CircleConst.cpp

Go to the documentation of this file.

/*
 * Copyright (c) 2020 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 "luci/Import/Nodes/CircleConst.h"
 
#include "luci/Import/CircleReader.h"
 
#include <luci/IR/Nodes/CircleConst.h>
#include <luci/Log.h>
 
#include <loco.h>
#include <oops/UserExn.h>
 
#include <cassert>
#include <limits>
#include <ostream>
#include <string>
#include <vector>
 
#include <string.h>
 
namespace
{
 
std::ostream &operator<<(std::ostream &os, const luci::VectorWrapper<int32_t> &vect)
{
  uint32_t seq = 0;
  for (const auto &v : vect)
  {
    if (seq)
      os << ", ";
    os << v;
    seq++;
  }
  return os;
}
 
using namespace luci;
 
template <loco::DataType DT>
void copy_data(const VectorWrapper<uint8_t> &raw_data, uint32_t num_elements,
               CircleConst *const_node)
{
  using T = typename loco::DataTypeImpl<DT>::Type;
 
  // TODO calculate the exact buffer size of sparse tensor
  if (const_node->sparsityparam())
  {
    num_elements = raw_data.size() / sizeof(T);
  }
 
  assert(raw_data.size() == num_elements * sizeof(T));
  const auto *data = reinterpret_cast<const T *>(raw_data.data());
 
  const_node->size<DT>(num_elements);
  for (uint32_t i = 0; i < num_elements; ++i)
  {
    const_node->at<DT>(i) = data[i];
  }
}
 
template <>
void copy_data<loco::DataType::STRING>(const VectorWrapper<uint8_t> &raw_data,
                                       uint32_t num_elements, CircleConst *const_node)
{
  assert(const_node->sparsityparam() == nullptr);
 
  const auto *data = reinterpret_cast<const char *>(raw_data.data());
  const auto *i32d = reinterpret_cast<const int32_t *>(raw_data.data());
 
  // de-serialize string data
  //   int32_t count
  //   int32_t offsets[count + 1]
  //   string  values[count]
  assert(static_cast<uint32_t>(*i32d) == num_elements);
  i32d++; // skip count
 
  std::vector<int32_t> offsets;
  offsets.push_back(*i32d++);
  for (uint32_t i = 0; i < num_elements; ++i)
  {
    offsets.push_back(*i32d++);
  }
  assert(offsets.size() == num_elements + 1);
 
  const_node->size<loco::DataType::STRING>(num_elements);
  for (uint32_t i = 0; i < num_elements; ++i)
  {
    int32_t start = offsets[i];
    int32_t next = offsets[i + 1];
 
    std::string value(data + start, next - start);
    const_node->at<loco::DataType::STRING>(i) = value;
  }
}
 
// NOTE copy_data for S4, U4.
//      this method will unpack two 4bit elements, packed in 8bit,
//      to two 8bit elements, having values -8~7, for S4 and 0~15 for U4.
template <loco::DataType DT>
void copy_data_4(const VectorWrapper<uint8_t> &raw_data, uint32_t num_elements,
                 CircleConst *const_node)
{
  using T = typename loco::DataTypeImpl<DT>::Type;
 
  // TODO support sparse?
  assert(const_node->sparsityparam() == nullptr);
  if (const_node->sparsityparam())
    return;
 
  uint32_t raw_size = (num_elements + 1) / 2;
  assert(raw_data.size() == raw_size);
 
  const uint8_t *data = raw_data.data();
  const_node->size<DT>(num_elements);
  for (uint32_t i = 0; i < raw_size; ++i)
  {
    uint32_t idx = i * 2;
    // for S4, 1bit for sign, 3bit for value
    const_node->at<DT>(idx) = static_cast<T>(data[i] << 4) >> 4;
    if (idx < num_elements)
    {
      idx++;
      const_node->at<DT>(idx) = static_cast<T>(data[i]) >> 4;
    }
  }
}
 
} // namespace
 
namespace luci
{
 
CircleNode *CircleConstNodeBuilder::build(TensorIndex tensor_index,
                                          GraphBuilderContext *context) const
{
  assert(tensor_index >= 0);
  LOGGER(l);
 
  auto graph = context->graph();
  auto reader = context->reader();
  const auto tensors = reader->tensors();
  const auto const_tensor = tensors[tensor_index];
  assert(const_tensor != nullptr);
  if (const_tensor->is_variable())
  {
    // Create CircleVariable for variable
    return nullptr;
  }
 
