Skip to content

Use LoadN for short BF16 dot product inputs instead of scalar conversion #2703

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

Open
wants to merge 2 commits into
base: master
Choose a base branch
from
Open

Use LoadN for short BF16 dot product inputs instead of scalar conversion #2703

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
0478c13
Use `LoadN` for short BF16 dot product inputs instead of scalar conve…
vekterli Sep 9, 2025
df82cf0
Only use `LoadN` fallback when compiling with GCC
vekterli Sep 26, 2025
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
24 changes: 22 additions & 2 deletions hwy/contrib/dot/dot-inl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -174,9 +174,17 @@ struct Dot {
(kAssumptions & kMultipleOfVector) != 0;
constexpr bool kIsPaddedToVector = (kAssumptions & kPaddedToVector) != 0;

// Won't be able to do a full vector load without padding => scalar loop.
// Won't be able to do a full vector load without padding. Use a scalar
// loop under Clang. GCC has very suboptimal codegen for scalar BF16->float
// conversions, so use vector ops with LoadN instead.
// TODO: https://github.com/google/highway/pull/2703
if (!kIsAtLeastOneVector && !kIsMultipleOfVector && !kIsPaddedToVector &&
HWY_UNLIKELY(num_elements < NF)) {
#if HWY_COMPILER_GCC_ACTUAL
const VF a = LoadN(df, pa, num_elements);
const VF b = PromoteTo(df, LoadN(dbfh, pb, num_elements));
return ReduceSum(df, Mul(a, b));
#else
// Only 2x unroll to avoid excessive code size.
float sum0 = 0.0f;
float sum1 = 0.0f;
Expand All @@ -189,6 +197,7 @@ struct Dot {
sum1 += pa[i] * ConvertScalarTo<float>(pb[i]);
}
return sum0 + sum1;
#endif
}

// Compiler doesn't make independent sum* accumulators, so unroll manually.
Expand Down Expand Up @@ -279,9 +288,19 @@ struct Dot {
(kAssumptions & kMultipleOfVector) != 0;
constexpr bool kIsPaddedToVector = (kAssumptions & kPaddedToVector) != 0;

// Won't be able to do a full vector load without padding => scalar loop.
// Won't be able to do a full vector load without padding. Use a scalar
// loop under Clang. GCC has very suboptimal codegen for scalar BF16->float
// conversions, so use vector ops with LoadN instead.
// TODO: https://github.com/google/highway/pull/2703
if (!kIsAtLeastOneVector && !kIsMultipleOfVector && !kIsPaddedToVector &&
HWY_UNLIKELY(num_elements < N)) {
#if HWY_COMPILER_GCC_ACTUAL
const auto a = LoadN(d, pa, num_elements);
const auto b = LoadN(d, pb, num_elements);
V sum1 = Zero(df32);
V sum0 = ReorderWidenMulAccumulate(df32, a, b, Zero(df32), sum1);
return ReduceSum(df32, Add(sum0, sum1));
#else
float sum0 = 0.0f; // Only 2x unroll to avoid excessive code size for..
float sum1 = 0.0f; // this unlikely(?) case.
for (; i + 2 <= num_elements; i += 2) {
Expand All @@ -292,6 +311,7 @@ struct Dot {
sum1 += F32FromBF16(pa[i]) * F32FromBF16(pb[i]);
}
return sum0 + sum1;
#endif
}

// See comment in the other Compute() overload. Unroll 2x, but we need
Expand Down
Loading