/*
* 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/Tanh.h"
-
+#include "Builders.h"
#include "kernels/Utils.h"
+#include "SISOKernel.h"
-#include <tensorflow/lite/kernels/internal/reference/tanh.h>
+#include "PALTanh.h"
namespace luci_interpreter
{
-namespace kernels
-{
-Tanh::Tanh(const Tensor *input, Tensor *output) : Kernel({input}, {output}) {}
+#ifndef DIS_QUANT
-void Tanh::configure()
+namespace
+{
+void calculateArithmeticData(const circle::Tensor *input, const circle::Tensor *output,
+ int32_t &input_zero_point, int32_t &input_range_radius,
+ int32_t &input_multiplier, int &input_left_shift)
{
- LUCI_INTERPRETER_CHECK(input()->element_type() == output()->element_type());
- if (input()->element_type() == DataType::U8)
+ const auto input_dtype = Tensor::element_type(input);
+ switch (input_dtype)
{
- populateLookupTable();
+ // TODO: enable it
+#if 0
+ case DataType::S8:
+ {
+ static constexpr int input_integer_bits = 4;
+ const double input_real_multiplier = static_cast<double>(Tensor::scale(input)) *
+ static_cast<double>(1 << (31 - input_integer_bits));
+
+ const double q = std::frexp(input_real_multiplier, &input_left_shift);
+ input_multiplier = static_cast<int32_t>(std::round(q * (1ll << 31)));
+ input_range_radius = kernels::calculateInputRadius(input_integer_bits, input_left_shift, 31);
+ }
+ break;
+#endif
+ case DataType::S16:
+ {
+ static constexpr int input_integer_bits = 3;
+ static constexpr int output_fractional_bits = 15;
+
+ // These operators are implemented in fixed-point arithmetic,
+ // which intrinsically wants symmetric ranges (zero_point==0)
+ // and power-of-two scales (power-of-two is abbreviated below as POT).
+ // While more general support would be possible by means of rescaling,
+ // that would add some overhead and some loss of accuracy and wouldn't
+ // be used at the moment as current quantized LSTM applications are
+ // happy with symmetric, power-of-two-scales quantization. So we just
+ // implement that narrow case only for now.
+
+ int input_scale_log2_rounded;
+ bool param_scale_pot = kernels::checkedLog2(Tensor::scale(input), &input_scale_log2_rounded);
+
+ input_left_shift = (15 - input_integer_bits) + input_scale_log2_rounded;
+ param_scale_pot &= (input_left_shift == 0 || input_left_shift == 1);
+
+ if (param_scale_pot)
+ {
+ input_multiplier = 0;
+ }
+ else
+ {
+ // Calculate multiplier to change input scale to 1/(3*4096)
+ // as required by the table lookup.
+ // The number 3.0 in the multiplier comes from here,
+ // because the interval is [-10.7, 10.7] instead of [-8, 8].
+ // So, in this scaling +/-2^17 represents +/-10.7.
+
+ double multiplier = static_cast<double>(Tensor::scale(input)) * 4096.0 * 3.0;
+ input_left_shift = 0;
+
+ while (multiplier <= 32767.0 / 2.0 && input_left_shift <= 30)
+ {
+ input_left_shift++;
+ multiplier = multiplier * 2.0;
+ }
+
+ input_multiplier = static_cast<int32_t>(multiplier);
+ }
+
+ int output_scale_log2_rounded;
+ kernels::checkedLog2(Tensor::scale(output), &output_scale_log2_rounded);
+ assert(output_scale_log2_rounded == -output_fractional_bits);
+ }
+ break;
+ default:
+ assert(false && "Unsupported type");
}
- // TODO: enable it only if kernel with dynamic shapes
- output()->resize(input()->shape());
}
-void Tanh::execute() const
+} // namespace
+
+void evalInteger(const circle::Tensor *input, const circle::Tensor *output,
+ BaseRuntimeGraph *runtime_graph)
{
- switch (input()->element_type())
+ int32_t input_zero_point = 0;
+ int32_t input_range_radius = 0;
+ int32_t input_multiplier = 0;
+ int input_left_shift = 0;
+
+ calculateArithmeticData(input, output, input_zero_point, input_range_radius, input_multiplier,
+ input_left_shift);
+
