Bei-z commented on a change in pull request #8542:
URL: https://github.com/apache/arrow/pull/8542#discussion_r520841038
##########
File path: cpp/src/arrow/util/basic_decimal.cc
##########
@@ -490,49 +527,60 @@ static void FixDivisionSigns(BasicDecimal128* result,
BasicDecimal128* remainder
}
}
-/// \brief Build a BasicDecimal128 from a list of ints.
+/// \brief Build a little endian array of uint64_t from a big endian array of
uint32_t.
+template <size_t N>
+static DecimalStatus BuildFromArray(std::array<uint64_t, N>* result_array,
+ uint32_t* array, int64_t length) {
+ for (int64_t i = length - 2 * N - 1; i >= 0; i--) {
+ if (array[i] != 0) {
+ return DecimalStatus::kOverflow;
+ }
+ }
+ int64_t next_index = length - 1;
+ for (size_t i = 0; i < N; i++) {
+ uint64_t lower_bits = (next_index < 0) ? 0 : array[next_index--];
+ (*result_array)[i] =
+ (next_index < 0)
Review comment:
Thank you for your suggestion!
Changed this part of code to 2 loops, but there is a performance regression.
Is there anything I can improve on the current code to help with the
performance, or should we consider change back to one loop?
##########
File path: cpp/src/arrow/util/basic_decimal.cc
##########
@@ -490,49 +527,60 @@ static void FixDivisionSigns(BasicDecimal128* result,
BasicDecimal128* remainder
}
}
-/// \brief Build a BasicDecimal128 from a list of ints.
+/// \brief Build a little endian array of uint64_t from a big endian array of
uint32_t.
+template <size_t N>
+static DecimalStatus BuildFromArray(std::array<uint64_t, N>* result_array,
+ uint32_t* array, int64_t length) {
Review comment:
Changed accordingly, thank you!
##########
File path: cpp/src/arrow/util/basic_decimal.cc
##########
@@ -451,6 +467,25 @@ static int64_t FillInArray(const BasicDecimal128& value,
uint32_t* array,
return 1;
}
+/// Expands the given value into a big endian array of ints so that we can
work on
+/// it. The array will be converted to an absolute value and the was_negative
+/// flag will be set appropriately. The array will remove leading zeros from
+/// the value.
+/// \param array a big endian array of length 8 to set with the value
+/// \param was_negative a flag for whether the value was original negative
+/// \result the output length of the array
+static int64_t FillInArray(const BasicDecimal256& value, uint32_t* array,
+ bool& was_negative) {
+ BasicDecimal256 positive_value = value;
+ was_negative = false;
+ int64_t highest_bit = positive_value.little_endian_array()[3];
+ if (highest_bit < 0) {
Review comment:
BasicDecimal256 do have a sign method, but that method need shifting of
int64_t value. Would you recommend adding a is_negatvie method?
##########
File path: cpp/src/arrow/util/basic_decimal.cc
##########
@@ -395,33 +395,49 @@ BasicDecimal128& BasicDecimal128::operator*=(const
BasicDecimal128& right) {
return *this;
}
-/// Expands the given value into an array of ints so that we can work on
-/// it. The array will be converted to an absolute value and the wasNegative
+/// Expands the given little endian array of uint64_t into a big endian array
of
+/// uint32_t. The value of input array is expected to be non-negative. The
result_array
+/// will remove leading zeros from the input array.
+/// \param value_array a little endian array to represent the value
+/// \param result_array a big endian array of length N*2 to set with the value
+/// \result the output length of the array
+template <size_t N>
+static int64_t FillInArray(const std::array<uint64_t, N>& value_array,
+ uint32_t* result_array) {
+ int64_t next_index = 0;
+ for (int64_t i = N - 1; i >= 0; i--) {
+ if (value_array[i] != 0) {
+ if (value_array[i] <= std::numeric_limits<uint32_t>::max()) {
+ result_array[next_index++] = static_cast<uint32_t>(value_array[i]);
+ i--;
+ }
+ for (int64_t j = i; j >= 0; j--) {
+ result_array[next_index++] = static_cast<uint32_t>(value_array[j] >>
32);
+ result_array[next_index++] = static_cast<uint32_t>(value_array[j]);
+ }
+ break;
+ }
+ }
+ return next_index;
+}
+
+/// Expands the given value into a big endian array of ints so that we can
work on
+/// it. The array will be converted to an absolute value and the was_negative
/// flag will be set appropriately. The array will remove leading zeros from
/// the value.
-/// \param array an array of length 4 to set with the value
+/// \param array a big endian array of length 4 to set with the value
/// \param was_negative a flag for whether the value was original negative
/// \result the output length of the array
static int64_t FillInArray(const BasicDecimal128& value, uint32_t* array,
bool& was_negative) {
- uint64_t high;
- uint64_t low;
- const int64_t highbits = value.high_bits();
- const uint64_t lowbits = value.low_bits();
-
- if (highbits < 0) {
- low = ~lowbits + 1;
- high = static_cast<uint64_t>(~highbits);
- if (low == 0) {
- ++high;
- }
- was_negative = true;
- } else {
- low = lowbits;
- high = static_cast<uint64_t>(highbits);
- was_negative = false;
- }
-
+ BasicDecimal128 abs_value = BasicDecimal128::Abs(value);
+ was_negative = value.high_bits() < 0;
+ uint64_t high = static_cast<uint64_t>(abs_value.high_bits());
+ uint64_t low = abs_value.low_bits();
+
+ // FillInArray(std::array<uint64_t, N>& value_array, uint32_t* result_array)
is not
+ // called here as the following code has better performance, to avoid
regression on
+ // BasicDecimal128 Division.
if (high != 0) {
if (high > std::numeric_limits<uint32_t>::max()) {
Review comment:
This part of code is from the original version of Decimal128 division,
to eliminate the leading zeros from the uint64_t value (high & low) and into
uint32_t array.
So that to find the first non-negative uint32_t value, we check the high
first, and then low value.
##########
File path: cpp/src/arrow/util/basic_decimal.cc
##########
@@ -395,33 +395,49 @@ BasicDecimal128& BasicDecimal128::operator*=(const
BasicDecimal128& right) {
return *this;
}
-/// Expands the given value into an array of ints so that we can work on
-/// it. The array will be converted to an absolute value and the wasNegative
+/// Expands the given little endian array of uint64_t into a big endian array
of
+/// uint32_t. The value of input array is expected to be non-negative. The
result_array
+/// will remove leading zeros from the input array.
+/// \param value_array a little endian array to represent the value
+/// \param result_array a big endian array of length N*2 to set with the value
+/// \result the output length of the array
+template <size_t N>
+static int64_t FillInArray(const std::array<uint64_t, N>& value_array,
+ uint32_t* result_array) {
+ int64_t next_index = 0;
+ for (int64_t i = N - 1; i >= 0; i--) {
Review comment:
Thank you! Changed to use 2 separate loops.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