  const auto r_buffers = reader->buffers();
  const auto c_buffer = const_tensor->buffer();
  const auto r_buffer = r_buffers[c_buffer];
  assert(r_buffer != nullptr);
  if (r_buffer->offset() == 1 || r_buffer->size() == 1)
  {
    // NOTE this shouldn't happen
    throw std::runtime_error("CircleConst: Circle file with invalid extended Buffer.");
  }
  // temporary buffer to provide raw data from file
  // must have life time same or longer than 'buffer' variable
  std::vector<uint8_t> temp_buffer;
  luci::VectorWrapper<uint8_t> buffer(nullptr);
  if (r_buffer->offset() > 1)
  {
    if (r_buffer->size() >= std::numeric_limits<uint32_t>::max())
    {
      // NOTE uint32_t limit is to match "uoffset_t flatbuffers::Vector::size()"
      throw std::runtime_error("CircleConst: Circle file with invalid extended Buffer.");
    }
    uint32_t r_size = static_cast<uint32_t>(r_buffer->size());
    // match binary level to flatbuffers::Vector
    temp_buffer.resize(r_size + sizeof(uint32_t));
 
    uint8_t *t_data = temp_buffer.data();
    const uint8_t *f_data = reader->file_data(r_buffer->offset());
    if (f_data == nullptr)
    {
      // NOTE this shouldn't happen
      assert(false);
      return nullptr;
    }
    memcpy(t_data, &r_size, sizeof(r_size));
    t_data = t_data + sizeof(r_size);
    if (r_buffer->offset() + r_buffer->size() > reader->file_size())
    {
      // NOTE this shouldn't happen
      assert(false);
      return nullptr;
    }
    memcpy(t_data, f_data, r_buffer->size());
 
    using fbv_t = flatbuffers::Vector<uint8_t>;
    const fbv_t *v_data = reinterpret_cast<const fbv_t *>(temp_buffer.data());
    buffer = wrap(v_data);
 
    context->ext_buffer(true);
  }
  else
  {
    buffer = wrap(r_buffer->data());
  }
  const auto const_dims = wrap(const_tensor->shape()); // in NHWC
  if (const_dims.size() == 0 && buffer.empty())
  {
    // unknown shape tensor and scalar tensor
    return nullptr;
  }
 
  // if tensor_index is used as output to some other operator, this is not a constant
  auto tensoroutputs = context->tensoroutputs();
  if (tensoroutputs->find(tensor_index))
  {
    // other operator output tensor
    return nullptr;
  }
 
  uint32_t num_elements = 1;
  for (uint32_t r = 0; r < const_dims.size(); ++r)
  {
    num_elements = num_elements * const_dims[r];
  }
 
  if (buffer.empty() && num_elements > 0)
  {
    // normal empty tensor
    return nullptr;
  }
 
  auto const_node = graph->nodes()->create<CircleConst>();
  copy_tensor_attributes(const_tensor, const_node);
  const_node->shape_status(luci::ShapeStatus::VALID);
  INFO(l) << "[luci] NodeFinder const_node(" << tensor_index << ") -> " << const_node << " "
          << const_dims << std::endl;
  if (num_elements > 0)
  {
    switch (luci_datatype(const_tensor->type()))
    {
      case loco::DataType::FLOAT32:
        copy_data<loco::DataType::FLOAT32>(buffer, num_elements, const_node);
        break;
 
      case loco::DataType::FLOAT16:
        copy_data<loco::DataType::FLOAT16>(buffer, num_elements, const_node);
        break;
 
      case loco::DataType::U4:
        copy_data_4<loco::DataType::U4>(buffer, num_elements, const_node);
        break;
 
      case loco::DataType::U8:
        copy_data<loco::DataType::U8>(buffer, num_elements, const_node);
        break;
 
      case loco::DataType::S4:
        copy_data_4<loco::DataType::S4>(buffer, num_elements, const_node);
        break;
 
      case loco::DataType::S8:
        copy_data<loco::DataType::S8>(buffer, num_elements, const_node);
        break;
 
      case loco::DataType::S16:
        copy_data<loco::DataType::S16>(buffer, num_elements, const_node);
        break;
 
      case loco::DataType::S32:
        copy_data<loco::DataType::S32>(buffer, num_elements, const_node);
        break;
 
      case loco::DataType::S64:
        copy_data<loco::DataType::S64>(buffer, num_elements, const_node);
        break;
 
      case loco::DataType::BOOL:
        copy_data<loco::DataType::BOOL>(buffer, num_elements, const_node);
        break;
 
      case loco::DataType::STRING:
        copy_data<loco::DataType::STRING>(buffer, num_elements, const_node);
        break;
 
      default:
        throw oops::UserExn("Unsupported tensor type",
                            circle::EnumNameTensorType(const_tensor->type()));
    }
  }
 
  return const_node;
}
 
} // namespace luci