	kGenericOptimized, // Neon-free
	kNeonOptimized,
	};

	constexpr int kInputTensor1 = 0;
	constexpr int kInputTensor2 = 1;
	constexpr int kOutputTensor = 0;

	struct OpData {
	bool requires_broadcast;

	// These fields are used in both the general 8-bit -> 8bit quantized path,
	// and the special 16-bit -> 16bit quantized path
	int input1_shift;
	int input2_shift;
	int32 output_activation_min;
	int32 output_activation_max;

	// These fields are used only in the general 8-bit -> 8bit quantized path
	int32 input1_multiplier;
	int32 input2_multiplier;
	int32 output_multiplier;
	int output_shift;
	int left_shift;
	int32 input1_offset;
	int32 input2_offset;
	int32 output_offset;

	// This parameter is used to indicate whether
	// parameter scale is power of two.
	// It is used in 16-bit -> 16-bit quantization.
	bool pot_scale_16bit;
	};

	void* Init(TfLiteContext* context, const char* buffer, size_t length) {
	auto* data = new OpData;
	data->requires_broadcast = false;
	return data;
	}

	void Free(TfLiteContext* context, void* buffer) {
	delete reinterpret_cast<OpData*>(buffer);
	}

	TfLiteStatus PrepareGeneralSubOp(TfLiteContext* context,
	const TfLiteTensor* input_1,
	const TfLiteTensor* input_2,
	TfLiteTensor* output, TfLiteSubParams* params,
	OpData* op_params, int op_sign) {
	TF_LITE_ENSURE(context, output->type == kTfLiteUInt8 \|\|
	output->type == kTfLiteInt8 \|\|
	output->type == kTfLiteInt16);
	const auto& input1_quantization_params = input_1->params;
	const auto& input2_quantization_params = input_2->params;
	const auto& output_quantization_params = output->params;
	int32_t integer_type_min = 0;
	int32_t integer_type_max = 0;
	if (output->type == kTfLiteUInt8) {
	integer_type_min = std::numeric_limits<uint8_t>::min();
	integer_type_max = std::numeric_limits<uint8_t>::max();
	} else if (output->type == kTfLiteInt16) {
	integer_type_min = std::numeric_limits<int16_t>::min();
	integer_type_max = std::numeric_limits<int16_t>::max();
	} else {
	// output->type == kTfLiteInt8
	integer_type_min = std::numeric_limits<int8_t>::min();
	integer_type_max = std::numeric_limits<int8_t>::max();
	}

	TF_LITE_ENSURE(context,
	input1_quantization_params.zero_point >= integer_type_min);
	TF_LITE_ENSURE(context,
	input1_quantization_params.zero_point <= integer_type_max);
	TF_LITE_ENSURE(context,
	input2_quantization_params.zero_point >= integer_type_min);
	TF_LITE_ENSURE(context,
	input2_quantization_params.zero_point <= integer_type_max);
	TF_LITE_ENSURE(context,
	output_quantization_params.zero_point >= integer_type_min);
	TF_LITE_ENSURE(context,
	output_quantization_params.zero_point <= integer_type_max);

	op_params->input1_offset = -input1_quantization_params.zero_point;
	op_params->input2_offset = -input2_quantization_params.zero_point;
	op_params->output_offset = output_quantization_params.zero_point;

	// The shift is set to 15 in case of 16-bit and 20 in case of 8-bit,
	// accordingly. In case of 16-bit we have 65535 << 15 which is less than 1 <<
	// 31, therefore the addition will still fit in a 32 bit accumulator.
	op_params->left_shift = output->type == kTfLiteInt16 ? 15 : 20;
	const double twice_max_input_scale =
	2 * std::max(input1_quantization_params.scale,
	input2_quantization_params.scale);
	const double real_input1_multiplier =
	input1_quantization_params.scale / twice_max_input_scale;
	const double real_input2_multiplier =
	input2_quantization_params.scale / twice_max_input_scale;
	const double real_output_multiplier =
	twice_max_input_scale /
	((1 << op_params->left_shift) * output_quantization_params.scale);

	tflite::QuantizeMultiplierSmallerThanOneExp(real_input1_multiplier,
	&op_params->input1_multiplier,
	&op_params->input1_shift);
	tflite::QuantizeMultiplierSmallerThanOneExp(real_input2_multiplier,
	&op_params->input2_multiplier,
	&op_params->input2_shift);
	op_params->input2_multiplier *= op_sign;
	tflite::QuantizeMultiplierSmallerThanOneExp(real_output_multiplier,
	&op_params->output_multiplier,
	&op_params->output_shift);

	TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized(
	context, params->activation, output, &op_params->output_activation_min,
	&op_params->output_activation_max));

	return kTfLiteOk;
	}

	TfLiteStatus PrepareInt16SubOpPOT(TfLiteContext* context,
	const TfLiteTensor* input1,
	const TfLiteTensor* input2,
	TfLiteTensor* output, TfLiteSubParams* params,
	OpData* data) {
	// 16bit -> 16bit special quantized path, supporting only a rather
	// narrow case of quantization parameters: zero_points must all be 0
	// ("symmetric quantization") and scales must be power-of-two (which
	// we abbreviate as "POT" below). The intended use case for this path
	// is in LSTM cells, where, due to the constraints of implementing
	// some of the math in these LSTM cells in fixed-point arithmetic,
	// we need to have such symmetric, power-of-two quantization
	// (Fixed-point formats are inherently symmetric, power-of-two).
	TF_LITE_ENSURE_EQ(context, input1->params.zero_point, 0);
	TF_LITE_ENSURE_EQ(context, input2->params.zero_point, 0);
	TF_LITE_ENSURE_EQ(context, output->params.zero_point, 0);

	int input1_scale_log2_rounded;
	bool input1_scale_is_pot =
	CheckedLog2(input1->params.scale, &input1_scale_log2_rounded);
	TF_LITE_ENSURE(context, input1_scale_is_pot);

	int input2_scale_log2_rounded;
	bool input2_scale_is_pot =
	CheckedLog2(input2->params.scale, &input2_scale_log2_rounded);
	TF_LITE_ENSURE(context, input2_scale_is_pot);

	int output_scale_log2_rounded;
	bool output_scale_is_pot =
	CheckedLog2(output->params.scale, &output_scale_log2_rounded);
	TF_LITE_ENSURE(context, output_scale_is_pot);

	data->input1_shift = input1_scale_log2_rounded - output_scale_log2_rounded;
	data->input2_shift = input2_scale_log2_rounded - output_scale_log2_rounded;

	// Shifting of one input is supported. The graph quantization should ensure
	// that the other input matches the output.
	TF_LITE_ENSURE(context, data->input1_shift == 0 \|\| data->input2_shift == 0);
	TF_LITE_ENSURE(context, data->input1_shift <= 0);
	TF_LITE_ENSURE(context, data->input2_shift <= 0);

	TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized(
	context, params->activation, output, &data->output_activation_min,
	&data->output_activation_max));
	return kTfLiteOk;
	}

	TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
	OpData* data = reinterpret_cast<OpData*>(node->user_data);
	auto* params = reinterpret_cast<TfLiteSubParams*>(node->builtin_data);

	TF_LITE_ENSURE_EQ(context, NumInputs(node), 2);
	TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1);

	const TfLiteTensor* input1 = GetInput(context, node, kInputTensor1);
	const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2);
	TfLiteTensor* output = GetOutput(context, node, kOutputTensor);

	TF_LITE_ENSURE_EQ(context, input1->type, input2->type);
	output->type = input2->type;

	data->requires_broadcast = !HaveSameShapes(input1, input2);

	TfLiteIntArray* output_size = nullptr;
	if (data->requires_broadcast) {
	TF_LITE_ENSURE_OK(context, CalculateShapeForBroadcast(
	context, input1, input2, &output_size));
	} else {
	output_size = TfLiteIntArrayCopy(input1->dims);
	}

