happyz
/
gemma.cpp
mirror of https://github.com/google/gemma.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Fix unaligned buffer causing crash on GCC. Thanks @ufownl, fixes #508 

PiperOrigin-RevId: 741590339
This commit is contained in:
Jan Wassenberg 2025-03-28 11:24:53 -07:00 committed by Copybara-Service
parent e55734219d
commit 76a81ac2d6
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
ops/matmul-inl.h
View File

@ -111,7 +111,7 @@ class MMStoreHorizontalSumsIntoC {
VF C30, VF C31, VF C32, VF C33, // VF C30, VF C31, VF C32, VF C33, //
const size_t row_c, const size_t col_c, const size_t row_c, const size_t col_c,
const MMArgs& args, const RowPtr<TC>& C) const { const MMArgs& args, const RowPtr<TC>& C) const {
float buf[16 * hn::MaxLanes(df)]; HWY_ALIGN float buf[16 * hn::MaxLanes(df)];
const size_t N = hn::Lanes(df); const size_t N = hn::Lanes(df);
// Horizontal reductions (`ReduceSum`) are rather expensive, entailing // Horizontal reductions (`ReduceSum`) are rather expensive, entailing
// log(N) operations for vectors of length N. Because `kNR` == 4, we // log(N) operations for vectors of length N. Because `kNR` == 4, we
@ -226,7 +226,7 @@ class MMAddHorizontalSumsIntoPartial {
static_assert(HWY_HAVE_FLOAT64, "Disable Armv7 NEON: we require fp64"); static_assert(HWY_HAVE_FLOAT64, "Disable Armv7 NEON: we require fp64");
const hn::Repartition<double, DF> dd; const hn::Repartition<double, DF> dd;
double buf[16 * hn::MaxLanes(dd)]; HWY_ALIGN double buf[16 * hn::MaxLanes(dd)];
using VD = hn::Vec<decltype(dd)>; using VD = hn::Vec<decltype(dd)>;
const size_t ND = hn::Lanes(dd); const size_t ND = hn::Lanes(dd);
VD C00 = SumOfPromotedPairs(dd, F00); VD C00 = SumOfPromotedPairs(dd, F00);