ExtensionFunctionsBinding.cpp File Reference

#include "ExtensionFunctionsBinding.h"
#include <algorithm>
#include "ExternalExecutor.h"

Include dependency graph for ExtensionFunctionsBinding.cpp:

Go to the source code of this file.

Namespaces
	anonymous_namespace{ExtensionFunctionsBinding.cpp}

Functions
ExtArgumentType	anonymous_namespace{ExtensionFunctionsBinding.cpp}::get_column_arg_elem_type (const ExtArgumentType ext_arg_column_type)
ExtArgumentType	anonymous_namespace{ExtensionFunctionsBinding.cpp}::get_column_list_arg_elem_type (const ExtArgumentType ext_arg_column_list_type)
ExtArgumentType	anonymous_namespace{ExtensionFunctionsBinding.cpp}::get_array_arg_elem_type (const ExtArgumentType ext_arg_array_type)
static int	anonymous_namespace{ExtensionFunctionsBinding.cpp}::match_numeric_argument (const SQLTypeInfo &arg_type_info, const bool is_arg_literal, const ExtArgumentType &sig_ext_arg_type, int32_t &penalty_score)
static int	anonymous_namespace{ExtensionFunctionsBinding.cpp}::match_arguments (const SQLTypeInfo &arg_type, const bool is_arg_literal, int sig_pos, const std::vector< ExtArgumentType > &sig_types, int &penalty_score)
bool	anonymous_namespace{ExtensionFunctionsBinding.cpp}::is_valid_identifier (std::string str)
template<typename T >
std::tuple< T, std::vector < SQLTypeInfo > >	bind_function (std::string name, Analyzer::ExpressionPtrVector func_args, const std::vector< T > &ext_funcs, const std::string processor)
const std::tuple < table_functions::TableFunction, std::vector< SQLTypeInfo > >	bind_table_function (std::string name, Analyzer::ExpressionPtrVector input_args, const std::vector< table_functions::TableFunction > &table_funcs, const bool is_gpu)
ExtensionFunction	bind_function (std::string name, Analyzer::ExpressionPtrVector func_args)
ExtensionFunction	bind_function (std::string name, Analyzer::ExpressionPtrVector func_args, const bool is_gpu)
ExtensionFunction	bind_function (const Analyzer::FunctionOper *function_oper, const bool is_gpu)
const std::tuple < table_functions::TableFunction, std::vector< SQLTypeInfo > >	bind_table_function (std::string name, Analyzer::ExpressionPtrVector input_args, const bool is_gpu)

Function Documentation

template<typename T >

std::tuple<T, std::vector<SQLTypeInfo> > bind_function	(	std::string	name,
		Analyzer::ExpressionPtrVector	func_args,
		const std::vector< T > &	ext_funcs,
		const std::string	processor
	)

Definition at line 515 of file ExtensionFunctionsBinding.cpp.

References CHECK, CHECK_EQ, CHECK_LE, DEFAULT_ROW_MULTIPLIER_SUFFIX, ext_arg_type_to_type_info(), logger::FATAL, generate_column_list_type(), generate_column_type(), SQLTypeInfo::get_type(), is_ext_arg_type_column_list(), anonymous_namespace{ExtensionFunctionsBinding.cpp}::is_valid_identifier(), kCOLUMN_LIST, kINT, kNULLT, kTEXT, LOG, anonymous_namespace{ExtensionFunctionsBinding.cpp}::match_arguments(), setup::name, heavydb.dtypes::T, to_string(), SQLTypeInfo::to_string(), ExtensionFunctionsWhitelist::toString(), and UNREACHABLE.

Referenced by bind_function(), CodeGenerator::codegenFunctionOper(), and RelAlgTranslator::translateFunction().

                               {
  /* worker function

     Template type T must implement the following methods:

       std::vector<ExtArgumentType> getInputArgs()
   */
  /*
    Return extension function/table function that has the following
    properties

    1. each argument type in `arg_types` matches with extension
       function argument types.

       For scalar types, the matching means that the types are either
       equal or the argument type is smaller than the corresponding
       the extension function argument type. This ensures that no
       information is lost when casting of argument values is
       required.

       For array and geo types, the matching means that the argument
       type matches exactly with a group of extension function
       argument types. See `match_arguments`.

    2. has minimal penalty score among all implementations of the
       extension function with given `name`, see `get_penalty_score`
       for the definition of penalty score.