	// 8bit -> 8bit general quantized path, with general rescalings
	// as well as, 16bit -> 16bit with general rescalings
	bool pot_scale_16bit = false;

	bool input1_scale_is_pot = false;
	bool input2_scale_is_pot = false;
	bool output_scale_is_pot = false;

	int input1_scale_log2_rounded;
	int input2_scale_log2_rounded;
	int output_scale_log2_rounded;

	if (input1->type == kTfLiteInt16 && input2->type == kTfLiteInt16 &&
	output->type == kTfLiteInt16) {
	// Check that param scale is POT
	input1_scale_is_pot =
	CheckedLog2(input1->params.scale, &input1_scale_log2_rounded);

	input2_scale_is_pot =
	CheckedLog2(input2->params.scale, &input2_scale_log2_rounded);

	output_scale_is_pot =
	CheckedLog2(output->params.scale, &output_scale_log2_rounded);

	pot_scale_16bit = input1_scale_log2_rounded && input2_scale_log2_rounded &&
	output_scale_log2_rounded;
	}

	data->pot_scale_16bit = pot_scale_16bit;

	if (output->type == kTfLiteUInt8 \|\| output->type == kTfLiteInt8 \|\|
	pot_scale_16bit) {
	TF_LITE_ENSURE_OK(context, PrepareGeneralSubOp(context, input1, input2,
	output, params, data, -1));
	} else if (output->type == kTfLiteInt16) {
	// LSTM-special case with scale parameter of POT
	TF_LITE_ENSURE_OK(context, PrepareInt16SubOpPOT(context, input1, input2,
	output, params, data));
	}

	return context->ResizeTensor(context, output, output_size);
	}

	template <KernelType kernel_type>
	void EvalSub(TfLiteContext* context, TfLiteNode* node, TfLiteSubParams* params,
	const OpData* data, const TfLiteTensor* input1,
	const TfLiteTensor* input2, TfLiteTensor* output) {
	#define TF_LITE_SUB(type, opname, data_type) \
	data_type output_activation_min, output_activation_max; \
	CalculateActivationRange(params->activation, &output_activation_min, \
	&output_activation_max); \
	tflite::ArithmeticParams op_params; \
	SetActivationParams(output_activation_min, output_activation_max, \
	&op_params); \
	type::opname(op_params, GetTensorShape(input1), \
	GetTensorData<data_type>(input1), GetTensorShape(input2), \
	GetTensorData<data_type>(input2), GetTensorShape(output), \
	GetTensorData<data_type>(output))
	if (output->type == kTfLiteInt32) {
	if (kernel_type == kReference) {
	if (data->requires_broadcast) {
	TF_LITE_SUB(reference_ops, BroadcastSub4DSlow, int32_t);
	} else {
	TF_LITE_SUB(reference_ops, SubWithActivation, int32_t);
	}
	} else {
	if (data->requires_broadcast) {
	TF_LITE_SUB(optimized_ops, BroadcastSub4DSlow, int32_t);
	} else {
	TF_LITE_SUB(optimized_ops, SubWithActivation, int32_t);
	}
	}
	} else if (output->type == kTfLiteFloat32) {
	if (kernel_type == kReference) {
	if (data->requires_broadcast) {
	TF_LITE_SUB(reference_ops, BroadcastSub4DSlow, float);
	} else {
	TF_LITE_SUB(reference_ops, SubWithActivation, float);
	}
	} else {
	if (data->requires_broadcast) {
	TF_LITE_SUB(optimized_ops, BroadcastSub4DSlow, float);
	} else {
	TF_LITE_SUB(optimized_ops, SubWithActivation, float);
	}
	}
	}
	#undef TF_LITE_SUB
	}

	template <KernelType kernel_type>
	void EvalQuantized(TfLiteContext* context, TfLiteNode* node,
	TfLiteSubParams* params, const OpData* data,
	const TfLiteTensor* input1, const TfLiteTensor* input2,
	TfLiteTensor* output) {
	tflite::ArithmeticParams op_params;
	op_params.left_shift = data->left_shift;
	op_params.input1_offset = data->input1_offset;
	op_params.input1_multiplier = data->input1_multiplier;
	op_params.input1_shift = data->input1_shift;
	op_params.input2_offset = data->input2_offset;
	op_params.input2_multiplier = data->input2_multiplier;
	op_params.input2_shift = data->input2_shift;
	op_params.output_offset = data->output_offset;
	op_params.output_multiplier = data->output_multiplier;
	op_params.output_shift = data->output_shift;
	SetActivationParams(data->output_activation_min, data->output_activation_max,
	&op_params);

	const bool need_broadcast = optimized_ops::ProcessBroadcastShapes(
	GetTensorShape(input1), GetTensorShape(input2), &op_params);

