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

Optimize the computation of log_softmax #40612

Merged
merged 2 commits into from
Mar 16, 2022
Merged

Optimize the computation of log_softmax #40612

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
92e199e
Optimize the computation of log_softmax
ZzSean Mar 16, 2022
bde814b
modify the var name
ZzSean Mar 16, 2022
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
43 changes: 20 additions & 23 deletions paddle/phi/kernels/gpudnn/softmax_gpudnn.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -121,17 +121,10 @@ struct ReduceMaxFunctor {
};

template <typename Tx, typename Ty = Tx>
struct ExpSubFunctor {
HOSTDEVICE inline ExpSubFunctor() { y = static_cast<Tx>(0.0f); }

HOSTDEVICE explicit inline ExpSubFunctor(Tx y) : y((Tx)(y)) {}

struct ExpFunctor {
HOSTDEVICE inline Ty operator()(const Tx& x) const {
return static_cast<Ty>(std::exp(x - y));
return static_cast<Ty>(std::exp(x));
}

private:
Tx y;
};

template <typename Tx, typename Ty = Tx>
Expand Down Expand Up @@ -293,10 +286,14 @@ __global__ void WarpSoftmaxForward(T* softmax,
}

// data src
AccT srcdata[kBatchSize][kLoopsV][kVSize];
T src_tmp[kBatchSize][kLoopsV][kVSize];
kps::Init<AccT, kStep>(&srcdata[0][0][0], kLowInf);
kps::Init<T, kStep>(&src_tmp[0][0][0], -std::numeric_limits<T>::infinity());
// src_data: the raw data form global memory
// sub_data: store the data obtained by (src_data - max), used by log_softmax
// exp_data: store the data obtained by (exp(sub_data)), used by softmax
T src_data[kBatchSize][kLoopsV][kVSize];
AccT sub_data[kBatchSize][kLoopsV][kVSize];
AccT exp_data[kBatchSize][kLoopsV][kVSize];
kps::Init<AccT, kStep>(&sub_data[0][0][0], kLowInf);
kps::Init<T, kStep>(&src_data[0][0][0], -std::numeric_limits<T>::infinity());

// data dst
T out_tmp[kBatchSize][kLoopsV][kVSize];
Expand All @@ -313,11 +310,11 @@ __global__ void WarpSoftmaxForward(T* softmax,
for (int i = 0; i < kBatchSize; ++i) {
const VecT* src_v =
reinterpret_cast<const VecT*>(&src[(first_batch + i) * stride]);
VecT* reg_v = reinterpret_cast<VecT*>(&src_tmp[i][0][0]);
VecT* reg_v = reinterpret_cast<VecT*>(&src_data[i][0][0]);
kps::ReadData<VecT, VecT, kLoopsV, 1, 1, true>(
&reg_v[0], &src_v[0], idx_max_v[i], 0, kWarpSize, 1);
kps::ElementwiseUnary<T, AccT, kVItem, 1, 1, DataTransFunctor<T, AccT>>(
&srcdata[i][0][0], &src_tmp[i][0][0], DataTransFunctor<T, AccT>());
&sub_data[i][0][0], &src_data[i][0][0], DataTransFunctor<T, AccT>());
}

// compute max
Expand All @@ -327,22 +324,24 @@ __global__ void WarpSoftmaxForward(T* softmax,
1,
ReduceMaxFunctor<AccT>,
kMode::kLocalMode>(
&max[0], &srcdata[0][0][0], ReduceMaxFunctor<AccT>(), true);
&max[0], &sub_data[0][0][0], ReduceMaxFunctor<AccT>(), true);
WarpReduceMax<AccT, kBatchSize, kWarpSize>(max);

// compute sum
#pragma unroll
for (int i = 0; i < kBatchSize; ++i) {
kps::ElementwiseUnary<AccT, AccT, kVItem, 1, 1, ExpSubFunctor<AccT>>(
&srcdata[i][0][0], &srcdata[i][0][0], ExpSubFunctor<AccT>(max[i]));
kps::ElementwiseUnary<AccT, AccT, kVItem, 1, 1, UnarySubFunctor<AccT>>(
&sub_data[i][0][0], &sub_data[i][0][0], UnarySubFunctor<AccT>(max[i]));
kps::ElementwiseUnary<AccT, AccT, kVItem, 1, 1, ExpFunctor<AccT>>(
&exp_data[i][0][0], &sub_data[i][0][0], ExpFunctor<AccT>());
}
kps::Reduce<AccT,
kVItem,
kBatchSize,
1,
kps::AddFunctor<AccT>,
kMode::kLocalMode>(
&sum[0], &srcdata[0][0][0], kps::AddFunctor<AccT>(), true);
&sum[0], &exp_data[0][0][0], kps::AddFunctor<AccT>(), true);
WarpReduceSum<AccT, kBatchSize, kWarpSize>(sum);

// write data to global memory
Expand All @@ -352,15 +351,13 @@ __global__ void WarpSoftmaxForward(T* softmax,
reinterpret_cast<VecT*>(&softmax[(first_batch + i) * stride]);
VecT* reg_v = reinterpret_cast<VecT*>(&out_tmp[i][0][0]);
if (LogMode) {
kps::ElementwiseUnary<AccT, AccT, kVItem, 1, 1, UnaryLogFunctor<AccT>>(
&srcdata[i][0][0], &srcdata[i][0][0], UnaryLogFunctor<AccT>());
kps::ElementwiseUnary<AccT, T, kVItem, 1, 1, UnarySubFunctor<AccT>>(
&out_tmp[i][0][0],
&srcdata[i][0][0],
&sub_data[i][0][0],
UnarySubFunctor<AccT>(std::log(sum[i])));
} else {
kps::ElementwiseUnary<AccT, T, kVItem, 1, 1, UnaryDivFunctor<AccT>>(
&out_tmp[i][0][0], &srcdata[i][0][0], UnaryDivFunctor<AccT>(sum[i]));
&out_tmp[i][0][0], &exp_data[i][0][0], UnaryDivFunctor<AccT>(sum[i]));
}
kps::WriteData<VecT, VecT, kLoopsV, 1, 1, true>(
&softmax_v[0], &reg_v[0], idx_max_v[i], 0, kWarpSize, 1);
Expand Down