    It is assumed that function_oper and extension functions in
    ext_funcs have the same name.
   */
  if (!is_valid_identifier(name)) {
    throw NativeExecutionError(
        "Cannot bind function with invalid UDF/UDTF function name: " + name);
  }

  int minimal_score = std::numeric_limits<int>::max();
  int index = -1;
  int optimal = -1;
  int optimal_variant = -1;

  std::vector<SQLTypeInfo> type_infos_input;
  std::vector<bool> args_are_constants;
  for (auto atype : func_args) {
    if constexpr (std::is_same_v<T, table_functions::TableFunction>) {
      if (dynamic_cast<const Analyzer::ColumnVar*>(atype.get())) {
        SQLTypeInfo type_info = atype->get_type_info();
        auto ti = generate_column_type(type_info);
        if (ti.get_subtype() == kNULLT) {
          throw std::runtime_error(std::string(__FILE__) + "#" +
                                   std::to_string(__LINE__) +
                                   ": column support for type info " +
                                   type_info.to_string() + " is not implemented");
        }
        type_infos_input.push_back(ti);
        args_are_constants.push_back(type_info.get_type() != kTEXT);
        continue;
      }
    }
    type_infos_input.push_back(atype->get_type_info());
    if (dynamic_cast<const Analyzer::Constant*>(atype.get())) {
      args_are_constants.push_back(true);
    } else {
      args_are_constants.push_back(false);
    }
  }
  CHECK_EQ(type_infos_input.size(), args_are_constants.size());

  if (type_infos_input.size() == 0 && ext_funcs.size() > 0) {
    CHECK_EQ(ext_funcs.size(), static_cast<size_t>(1));
    CHECK_EQ(ext_funcs[0].getInputArgs().size(), static_cast<size_t>(0));
    if constexpr (std::is_same_v<T, table_functions::TableFunction>) {
      CHECK(ext_funcs[0].hasNonUserSpecifiedOutputSize());
    }
    std::vector<SQLTypeInfo> empty_type_info_variant(0);
    return {ext_funcs[0], empty_type_info_variant};
  }

  // clang-format off
  /*
    Table functions may have arguments such as ColumnList that collect
    neighboring columns with the same data type into a single object.
    Here we compute all possible combinations of mapping a subset of
    columns into columns sets. For example, if the types of function
    arguments are (as given in func_args argument)

      (Column<int>, Column<int>, Column<int>, int)

    then the computed variants will be

      (Column<int>, Column<int>, Column<int>, int)
      (Column<int>, Column<int>, ColumnList[1]<int>, int)
      (Column<int>, ColumnList[1]<int>, Column<int>, int)
      (Column<int>, ColumnList[2]<int>, int)
      (ColumnList[1]<int>, Column<int>, Column<int>, int)
      (ColumnList[1]<int>, Column<int>, ColumnList[1]<int>, int)
      (ColumnList[2]<int>, Column<int>, int)
      (ColumnList[3]<int>, int)

    where the integers in [..] indicate the number of collected
    columns. In the SQLTypeInfo instance, this number is stored in the
    SQLTypeInfo dimension attribute.

    As an example, let us consider a SQL query containing the
    following expression calling a UDTF foo:

      table(foo(cursor(select a, b, c from tableofints), 1))

    Here follows a list of table functions and the corresponding
    optimal argument type variants that are computed for the given
    query expression:

    UDTF:  foo(ColumnList<int>, RowMultiplier) -> Column<int>
           (ColumnList[3]<int>, int)               # a, b, c are all collected to column_list

    UDTF:  foo(Column<int>, ColumnList<int>, RowMultiplier) -> Column<int>
           (Column<int>, ColumnList[2]<int>, int)  # b and c are collected to column_list

    UDTF:  foo(Column<int>, Column<int>, Column<int>, RowMultiplier) -> Column<int>
           (Column<int>, Column<int>, Column<int>, int)
   */
  // clang-format on

  // We first check if any of the matched extension functions
  // in the ext_funcs list allow for column lists, as if they do not,
  // we do not need to account for possible input permutations with
  // ColumnList arguments, which currently can be slow time
  // to match with the extension functions if the number of arguments
  // is high

  // Todo: Develop faster matching algorithm to avoid such performance
  // hits when ColumnLists are allowed

  bool ext_funcs_allow_column_lists{false};
  for (const auto& ext_func : ext_funcs) {
    auto ext_func_args = ext_func.getInputArgs();
    for (const auto& arg : ext_func_args) {
      if (is_ext_arg_type_column_list(arg)) {
        ext_funcs_allow_column_lists = true;
        break;
      }
    }
    if (ext_funcs_allow_column_lists) {
      break;
    }
  }

