/*
* Copyright (c) 2020 Samsung Electronics Co., Ltd. All Rights Reserved
- * Copyright 2019 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.
* limitations under the License.
*/
-#include "kernels/Relu.h"
+#include "Builders.h"
#include "kernels/Utils.h"
+#include "SISOKernel.h"
-#include "PALRelu.h"
+#include "PALReluCommon.h"
namespace luci_interpreter
{
-namespace kernels
+void configure_kernel_CircleRelu(const circle::Operator *cur_op, BaseRuntimeGraph *runtime_graph)
{
+ kernels::SISOKernel kernel(cur_op, runtime_graph);
-Relu::Relu(const Tensor *input, Tensor *output) : Kernel({input}, {output}) {}
+ LUCI_INTERPRETER_CHECK(Tensor::element_type(kernel.input()) ==
+ Tensor::element_type(kernel.output()));
+ LUCI_INTERPRETER_CHECK(Tensor::num_dims(kernel.input()) == Tensor::num_dims(kernel.output()));
+ LUCI_INTERPRETER_CHECK(Tensor::num_elements(kernel.input()) ==
+ Tensor::num_elements(kernel.output()));
+}
-void Relu::configure()
+void execute_kernel_CircleRelu(const circle::Operator *cur_op, BaseRuntimeGraph *runtime_graph)
{
- LUCI_INTERPRETER_CHECK(input()->element_type() == output()->element_type());
- if (input()->element_type() == DataType::S16)
- {
- LUCI_INTERPRETER_CHECK(input()->zero_point() == 0 && output()->zero_point() == 0);
- }
+ kernels::SISOKernel kernel(cur_op, runtime_graph);
- if (input()->element_type() == DataType::U8 || input()->element_type() == DataType::S16)
- {
- double multiplier = input()->scale() / output()->scale();
- quantizeMultiplier(multiplier, &_output_multiplier, &_output_shift);
- }
- // TODO: enable it only if kernel with dynamic shapes
- output()->resize(input()->shape());
-}
+ const auto *input_data = runtime_graph->getDataByTensor(kernel.input());
+ assert(input_data);
-void Relu::execute() const
-{
- switch (input()->element_type())
+ auto *output_data = runtime_graph->getDataByTensor(kernel.output());
+
+ bool is_inplace = runtime_graph->is_inplace_op(cur_op);
+
+ switch (Tensor::element_type(kernel.input()))
{
+#ifndef DIS_FLOAT
case DataType::FLOAT32:
- evalFloat();
- break;
- case DataType::U8:
- evalQuantized();
- break;
- case DataType::S16:
- evalQuantizedS16();
+ {
+ const float *input_data_float = kernels::getTensorData<float>(input_data);
+ float *output_data_float = kernels::getTensorData<float>(output_data);
+ if (is_inplace)
+ {
+ output_data_float = const_cast<float *>(input_data_float);
+ }
+
+ assert(output_data_float);
+ const int flat_size =
+ kernels::getTensorRuntimeShape(kernel.input(), runtime_graph).flatSize();
+
+ luci_interpreter_pal::ReLUCommon(flat_size, input_data_float, output_data_float, 0.0f, false);
break;
+ }
+#endif // DIS_FLOAT
default:
- assert(false && "Unsupported type.");
+ assert(false && "Unsupported type");
}
-}
-
-void Relu::evalFloat() const
-{
- const auto input_data = getTensorData<float>(input());
- const auto input_shape = getTensorShape(input());
- auto output_data = getTensorData<float>(output());
- auto output_shape = getTensorShape(output());
- luci_interpreter_pal::Relu(input_shape, input_data, output_shape, output_data);
-}
-
-void Relu::evalQuantized() const
-{
- tflite::ReluParams params;
- params.input_offset = input()->zero_point();
- params.output_offset = output()->zero_point();
- params.output_multiplier = _output_multiplier;
- params.output_shift = _output_shift;
-
- params.quantized_activation_min =
- std::max(static_cast<int32_t>(std::numeric_limits<uint8_t>::min()), params.output_offset);
- params.quantized_activation_max = static_cast<int32_t>(std::numeric_limits<uint8_t>::max());
-
- luci_interpreter_pal::ReluX(params, getTensorShape(input()), getTensorData<uint8_t>(input()),
- getTensorShape(output()), getTensorData<uint8_t>(output()));
-}
-
-void Relu::evalQuantizedS16() const
-{
- const auto *input_data = getTensorData<int16_t>(input());
- auto *output_data = getTensorData<int16_t>(output());
-
- constexpr int32_t output_min = 0;
- constexpr int32_t output_max = std::numeric_limits<int16_t>::max();
-
- const int32_t num_elements = input()->shape().num_elements();
-
- for (int32_t i = 0; i < num_elements; ++i)
- {
- const int32_t input_val = input_data[i];
- int32_t output_val =
- tflite::MultiplyByQuantizedMultiplier(input_val, _output_multiplier, _output_shift);
- output_val = std::max(output_val, output_min);
- output_val = std::min(output_val, output_max);
- output_data[i] = static_cast<int16_t>(output_val);
- }
+ if (is_inplace)
+ runtime_graph->makeInplaceOperation(kernel.input(), kernel.output());
}
-} // namespace kernels
} // namespace luci_interpreter