-
Notifications
You must be signed in to change notification settings - Fork 11.8k
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
[flang][runtime] Use std::fmod for most MOD/MODULO #78745
Conversation
@llvm/pr-subscribers-flang-runtime Author: Peter Klausler (klausler) ChangesThe new accurate algorithm for real MOD and MODULO in the runtime is not as fast as std::fmod(), which is also accurate. So use std::fmod() for those floating-point types that it supports. Fixes #78641. Full diff: https://github.com/llvm/llvm-project/pull/78745.diff 1 Files Affected:
diff --git a/flang/runtime/numeric.cpp b/flang/runtime/numeric.cpp
index 3f6f553e7bb554d..ad6b0e854522496 100644
--- a/flang/runtime/numeric.cpp
+++ b/flang/runtime/numeric.cpp
@@ -145,25 +145,33 @@ inline RT_API_ATTRS T RealMod(
} else if (std::isinf(p)) {
return a;
} else {
- // The standard defines MOD(a,p)=a-AINT(a/p)*p and
- // MODULO(a,p)=a-FLOOR(a/p)*p, but those definitions lose
- // precision badly due to cancellation when ABS(a) is
- // much larger than ABS(p).
- // Insights:
- // - MOD(a,p)=MOD(a-n*p,p) when a>0, p>0, integer n>0, and a>=n*p
- // - when n is a power of two, n*p is exact
- // - as a>=n*p, a-n*p does not round.
- // So repeatedly reduce a by all n*p in decreasing order of n;
- // what's left is the desired remainder. This is basically
- // the same algorithm as arbitrary precision binary long division,
- // discarding the quotient.
T tmp{std::abs(a)};
T pAbs{std::abs(p)};
- for (T adj{SetExponent(pAbs, Exponent<int>(tmp))}; tmp >= pAbs; adj /= 2) {
- if (tmp >= adj) {
- tmp -= adj;
- if (tmp == 0) {
- break;
+ if (tmp < pAbs) {
+ } else if constexpr (std::is_same_v<T, float> ||
+ std::is_same_v<T, double> || std::is_same_v<T, long double>) {
+ tmp = std::fmod(tmp, pAbs);
+ } else {
+ // The standard defines MOD(a,p)=a-AINT(a/p)*p and
+ // MODULO(a,p)=a-FLOOR(a/p)*p, but those definitions lose
+ // precision badly due to cancellation when ABS(a) is
+ // much larger than ABS(p) and the values are not
+ // integers
+ // Insights:
+ // - MOD(a,p)=MOD(a-n*p,p) when a>0, p>0, integer n>0, and a>=n*p
+ // - when n is a power of two, n*p is exact
+ // - as a>=n*p, a-n*p does not round.
+ // So repeatedly reduce a by all n*p in decreasing order of n;
+ // what's left is the desired remainder. This is basically
+ // the same algorithm as arbitrary precision binary long division,
+ // discarding the quotient.
+ for (T adj{SetExponent(pAbs, Exponent<int>(tmp))}; tmp >= pAbs;
+ adj /= 2) {
+ if (tmp >= adj) {
+ tmp -= adj;
+ if (tmp == 0) {
+ break;
+ }
}
}
}
|
flang/runtime/numeric.cpp
Outdated
if (tmp < pAbs) { | ||
} else if constexpr (std::is_same_v<T, float> || | ||
std::is_same_v<T, double> || std::is_same_v<T, long double>) { | ||
tmp = std::fmod(tmp, pAbs); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Maybe we should just use fmod(a, p)
for these types, and avoid the abs
computations, tmp < pAbs
evaluation, and the tmp = -tmp
fixup below. I think this may remove some overhead for the likely cases (e.g. when a >= p
). What do you think?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Getting the result right for both MOD and MODULO when either a or p or both are negative is tricky. Will try.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Updated.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thank you, Peter! It looks good to me, though, there might be some unneeded overhead (see my comment).
I suggest merging this, and I will check out performance after. It looks like using fmod
is the standard way in other compilers as well.
The new accurate algorithm for real MOD and MODULO in the runtime is not as fast as std::fmod(), which is also accurate. So use std::fmod() for those floating-point types that it supports. Fixes llvm#78641.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thank you!
The new accurate algorithm for real MOD and MODULO in the runtime is not as fast as std::fmod(), which is also accurate. So use std::fmod() for those floating-point types that it supports.
Fixes #78641.