+ const auto *input_data = runtime_graph->getDataByTensor(input);
+ assert(input_data);
+
+ auto *output_data = runtime_graph->getDataByTensor(output);
+ assert(output_data);
+
+ const int flat_size = kernels::getTensorRuntimeShape(input, runtime_graph).flatSize();
+
+ const auto input_dtype = Tensor::element_type(input);
+ switch (input_dtype)
{
- case DataType::FLOAT32:
- evalFloat();
+ // TODO: enable it
+#if 0
+ case DataType::S8:
+ luci_interpreter_pal::Tanh(
+ input_zero_point, input_range_radius, input_multiplier, input_left_shift,
+ flat_size, kernels::getTensorData<int8_t>(input_data), kernels::getTensorData<int8_t>(output_data));
break;
- case DataType::U8:
- evalQuantized();
+#endif // 0
+ case DataType::S16:
+ luci_interpreter_pal::Tanh(input_multiplier, input_left_shift, flat_size,
+ kernels::getTensorData<int16_t>(input_data),
+ kernels::getTensorData<int16_t>(output_data));
break;
default:
- assert(false && "Unsupported type.");
+ assert(false && "Not support yet");
}
}
+#endif // DIS_QUANT
-void Tanh::evalFloat() const
+void configure_kernel_CircleTanh(const circle::Operator *cur_op, BaseRuntimeGraph *runtime_graph)
{
- tflite::reference_ops::Tanh(getTensorShape(input()), getTensorData<float>(input()),
- getTensorShape(output()), getTensorData<float>(output()));
-}
+ kernels::SISOKernel kernel(cur_op, runtime_graph);
-void Tanh::evalQuantized() const
-{
- const int size = tflite::MatchingFlatSize(getTensorShape(input()), getTensorShape(output()));
- uint8_t *output_data = getTensorData<uint8_t>(output());
- const uint8_t *input_data = getTensorData<uint8_t>(input());
- for (int i = 0; i < size; ++i)
- {
- output_data[i] = getTableValue(input_data[i]);
- }
+ LUCI_INTERPRETER_CHECK(Tensor::element_type(kernel.input()) ==
+ Tensor::element_type(kernel.output()));
}
-void Tanh::populateLookupTable()
+void execute_kernel_CircleTanh(const circle::Operator *cur_op, BaseRuntimeGraph *runtime_graph)
{
- const auto input_scale = static_cast<double>(input()->scale());
- const auto input_zero_point = static_cast<int32_t>(input()->zero_point());
- const auto output_scale = static_cast<double>(output()->scale());
- const auto output_zero_point = static_cast<int32_t>(output()->zero_point());
- const float inverse_scale = 1 / output_scale;
- int32_t maxval = std::numeric_limits<uint8_t>::max();
- int32_t minval = std::numeric_limits<uint8_t>::min();
- for (int32_t val = minval; val <= maxval; ++val)
+ kernels::SISOKernel kernel(cur_op, runtime_graph);
+
+ const auto *input_data = runtime_graph->getDataByTensor(kernel.input());
+ assert(input_data);
+
+ auto *output_data = runtime_graph->getDataByTensor(kernel.output());
+
+ bool is_inplace = runtime_graph->is_inplace_op(cur_op);
+
+ switch (Tensor::element_type(kernel.input()))
{
- const float dequantized = input_scale * (val - input_zero_point);
- const float transformed = std::tanh(dequantized);
- const float rescaled = std::round(transformed * inverse_scale);
- const int32_t quantized = static_cast<int32_t>(rescaled + output_zero_point);
- setTableValue(static_cast<uint8_t>(std::max(std::min(maxval, quantized), minval)),
- static_cast<uint8_t>(val));
+#ifndef DIS_FLOAT
+ case DataType::FLOAT32:
+ {
+ 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::Tanh(flat_size, input_data_float, output_data_float);
+ break;
+ }
+#endif // DIS_FLOAT
+#ifndef DIS_QUANT
+ case DataType::S16:
+ // TODO: enable it
+#if 0
+ case DataType::S8:
+#endif
+ evalInteger(kernel.input(), kernel.output(), runtime_graph);
+ break;
+#endif // DIS_QUANT
+ default:
+ assert(false && "Unsupported type");
}
-}
-} // namespace kernels
+ if (is_inplace)
+ runtime_graph->makeInplaceOperation(kernel.input(), kernel.output());
+}
} // namespace luci_interpreter