saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-05-29 23:06:41 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Optimize use of Eigen::DenseBase::select() for PRelu (#15287)

Browse source 
MSVC and gcc are both not good at optimizing select(), even in trivial
usage outside of ORT.
gcc seems to do better with -ffast-math (not used by ORT) but /fp:fast
does nothing for MSVC
This PR delivers a 33% speedup on the same model (360us -> 270us on
Windows; 205 us -> 153 us on Linux; measured on different systems).

TODO: Examine and fix Elu and other similar activation functions for the
use of `Eigen::select`

  Co-authored-by: @fpribeiro

### Description
<!-- Describe your changes. -->



### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
This commit is contained in:
Dmitri Smirnov 2023-03-31 11:20:07 -07:00 committed by GitHub
parent fff75a301c
commit c06ab5e353
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 26 additions and 16 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

42
onnxruntime/core/providers/cpu/math/element_wise_ops.cc
View file

@ -370,7 +370,7 @@ REG_ELEMENTWISE_VERSIONED_TYPED_KERNEL(Mean, 8, 12, float, Mean_8);
REG_ELEMENTWISE_TYPED_KERNEL(Mean, 13, float, Mean_8);
REG_ELEMENTWISE_TYPED_KERNEL(BitShift, 11, uint8_t, BitShift);
//REG_ELEMENTWISE_TYPED_KERNEL(BitShift, 11, uint16_t, BitShift);
// REG_ELEMENTWISE_TYPED_KERNEL(BitShift, 11, uint16_t, BitShift);
REG_ELEMENTWISE_TYPED_KERNEL(BitShift, 11, uint32_t, BitShift);
REG_ELEMENTWISE_TYPED_KERNEL(BitShift, 11, uint64_t, BitShift);
@ -1192,16 +1192,16 @@ Status BitShift<T>::Compute(OpKernelContext* context) const {
template <typename T>
Status BitwiseAnd<T>::Compute(OpKernelContext* context) const {
ProcessBroadcastSpanFuncs funcs {
ProcessBroadcastSpanFuncs funcs{
[](BroadcastHelper& per_iter_bh) {
const T X = per_iter_bh.ScalarInput0<T>();
auto Y = per_iter_bh.SpanInput1<T>();
auto output = per_iter_bh.OutputSpan<T>();
std::transform(Y.begin(), Y.end(), output.begin(),
[X](T y) {
return std::bit_and<T>()(X, y);
});
[X](T y) {
return std::bit_and<T>()(X, y);
});
},
[](BroadcastHelper& per_iter_bh) {
auto X = per_iter_bh.SpanInput0<T>();
@ -1209,9 +1209,9 @@ Status BitwiseAnd<T>::Compute(OpKernelContext* context) const {
auto output = per_iter_bh.OutputSpan<T>();
std::transform(X.begin(), X.end(), output.begin(),
[Y](T x) {
return static_cast<T>(std::bit_and<T>()(x, Y));
});
[Y](T x) {
return static_cast<T>(std::bit_and<T>()(x, Y));
});
},
[](BroadcastHelper& per_iter_bh) {
auto X = per_iter_bh.SpanInput0<T>();
@ -1220,7 +1220,7 @@ Status BitwiseAnd<T>::Compute(OpKernelContext* context) const {
std::transform(X.begin(), X.end(), Y.begin(), output.begin(), std::bit_and<T>());
}};
UntypedBroadcastTwo(*context, funcs, 1.0f);
return Status::OK();
}
@ -1306,7 +1306,7 @@ Status BitwiseXor<T>::Compute(OpKernelContext* context) const {
}
template <typename T>
class Sin final : public OpKernel {
class Sin final : public OpKernel {
public:
Sin(const OpKernelInfo& info) : OpKernel(info) {
}
@ -1580,14 +1580,24 @@ Status PRelu<float>::Compute(OpKernelContext* context) const {
per_iter_bh.OutputEigen<float>() = input0 * per_iter_bh.EigenInput1<float>().array();
},
[](BroadcastHelper& per_iter_bh) {
auto input0 = per_iter_bh.EigenInput0<float>();
float input1 = per_iter_bh.ScalarInput1<float>();
per_iter_bh.OutputEigen<float>() = (input0.array() > 0).select(input0, input0 * input1);
const float* input0 = per_iter_bh.EigenInput0<float>().data();
const float input1 = per_iter_bh.ScalarInput1<float>();
float* output = per_iter_bh.OutputEigen<float>().data();
size_t size = per_iter_bh.OutputEigen<float>().size();
for (size_t i = 0; i < size; i++) {
output[i] = static_cast<float>(input0[i] > 0) * input0[i] +
(1.0f - static_cast<float>(input0[i] > 0)) * input0[i] * input1;
}
},
[](BroadcastHelper& per_iter_bh) {
auto input0 = per_iter_bh.EigenInput0<float>();
auto input1 = per_iter_bh.EigenInput1<float>();
per_iter_bh.OutputEigen<float>() = (input0.array() > 0).select(input0, input0.cwiseProduct(input1));
const float* input0 = per_iter_bh.EigenInput0<float>().data();
const float* input1 = per_iter_bh.EigenInput1<float>().data();
float* output = per_iter_bh.OutputEigen<float>().data();
size_t size = per_iter_bh.OutputEigen<float>().size();
for (size_t i = 0; i < size; i++) {
output[i] = static_cast<float>(input0[i] > 0) * input0[i] +
(1.0f - static_cast<float>(input0[i] > 0)) * input0[i] * input1[i];
}
}};
UntypedBroadcastTwo(*context, funcs, 1.0);