Skip to content

[XPU] reuse logits and softmax to avoid redundant memory allocation #68906

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
merged 1 commit into from
Oct 25, 2024
Merged

[XPU] reuse logits and softmax to avoid redundant memory allocation #68906

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
29 changes: 4 additions & 25 deletions paddle/fluid/operators/collective/c_softmax_with_cross_entropy_op_xpu.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -152,18 +152,13 @@ struct CSoftmaxWithCrossEntropyProcessGroupFunctor<phi::XPUContext, T> {
dev_ctx.template Alloc(loss, logits->dtype());

const auto& logits_dims = logits->dims();

const int axis = logits_dims.size() - 1;
const int64_t N = phi::funcs::SizeToAxis(axis, logits_dims);
const int64_t D = phi::funcs::SizeFromAxis(axis, logits_dims);

phi::DenseTensor logits_2d, softmax_2d;
framework::TensorCopy(
*logits, ctx.GetPlace(), ctx.device_context(), &logits_2d);
framework::TensorCopy(
*softmax, ctx.GetPlace(), ctx.device_context(), &softmax_2d);
logits_2d.Resize({N, D});
softmax_2d.Resize({N, D});
logits_2d.ShareDataWith(*logits).Resize({N, D});
softmax_2d.ShareDataWith(*softmax).Resize({N, D});
Copy link
Contributor

@lj970926 lj970926 Oct 24, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

下面CSoftmaxWithCrossEntropyFunctor（line 389~line 395）那里是不是也可以一起改下？

Copy link
Contributor

@HarperCy HarperCy Oct 24, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

sharedatawith我记得paddle同学说过，是访问同一块地址，会有潜在的风险。之前paddle cast，在数据类型相同时，就是用了sharedatawith，模型训练有时就会有奇怪的现象

Copy link
Contributor Author

@runzhech runzhech Oct 25, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

这个算子比较特殊，sharedatawith的是两个临时tensor。其中logits_2d不会被修改，softmax_2d在修改后原本就会被写回原地址，逻辑上跟之前一样，所以不会有问题。


int ret = -1;
// step 1, obtain logit_max
Expand Down Expand Up @@ -324,12 +319,6 @@ struct CSoftmaxWithCrossEntropyProcessGroupFunctor<phi::XPUContext, T> {
// 将label和ignore_index相同的那些loss，置为0
FixLossAccordingToIgnoreIndex<T>(
ctx, labels, &predicted_logits, loss, N, ignore_index);

phi::memory_utils::Copy(ctx.GetPlace(),
softmax->data(),
ctx.GetPlace(),
softmax_2d.data(),
N * D * sizeof(T));
}
};

Expand Down Expand Up @@ -398,12 +387,8 @@ struct CSoftmaxWithCrossEntropyFunctor<phi::XPUContext, T> {
const int64_t D = phi::funcs::SizeFromAxis(axis, logits_dims);

phi::DenseTensor logits_2d, softmax_2d;
framework::TensorCopy(
*logits, ctx.GetPlace(), ctx.device_context(), &logits_2d);
framework::TensorCopy(
*softmax, ctx.GetPlace(), ctx.device_context(), &softmax_2d);
logits_2d.Resize({N, D});
softmax_2d.Resize({N, D});
logits_2d.ShareDataWith(*logits).Resize({N, D});
softmax_2d.ShareDataWith(*softmax).Resize({N, D});

int ret = -1;
// step 1, obtain logit_max
Expand Down Expand Up @@ -593,12 +578,6 @@ struct CSoftmaxWithCrossEntropyFunctor<phi::XPUContext, T> {
// 将label和ignore_index相同的那些loss，置为0
FixLossAccordingToIgnoreIndex<T>(
ctx, labels, &predicted_logits, loss, N, ignore_index);

phi::memory_utils::Copy(ctx.GetPlace(),
softmax->data(),
ctx.GetPlace(),
softmax_2d.data(),
N * D * sizeof(T));
}
};

Expand Down