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.

Sign up for GitHub

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

Fix rounding to zero error in stod on very small float numbers #10672

Merged
merged 4 commits into from
Apr 23, 2022
Merged

Fix rounding to zero error in stod on very small float numbers #10672

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
f6f246c
Fix rounding error in stod on very small float numbers
davidwendt Apr 15, 2022
75db733
Merge branch 'branch-22.06' into stod-exp-min-check
davidwendt Apr 18, 2022
958f3f9
code for supporting subnormal float values from bdice
davidwendt Apr 18, 2022
a10e432
update comments
davidwendt Apr 18, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
21 changes: 16 additions & 5 deletions cpp/src/strings/convert/convert_floats.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -124,16 +124,27 @@ __device__ inline double stod(string_view const& d_str)
exp_ten *= exp_sign;
exp_ten += exp_off;
exp_ten += num_digits - 1;
if (exp_ten > std::numeric_limits<double>::max_exponent10)
if (exp_ten > std::numeric_limits<double>::max_exponent10) {
return sign > 0 ? std::numeric_limits<double>::infinity()
: -std::numeric_limits<double>::infinity();
else if (exp_ten < std::numeric_limits<double>::min_exponent10)
return double{0};
}

double base = sign * static_cast<double>(digits);

exp_ten += 1 - num_digits;
// exp10() is faster than pow(10.0,exp_ten)
// If 10^exp_ten would result in a subnormal value, the base and
// exponent should be adjusted so that 10^exp_ten is a normal value
auto const subnormal_shift = std::numeric_limits<double>::min_exponent10 - exp_ten;
if (subnormal_shift > 0) {
// Handle subnormal values. Ensure that both base and exponent are
// normal values before computing their product.
base = base / exp10(static_cast<double>(num_digits - 1 + subnormal_shift));
exp_ten += num_digits - 1; // adjust exponent
auto const exponent = exp10(static_cast<double>(exp_ten + subnormal_shift));
return base * exponent;
}

double const exponent = exp10(static_cast<double>(std::abs(exp_ten)));
double const base = sign * static_cast<double>(digits);
return exp_ten < 0 ? base / exponent : base * exponent;
}

Expand Down
6 changes: 5 additions & 1 deletion cpp/tests/strings/floats_tests.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -125,11 +125,15 @@ TEST_F(StringsConvertTest, FromFloats32)

TEST_F(StringsConvertTest, ToFloats64)
{
// clang-format off
std::vector<const char*> h_strings{
"1234", nullptr, "-876", "543.2", "-0.12", ".25",
"-.002", "", "-0.0", "1.28e256", "NaN", "abc123",
"123abc", "456e", "-1.78e+5", "-122.33644782", "12e+309", "1.7976931348623159E308",
"-Inf", "-INFINITY"};
"-Inf", "-INFINITY", "1.0", "1.7976931348623157e+308", "1.7976931348623157e-307",
// subnormal numbers: v--- smallest double v--- result is 0
"4e-308", "3.3333333333e-320", "4.940656458412465441765688e-324", "1.e-324" };
davidwendt marked this conversation as resolved.
Show resolved Hide resolved
// clang-format on
cudf::test::strings_column_wrapper strings(
h_strings.begin(),
h_strings.end(),
Expand Down