  std::vector<std::vector<SQLTypeInfo>> type_infos_variants;
  if (ext_funcs_allow_column_lists) {
    for (const auto& ti : type_infos_input) {
      if (type_infos_variants.begin() == type_infos_variants.end()) {
        type_infos_variants.push_back({ti});
        if constexpr (std::is_same_v<T, table_functions::TableFunction>) {
          if (ti.is_column()) {
            auto mti = generate_column_list_type(ti);
            if (mti.get_subtype() == kNULLT) {
              continue;  // skip unsupported element type.
            }
            mti.set_dimension(1);
            type_infos_variants.push_back({mti});
          }
        }
        continue;
      }
      std::vector<std::vector<SQLTypeInfo>> new_type_infos_variants;
      for (auto& type_infos : type_infos_variants) {
        if constexpr (std::is_same_v<T, table_functions::TableFunction>) {
          if (ti.is_column()) {
            auto new_type_infos = type_infos;  // makes a copy
            const auto& last = type_infos.back();
            if (last.is_column_list() && last.has_same_itemtype(ti)) {
              // last column_list consumes column argument if item types match
              new_type_infos.back().set_dimension(last.get_dimension() + 1);
            } else {
              // add column as column_list argument
              auto mti = generate_column_list_type(ti);
              if (mti.get_subtype() == kNULLT) {
                // skip unsupported element type
                type_infos.push_back(ti);
                continue;
              }
              mti.set_dimension(1);
              new_type_infos.push_back(mti);
            }
            new_type_infos_variants.push_back(new_type_infos);
          }
        }
        type_infos.push_back(ti);
      }
      type_infos_variants.insert(type_infos_variants.end(),
                                 new_type_infos_variants.begin(),
                                 new_type_infos_variants.end());
    }
  } else {
    type_infos_variants.emplace_back(type_infos_input);
  }

  // Find extension function that gives the best match on the set of
  // argument type variants:
  for (const auto& ext_func : ext_funcs) {
    index++;

    const auto& ext_func_args = ext_func.getInputArgs();
    int index_variant = -1;
    for (const auto& type_infos : type_infos_variants) {
      index_variant++;
      int penalty_score = 0;
      int pos = 0;
      int original_input_idx = 0;
      CHECK_LE(type_infos.size(), args_are_constants.size());
      for (const auto& ti : type_infos) {
        int offset = match_arguments(ti,
                                     args_are_constants[original_input_idx],
                                     pos,
                                     ext_func_args,
                                     penalty_score);
        if (offset < 0) {
          // atype does not match with ext_func argument
          pos = -1;
          break;
        }
        if (ti.get_type() == kCOLUMN_LIST) {
          original_input_idx += ti.get_dimension();
        } else {
          original_input_idx++;
        }
        pos += offset;
      }

      if ((size_t)pos == ext_func_args.size()) {
        CHECK_EQ(args_are_constants.size(), original_input_idx);
        // prefer smaller return types
        penalty_score += ext_arg_type_to_type_info(ext_func.getRet()).get_logical_size();
        if (penalty_score < minimal_score) {
          optimal = index;
          minimal_score = penalty_score;
          optimal_variant = index_variant;
        }
      }
    }
  }

  if (optimal == -1) {
    /* no extension function found that argument types would match
       with types in `arg_types` */
    auto sarg_types = ExtensionFunctionsWhitelist::toString(type_infos_input);
    std::string message;
    if (!ext_funcs.size()) {
      message = "Function " + name + "(" + sarg_types + ") not supported.";
      throw ExtensionFunctionBindingError(message);
    } else {
      if constexpr (std::is_same_v<T, table_functions::TableFunction>) {
        message = "Could not bind " + name + "(" + sarg_types + ") to any " + processor +
                  " UDTF implementation.";
      } else if constexpr (std::is_same_v<T, ExtensionFunction>) {
        message = "Could not bind " + name + "(" + sarg_types + ") to any " + processor +
                  " UDF implementation.";
      } else {
        LOG(FATAL) << "bind_function: unknown extension function type "
                   << typeid(T).name();
      }
      message += "\n  Existing extension function implementations:";
      for (const auto& ext_func : ext_funcs) {
        // Do not show functions missing the sizer argument
        if constexpr (std::is_same_v<T, table_functions::TableFunction>)
          if (ext_func.useDefaultSizer())
            continue;
        message += "\n    " + ext_func.toStringSQL();
      }
    }
    throw ExtensionFunctionBindingError(message);
  }

  // Functions with "_default_" suffix only exist for calcite
  if constexpr (std::is_same_v<T, table_functions::TableFunction>) {
    if (ext_funcs[optimal].hasUserSpecifiedOutputSizeMultiplier() &&
        ext_funcs[optimal].useDefaultSizer()) {
      std::string name = ext_funcs[optimal].getName();
      name.erase(name.find(DEFAULT_ROW_MULTIPLIER_SUFFIX),
                 sizeof(DEFAULT_ROW_MULTIPLIER_SUFFIX));
      for (size_t i = 0; i < ext_funcs.size(); i++) {
        if (ext_funcs[i].getName() == name) {
          optimal = i;
          std::vector<SQLTypeInfo> type_info = type_infos_variants[optimal_variant];
          size_t sizer = ext_funcs[optimal].getOutputRowSizeParameter();
          type_info.insert(type_info.begin() + sizer - 1, SQLTypeInfo(kINT, true));
          return {ext_funcs[optimal], type_info};
        }
      }
      UNREACHABLE();
    }
  }

  return {ext_funcs[optimal], type_infos_variants[optimal_variant]};
}

Here is the call graph for this function:

