Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Compile-time ipow computation with array lookup #15110

Merged
merged 7 commits into from
Feb 28, 2024
Merged
63 changes: 46 additions & 17 deletions cpp/include/cudf/fixed_point/fixed_point.hpp
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
/*
* Copyright (c) 2020-2023, NVIDIA CORPORATION.
* Copyright (c) 2020-2024, NVIDIA CORPORATION.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
Expand All @@ -22,6 +22,7 @@

#include <cuda/std/limits>
#include <cuda/std/type_traits>
#include <cuda/std/utility>

#include <algorithm>
#include <cassert>
Expand Down Expand Up @@ -82,12 +83,43 @@ constexpr inline auto is_supported_construction_value_type()

// Helper functions for `fixed_point` type
namespace detail {

/**
* @brief A function for integer exponentiation by squaring
* @brief Recursively computes integer exponentiation
*
* https://simple.wikipedia.org/wiki/Exponentiation_by_squaring <br>
* Note: this is the iterative equivalent of the recursive definition (faster) <br>
* Quick-bench: http://quick-bench.com/Wg7o7HYQC9FW5M0CO0wQAjSwP_Y
* Note: This is intended to be run at compile time
davidwendt marked this conversation as resolved.
Show resolved Hide resolved
pmattione-nvidia marked this conversation as resolved.
Show resolved Hide resolved
*
* @tparam Rep Representation type for return type
* @tparam Base The base to be exponentiated
* @param exp The exponent to be used for exponentiation
* @return Result of `Base` to the power of `exponent` of type `Rep`
*/
template <typename Rep, int32_t Base>
CUDF_HOST_DEVICE inline constexpr Rep get_power(int32_t exp)
{
// Compute power recursively
return (exp > 0) ? Rep(Base) * get_power<Rep, Base>(exp - 1) : 1;
}

/**
* @brief Implementation of integer exponentiation by array lookup
*
* @tparam Rep Representation type for return type
* @tparam Base The base to be exponentiated
* @tparam Exponents The exponents for the array entries
* @param exponent The exponent to be used for exponentiation
* @return Result of `Base` to the power of `exponent` of type `Rep`
*/
template <typename Rep, int32_t Base, std::size_t... Exponents>
CUDF_HOST_DEVICE inline Rep ipow_impl(int32_t exponent, cuda::std::index_sequence<Exponents...>)
{
// Compute powers at compile time, storing into array
static constexpr Rep powers[] = {get_power<Rep, Base>(Exponents)...};
return powers[exponent];
}

/**
* @brief A function for integer exponentiation by array lookup
*
* @tparam Rep Representation type for return type
* @tparam Base The base to be exponentiated
Expand All @@ -102,19 +134,16 @@ template <typename Rep,
CUDF_HOST_DEVICE inline Rep ipow(T exponent)
{
cudf_assert(exponent >= 0 && "integer exponentiation with negative exponent is not possible.");
if (exponent == 0) { return static_cast<Rep>(1); }

auto extra = static_cast<Rep>(1);
auto square = static_cast<Rep>(Base);
while (exponent > 1) {
if (exponent & 1 /* odd */) {
extra *= square;
exponent -= 1;
}
exponent /= 2;
square *= square;
if constexpr (Base == numeric::Radix::BASE_2) {
return static_cast<Rep>(1) << exponent;
} else { // BASE_10
// Build index sequence for building power array at compile time
static constexpr auto max_exp = cuda::std::numeric_limits<Rep>::digits10;
static constexpr auto exponents = cuda::std::make_index_sequence<max_exp + 1>{};

// Get compile-time result
return ipow_impl<Rep, static_cast<int32_t>(Base)>(exponent, exponents);
}
return square * extra;
}

/** @brief Function that performs a `right shift` scale "times" on the `val`
Expand Down
7 changes: 4 additions & 3 deletions cpp/include/cudf/round.hpp
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
/*
* Copyright (c) 2020-2023, NVIDIA CORPORATION.
* Copyright (c) 2020-2024, NVIDIA CORPORATION.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
Expand Down Expand Up @@ -32,8 +32,9 @@ namespace cudf {
/**
* @brief Different rounding methods for `cudf::round`
*
* Info on HALF_UP rounding: https://en.wikipedia.org/wiki/Rounding#Round_half_up
* Info on HALF_EVEN rounding: https://en.wikipedia.org/wiki/Rounding#Round_half_to_even
* Info on HALF_EVEN rounding: https://en.wikipedia.org/wiki/Rounding#Rounding_half_to_even
* Info on HALF_UP rounding: https://en.wikipedia.org/wiki/Rounding#Rounding_half_away_from_zero
* Note: HALF_UP means up in MAGNITUDE: Away from zero! Because of how Java and python define it
*/
enum class rounding_method : int32_t { HALF_UP, HALF_EVEN };

Expand Down
5 changes: 3 additions & 2 deletions cpp/tests/strings/fixed_point_tests.cpp
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
/*
* Copyright (c) 2021-2023, NVIDIA CORPORATION.
* Copyright (c) 2021-2024, NVIDIA CORPORATION.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
Expand Down Expand Up @@ -324,7 +324,8 @@ TEST_F(StringsConvertTest, DISABLED_FixedPointStringConversionOperator)
{
auto const max = cuda::std::numeric_limits<__int128_t>::max();

auto const x = numeric::decimal128{max, numeric::scale_type{-10}};
//Must use scaled_integer, else shift (multiply) is undefined behavior (integer overflow)
auto const x = numeric::decimal128(numeric::scaled_integer{max, numeric::scale_type{-10}});
EXPECT_EQ(static_cast<std::string>(x), "17014118346046923173168730371.5884105727");

auto const y = numeric::decimal128{max, numeric::scale_type{10}};
Expand Down
Loading