	#define TF_LITE_SUB(type, opname, data_type) \
	type::opname(op_params, GetTensorShape(input1), \
	GetTensorData<data_type>(input1), GetTensorShape(input2), \
	GetTensorData<data_type>(input2), GetTensorShape(output), \
	GetTensorData<data_type>(output))
	// NOTE: We are using the add kernels. This is possible as the second values
	// multiplier is negated before being passed down.
	if (output->type == kTfLiteInt8) {
	if (need_broadcast) {
	TF_LITE_SUB(reference_integer_ops, BroadcastAdd4DSlow, int8_t);
	} else {
	TF_LITE_SUB(reference_integer_ops, Add, int8_t);
	}
	} else if (data->pot_scale_16bit) {
	if (need_broadcast) {
	TF_LITE_SUB(reference_ops, BroadcastAdd4DSlow, int16_t);
	} else {
	reference_ops::Add(op_params, GetTensorShape(input1),
	GetTensorData<int16_t>(input1), GetTensorShape(input2),
	GetTensorData<int16_t>(input2), GetTensorShape(output),
	GetTensorData<int16_t>(output), false);
	}
	} else if (output->type == kTfLiteUInt8) {
	if (kernel_type == kReference) {
	if (need_broadcast) {
	TF_LITE_SUB(reference_ops, BroadcastAdd4DSlow, uint8_t);
	} else {
	TF_LITE_SUB(reference_ops, Add, uint8_t);
	}
	} else {
	if (op_params.broadcast_category ==
	BroadcastableOpCategory::kGenericBroadcast) {
	TF_LITE_SUB(optimized_ops, BroadcastAdd4DSlow, uint8_t);
	} else if (need_broadcast) {
	TF_LITE_SUB(optimized_ops, BroadcastAddFivefold, uint8_t);
	} else {
	TF_LITE_SUB(optimized_ops, Add, uint8_t);
	}
	}
	} else {
	if (kernel_type == kReference) {
	if (need_broadcast) {
	TF_LITE_SUB(reference_ops, BroadcastSub4DSlow, int16_t);
	} else {
	TF_LITE_SUB(reference_ops, Sub16, int16_t);
	}
	} else {
	if (need_broadcast) {
	TF_LITE_SUB(optimized_ops, BroadcastSub4DSlow, int16_t);
	} else {
	TF_LITE_SUB(optimized_ops, Sub16, int16_t);
	}
	}
	}
	#undef TF_LITE_SUB
	}

	template <KernelType kernel_type>
	TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
	auto* params = reinterpret_cast<TfLiteSubParams*>(node->builtin_data);
	OpData* data = reinterpret_cast<OpData*>(node->user_data);

	const TfLiteTensor* input1 = GetInput(context, node, kInputTensor1);
	const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2);
	TfLiteTensor* output = GetOutput(context, node, kOutputTensor);

	if (output->type == kTfLiteFloat32 \|\| output->type == kTfLiteInt32) {
	EvalSub<kernel_type>(context, node, params, data, input1, input2, output);
	} else if (output->type == kTfLiteUInt8 \|\| output->type == kTfLiteInt8 \|\|
	output->type == kTfLiteInt16) {
	EvalQuantized<kernel_type>(context, node, params, data, input1, input2,
	output);
	} else {
	context->ReportError(
	context,
	"output type %d is not supported, requires float\|uint8\|int32 types.",
	output->type);
	return kTfLiteError;
	}

	return kTfLiteOk;
	}

	} // namespace sub

	TfLiteRegistration* Register_SUB_REF() {
	static TfLiteRegistration r = {sub::Init, sub::Free, sub::Prepare,
	sub::Eval<sub::kReference>};
	return &r;
	}

	TfLiteRegistration* Register_SUB_GENERIC_OPT() {
	static TfLiteRegistration r = {sub::Init, sub::Free, sub::Prepare,
	sub::Eval<sub::kGenericOptimized>};
	return &r;
	}

	TfLiteRegistration* Register_SUB_NEON_OPT() {
	static TfLiteRegistration r = {sub::Init, sub::Free, sub::Prepare,
	sub::Eval<sub::kNeonOptimized>};
	return &r;
	}

	TfLiteRegistration* Register_SUB() {
	#ifdef USE_NEON
	return Register_SUB_NEON_OPT();
	#else
	return Register_SUB_GENERIC_OPT();
	#endif
	}

	} // namespace builtin
	} // namespace ops
	} // namespace tflite