Here is the caller graph for this function:

ExtensionFunction bind_function	(	std::string	name,
		Analyzer::ExpressionPtrVector	func_args
	)

Definition at line 825 of file ExtensionFunctionsBinding.cpp.

References ExtensionFunctionsWhitelist::get_ext_funcs(), and setup::name.

                                                                       {
  // used in RelAlgTranslator.cpp, first try GPU UDFs, then fall back
  // to CPU UDFs.
  bool is_gpu = true;
  std::string processor = "GPU";
  auto ext_funcs = ExtensionFunctionsWhitelist::get_ext_funcs(name, is_gpu);
  if (!ext_funcs.size()) {
    is_gpu = false;
    processor = "CPU";
    ext_funcs = ExtensionFunctionsWhitelist::get_ext_funcs(name, is_gpu);
  }
  try {
    return std::get<0>(
        bind_function<ExtensionFunction>(name, func_args, ext_funcs, processor));
  } catch (ExtensionFunctionBindingError& e) {
    if (is_gpu) {
      is_gpu = false;
      processor = "GPU|CPU";
      ext_funcs = ExtensionFunctionsWhitelist::get_ext_funcs(name, is_gpu);
      return std::get<0>(
          bind_function<ExtensionFunction>(name, func_args, ext_funcs, processor));
    } else {
      throw;
    }
  }
}

Here is the call graph for this function:

ExtensionFunction bind_function	(	std::string	name,
		Analyzer::ExpressionPtrVector	func_args,
		const bool	is_gpu
	)

Definition at line 853 of file ExtensionFunctionsBinding.cpp.

References ExtensionFunctionsWhitelist::get_ext_funcs(), and setup::name.

                                                   {
  // used below
  std::vector<ExtensionFunction> ext_funcs =
      ExtensionFunctionsWhitelist::get_ext_funcs(name, is_gpu);
  std::string processor = (is_gpu ? "GPU" : "CPU");
  return std::get<0>(
      bind_function<ExtensionFunction>(name, func_args, ext_funcs, processor));
}

Here is the call graph for this function:

ExtensionFunction bind_function	(	const Analyzer::FunctionOper *	function_oper,
		const bool	is_gpu
	)

Definition at line 864 of file ExtensionFunctionsBinding.cpp.

References bind_function(), Analyzer::FunctionOper::getArity(), Analyzer::FunctionOper::getName(), Analyzer::FunctionOper::getOwnArg(), and setup::name.

                                                   {
  // used in ExtensionsIR.cpp
  auto name = function_oper->getName();
  Analyzer::ExpressionPtrVector func_args = {};
  for (size_t i = 0; i < function_oper->getArity(); ++i) {
    func_args.push_back(function_oper->getOwnArg(i));
  }
  return bind_function(name, func_args, is_gpu);
}

Here is the call graph for this function:

const std::tuple<table_functions::TableFunction, std::vector<SQLTypeInfo> > bind_table_function	(	std::string	name,
		Analyzer::ExpressionPtrVector	input_args,
		const std::vector< table_functions::TableFunction > &	table_funcs,
		const bool	is_gpu
	)

Definition at line 816 of file ExtensionFunctionsBinding.cpp.

References setup::name.

Referenced by bind_table_function(), and RelAlgExecutor::createTableFunctionWorkUnit().

                                       {
  std::string processor = (is_gpu ? "GPU" : "CPU");
  return bind_function<table_functions::TableFunction>(
      name, input_args, table_funcs, processor);
}

Here is the caller graph for this function:

const std::tuple<table_functions::TableFunction, std::vector<SQLTypeInfo> > bind_table_function	(	std::string	name,
		Analyzer::ExpressionPtrVector	input_args,
		const bool	is_gpu
	)

Definition at line 876 of file ExtensionFunctionsBinding.cpp.

References bind_table_function(), and table_functions::TableFunctionsFactory::get_table_funcs().

                                       {
  // used in RelAlgExecutor.cpp
  std::vector<table_functions::TableFunction> table_funcs =
      table_functions::TableFunctionsFactory::get_table_funcs(name, is_gpu);
  return bind_table_function(name, input_args, table_funcs, is_gpu);
}

Here is the call graph for this function: