Skip to content

Revert PR #76: NTT-CRT zero-fill change (no certifiable net win) #77

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

Merged
mmurshed merged 1 commit into from
May 31, 2026
Merged

Revert PR #76: NTT-CRT zero-fill change (no certifiable net win) #77

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
27 changes: 5 additions & 22 deletions include/biginteger/algorithms/multiplication/NTTMultiplicationCrt.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1207,25 +1207,11 @@ namespace BigMath
ULong coeffCount = aCoeffSize + bCoeffSize - 1;
Int n = (Int)std::bit_ceil(coeffCount);

// Three parallel transforms. PackOperand writes the coefficient head
// [0, *CoeffSize) of each buffer; only the zero-pad tail needs clearing.
// Using assign(n, 0) here would zero the whole buffer and then have
// PackOperand overwrite the head — double-writing ~aCoeffSize+bCoeffSize
// elements per call. On a bandwidth-bound kernel that waste is real, so
// resize (no fill at steady state, size already n) + tail-only fill.
// Three parallel transforms.
static thread_local std::vector<UInt> fa1, fb1, fa2, fb2, fa3, fb3;
fa1.resize(n); fb1.resize(n);
fa2.resize(n); fb2.resize(n);
fa3.resize(n); fb3.resize(n);

const std::ptrdiff_t aHead = (std::ptrdiff_t)aCoeffSize;
const std::ptrdiff_t bHead = (std::ptrdiff_t)bCoeffSize;
std::fill(fa1.begin() + aHead, fa1.end(), 0u);
std::fill(fa2.begin() + aHead, fa2.end(), 0u);
std::fill(fa3.begin() + aHead, fa3.end(), 0u);
std::fill(fb1.begin() + bHead, fb1.end(), 0u);
std::fill(fb2.begin() + bHead, fb2.end(), 0u);
std::fill(fb3.begin() + bHead, fb3.end(), 0u);
fa1.assign(n, 0); fb1.assign(n, 0);
fa2.assign(n, 0); fb2.assign(n, 0);
fa3.assign(n, 0); fb3.assign(n, 0);

PackOperand(a, base, fa1, fa2, fa3);
PackOperand(b, base, fb1, fb2, fb3);
Expand All @@ -1245,10 +1231,7 @@ namespace BigMath
MfaPlanTree<F3> tree3;
if (useMfa)
{
// No zero-fill: each scratch is fully written by the step-1 Transpose
// before any read (recursive sub-calls likewise write their slice
// before reading). assign(n, 0) here was ~6n wasted writes per call.
for (int i = 0; i < 6; ++i) mfaScratch[i].resize(n);
for (int i = 0; i < 6; ++i) mfaScratch[i].assign(n, 0);
// Pre-warm all plans in main thread: worker threads cannot call
// GetPlan() safely from inside ParallelDo (BuildRoots reenters pool).
BuildMfaPlanTree<F1, G1>(n, tree1);
Expand Down
Loading