frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-20 21:14:14 +08:00
Code Issues Packages Projects Releases Wiki Activity

Revert "Fix boxcox to return same result for same input in one batch (#162772)"

Browse Source 
This reverts commit 49d30f9a234f0816a1ece278c8450d119e417714.

Reverted https://github.com/pytorch/pytorch/pull/162772 on behalf of https://github.com/facebook-github-bot due to Diff reverted internally ([comment](https://github.com/pytorch/pytorch/pull/162772#issuecomment-3313213011))
This commit is contained in:
PyTorch MergeBot
2025-09-19 17:58:29 +00:00
parent 2984bfe3da
commit a3b68c7c57
1 changed files with 12 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
caffe2/perfkernels/batch_box_cox_vec.h
View File

@ -73,19 +73,6 @@ void box_cox_zero_lambda(
}
}
template <typename T>
at::vec::Vectorized<T> box_cox_nonzero_lambda_impl(
at::vec::Vectorized<T> data,
at::vec::Vectorized<T> lambda1,
at::vec::Vectorized<T> lambda2,
at::vec::Vectorized<T> k_eps) {
auto sum = data + lambda2;
auto max = at::vec::max(sum, k_eps);
auto lambda_over_1 = at::vec::fast_recieprocal(lambda1);
auto pow = max.pow(lambda1);
return at::vec::fmsub(pow, lambda_over_1, lambda_over_1);
}
template <typename T>
void box_cox_nonzero_lambda(
int64_t D,
@ -101,18 +88,21 @@ void box_cox_nonzero_lambda(
auto k_eps_vec = Vec(k_eps);
for(; j + VLEN < D; j += VLEN) {
auto data = Vec::loadu(data_ptr + j);
auto lambda1 = Vec::loadu(lambda1_ptr + j);
auto lambda2 = Vec::loadu(lambda2_ptr + j);
auto res = box_cox_nonzero_lambda_impl(data, lambda1, lambda2, k_eps_vec);
auto sum = data + lambda2;
auto max = at::vec::max(sum, k_eps_vec);
auto lambda1 = Vec::loadu(lambda1_ptr + j);
auto lambda_over_1 = at::vec::fast_recieprocal(lambda1);
auto pow = max.pow(lambda1);
auto res = at::vec::fmsub(pow, lambda_over_1, lambda_over_1);
res.store(out + j);
}
if (j < D) {
auto remaining = D - j;
auto data = Vec::loadu(data_ptr + j, remaining);
auto lambda1 = Vec::loadu(lambda1_ptr + j, remaining);
auto lambda2 = Vec::loadu(lambda2_ptr + j, remaining);
auto res = box_cox_nonzero_lambda_impl(data, lambda1, lambda2, k_eps_vec);
res.store(out + j, remaining);
for ( ;j < D; ++j) {
auto sum = data_ptr[j] + lambda2_ptr[j];
auto max = std::max(sum, k_eps);
auto lambda_over_1 = at::vec::fast_recieprocal(lambda1_ptr[j]);
auto pow = std::pow(max, lambda1_ptr[j]);
out[j] = pow * lambda_over_1 - lambda_over_1;
}
}
#else