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 batch norm vectorize path accuracy issue by enforcing shape alignment #1238

Merged
merged 4 commits into from
Jan 3, 2025
Merged

Fix batch norm vectorize path accuracy issue by enforcing shape alignment #1238

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
4d3d601
Update BatchNormKernels.cpp
xytintel Dec 31, 2024
1dec145
Merge branch 'main' into xyt/batch_norm_vec_fix
mengfei25 Dec 31, 2024
8667ccc
Merge branch 'main' into xyt/batch_norm_vec_fix
xytintel Jan 2, 2025
6e82696
Update BatchNormKernels.cpp
xytintel Jan 2, 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
29 changes: 10 additions & 19 deletions src/ATen/native/xpu/sycl/BatchNormKernels.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1287,7 +1287,7 @@ struct BatchNormTransformInputVectorizedKernelFunctor {
} else {
invstd =
static_cast<stat_accscalar_t>(1) /
device_sqrt(
std::sqrt(
static_cast<stat_accscalar_t>(var_or_invstd_[plane]) + epsilon_);
}

Expand All @@ -1302,25 +1302,16 @@ struct BatchNormTransformInputVectorizedKernelFunctor {
for (index_t feature_vec_begin = item.get_local_id(1) * VEC_SIZE;
feature_vec_begin < fs;
feature_vec_begin += VEC_SIZE * item.get_local_range(1)) {
auto remaining = fs - feature_vec_begin;
if (remaining < VEC_SIZE) {
for (index_t idx = 0; idx < remaining; ++idx) {
index_t feature = feature_vec_begin + idx;
o[feature] = static_cast<input_scalar_t>(
gamma * (i[feature] - mean) * invstd + beta);
}
} else {
using vec_t = memory::aligned_vector<input_scalar_t, VEC_SIZE>;
vec_t vec;
using vec_t = memory::aligned_vector<input_scalar_t, VEC_SIZE>;
vec_t vec;
#pragma unroll
for (int vt = 0; vt < VEC_SIZE; ++vt) {
index_t feature = feature_vec_begin + vt;
vec[vt] = static_cast<input_scalar_t>(
gamma * (i[feature] - mean) * invstd + beta);
}
input_scalar_t* write_ptr = &o[feature_vec_begin];
*(reinterpret_cast<vec_t*>(write_ptr)) = vec;
for (int vt = 0; vt < VEC_SIZE; ++vt) {
index_t feature = feature_vec_begin + vt;
vec[vt] = static_cast<input_scalar_t>(
gamma * (i[feature] - mean) * invstd + beta);
}
input_scalar_t* write_ptr = &o[feature_vec_begin];
*(reinterpret_cast<vec_t*>(write_ptr)) = vec;
}
}
}
Expand Down Expand Up @@ -1459,7 +1450,7 @@ void batch_norm_elemt_template(
auto output_ptr = (char*)output_reshaped.data_ptr();
if (output_reshaped.is_contiguous() &&
memory::can_vectorize_up_to<input_scalar_t>(output_ptr) >= 4 &&
sizeof(input_scalar_t) < sizeof(float)) {
sizeof(input_scalar_t) < sizeof(float) && input.size(2) % 4 == 0) {
auto kfn = BatchNormTransformInputVectorizedKernelFunctor<
4,
input_scalar_t,
Expand Down
Loading