1 files changed, 172 insertions, 0 deletions
diff --git a/runtime/neurun/backend/cpu/kernel/SoftMaxLayer.cc b/runtime/neurun/backend/cpu/kernel/SoftMaxLayer.cc
new file mode 100644
index 000000000..58ba109b4
--- /dev/null
+++ b/runtime/neurun/backend/cpu/kernel/SoftMaxLayer.cc
@@ -0,0 +1,172 @@
+/*
+ * Copyright (c) 2018 Samsung Electronics Co., Ltd. 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 "SoftMaxLayer.h"
+
+#include <cker/operation/SoftMax.h>
+
+#include "OperationUtils.h"
+
+namespace neurun
+{
+namespace backend
+{
+namespace cpu
+{
+namespace kernel
+{
+
+SoftMaxLayer::SoftMaxLayer()
+    : _inputData(), _outputData(), _beta(0.0), _inputDescr(), _outputDescr(),
+      _inputType(OperandType::FLOAT32)
+{
+  // DO NOTHING
+}
+
+// Performs softmax along the input of size (input_size * batch_size).
+void Softmax(const float *in, const int input_size, const int batch_size, const float beta,
+             float *out)
+{
+  assert(input_size > 0);
+
+  // For each batch
+  for (int b = 0; b < batch_size; b++)
+  {
+    // Find the max coeff.
+    float max_coeff = in[0];
+    for (int i = 1; i < input_size; i++)
+    {
+      if (in[i] > max_coeff)
+        max_coeff = in[i];
+    }
+
+    // Compute the normalized sum of exps.
+    float exp_sum = 0.0;
+    for (int i = 0; i < input_size; i++)
+    {
+      out[i] = std::exp((in[i] - max_coeff) * beta);
+      exp_sum += out[i];
+    }
+
+    // Divide by the sum of exps.
+    float reciprocal_sum_exp = 1.f / exp_sum;
+    for (int i = 0; i < input_size; i++)
+    {
+      out[i] *= reciprocal_sum_exp;
+    }
+
+    // Advance in and out pointers for the next batch.
+    in += input_size;
+    out += input_size;
+  }
+}
+
+void SoftMaxLayer::softmaxFloat32()
+{
+  TensorDescriptor descrIn4D;
+
+  if (getNumberOfDimensions(_inputDescr) == 2)
+  {
+    uint32_t batch_size = getSizeOfDimension(_inputDescr, 0);
+    if (batch_size == 0)
+      throw std::runtime_error("batch_size should not be 0");
+
+    uint32_t input_size = getNumberOfElements(_inputDescr) / batch_size;
+    Softmax(_inputData.f, input_size, batch_size, _beta, _outputData.f);
+  }
+  else if (getNumberOfDimensions(_inputDescr) == 4)
+  {
+    nnfw::cker::SoftmaxParams op_params;
+    op_params.beta = _beta;
+    nnfw::cker::Softmax(op_params, convertTensorDescriptorToCkerShape(_inputDescr), _inputData.f,
+                        convertTensorDescriptorToCkerShape(_outputDescr), _outputData.f);
+  }
+  else
+  {
+    throw std::runtime_error{"only 2D and 4D tensors supported"};
+  }
+}
+
+void SoftMaxLayer::softmaxQuant8()
+{
+  TensorDescriptor descrIn4D = _inputDescr;
+
+  if (getNumberOfDimensions(_inputDescr) == 2)
+  {
+    uint32_t batch_size = getSizeOfDimension(_inputDescr, 0);
+    if (batch_size == 0)
+      throw std::runtime_error("batch_size should not be 0");
+
+    uint32_t input_size = getNumberOfElements(_inputDescr) / batch_size;
+    descrIn4D.dimensions = {batch_size, 1, 1, input_size};
+  }
+  else if (getNumberOfDimensions(_inputDescr) == 4)
+  {
+    descrIn4D = _inputDescr;
+  }
+  else
+  {
+    throw std::runtime_error{"only 2D and 4D tensors supported"};
+  }
+  if (_outputDescr.offset != 0 || _outputDescr.scale != 1.f / 256)
+  {
+    throw std::runtime_error{"incorrect scale / offset for output"};
+  }
+  static const int32_t kScaledDiffIntegerBits = 5;
+  const double input_beta_real_multiplier = std::min(
+      1.0 * _beta * _inputDescr.scale * (1 << (31 - kScaledDiffIntegerBits)), (1ll << 31) - 1.0);
+  int32_t input_multiplier = 0;
+  int32_t input_left_shift = 0;
+  QuantizeMultiplierGreaterThanOne(input_beta_real_multiplier, &input_multiplier,
+                                   &input_left_shift);
+  float diff_min = -1.0f * CalculateInputRadius(kScaledDiffIntegerBits, input_left_shift);
+
+  nnfw::cker::SoftmaxParams op_params;
+  op_params.input_multiplier = input_multiplier;
+  op_params.input_left_shift = input_left_shift;
+  op_params.diff_min = diff_min;
+  nnfw::cker::Softmax(op_params, convertTensorDescriptorToCkerShape(descrIn4D), _inputData.u8,
+                      convertTensorDescriptorToCkerShape(descrIn4D), _outputData.u8);
+}
+
+void SoftMaxLayer::configure(uint8_t *inputData, const TensorDescriptor &inputDescr,
+                             const float beta, uint8_t *outputData,
+                             const TensorDescriptor &outputDescr)
+{
+  _inputData.u8 = inputData;
+  _inputDescr = inputDescr;
+  _inputType = inputDescr.type;
+  _outputData.u8 = outputData;
+  _outputDescr = outputDescr;
+  _beta = beta;
+}
+
+void SoftMaxLayer::run()
+{
+  if (_inputType == OperandType::FLOAT32)
+  {
+    softmaxFloat32();
+  }
+  else if (_inputType == OperandType::QUANT8_ASYMM)
+  {
+    softmaxQuant8();
+  }
+}
+
+} // namespace kernel
+} // namespace cpu
+} // namespace backend
+} // namespace neurun