mir_interpreter::SoftmaxImpl< uint8_t > Struct Reference

Static Public Member Functions
static void	run (const mir::TensorVariant &input, int axis, mir::TensorVariant &output)

Detailed Description

Definition at line 77 of file Softmax.cpp.

Member Function Documentation

run()

void mir_interpreter::SoftmaxImpl< uint8_t >::run ( const mir::TensorVariant &  input, int  axis, mir::TensorVariant &  output  )

static

Definition at line 82 of file Softmax.cpp.

{
  const auto &input_type = input.getType();
  const auto &output_type = output.getType();
 
  assert(input_type.isQuantized());
  assert(output_type.isQuantized());
 
  const auto input_shape = input_type.getShape();
 
  assert(input_type.getElementType() == mir::DataType::UINT8);
  assert(axis == input_shape.rank() - 1); // supported only last dim axis
  (void)axis;
 
  double input_scale = input_type.getQuantization().getScale();
  double output_scale = output_type.getQuantization().getScale();
 
  const int trailing_dim = input_shape.rank() - 1;
  int excluding_last_dim = 1;
  for (int32_t i = 0; i < input_shape.rank() - 1; i++)
  {
    excluding_last_dim *= input_shape.dim(i);
  }
  const int last_dim = input_shape.dim(trailing_dim);
 
  const int32_t clamp_max = std::numeric_limits<uint8_t>::max();
  const int32_t clamp_min = std::numeric_limits<uint8_t>::min();
 
  uint8_t *input_data = reinterpret_cast<uint8_t *>(input.atOffset(0));
 
  float table[256];
  PopulateSoftmaxLookupTable(table, input_scale, 1.f);
 
  uint8_t *output_data = reinterpret_cast<uint8_t *>(output.atOffset(0));
 
  for (int i = 0; i < excluding_last_dim; ++i)
  {
    int32_t max_val = std::numeric_limits<uint8_t>::min();
    // Find max quantized value.
    for (int j = 0; j < last_dim; ++j)
    {
      max_val = std::max(max_val, static_cast<int32_t>(input_data[j]));
    }
 
    float sum_exp = 0.0f;
    const int32_t max_uint8 = std::numeric_limits<uint8_t>::max();
    const float *table_offset = &table[max_uint8 - max_val];
    // Calculate normalizer sum(exp(x)).
    for (int j = 0; j < last_dim; ++j)
    {
      sum_exp += table_offset[input_data[j]];
    }
 
    const float inv_sum_exp = 1.0f / (sum_exp * output_scale);
    // Normalize and quantize probabilities.
    for (int j = 0; j < last_dim; ++j)
    {
      const float prob_rescaled = table_offset[input_data[j]] * inv_sum_exp;
      const int32_t prob_quantized = static_cast<int32_t>(prob_rescaled + 0.5);
      output_data[j] =
        static_cast<uint8_t>(std::max(std::min(clamp_max, prob_quantized), clamp_min));
    }
    input_data += last_dim;
    output_data += last_dim;
  }
}

Referenced by package.infer.session::inference().

---

The documentation for this struct was generated from the following file:

• compiler/mir-interpreter/src/ops/Softmax.cpp