saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-05-31 23:27:43 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Move some vectors out of loops to minimize memory allocations/reallocations (#3062)

Browse source 
* Move some vectors out of loops to minimize memory allocations/reallocations.
Remove some unused/unnecessary vectors.
This commit is contained in:
Scott McKay 2020-02-22 06:35:13 +10:00 committed by GitHub
parent 7ffb36be44
commit 21cc2d88b4
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
2 changed files with 22 additions and 13 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

5
onnxruntime/core/providers/cpu/ml/linearregressor.cc
View file

@ -35,9 +35,12 @@ Status LinearRegressor<float>::Compute(OpKernelContext* ctx) const {
int64_t yindex = 0;
bool useIntercepts = intercepts_.size() == static_cast<size_t>(targets_);
std::vector<float> scores;
scores.reserve(targets_);
for (int64_t i = 0; i < N; i++) //for each point
{
std::vector<float> scores;
scores.clear();
int64_t current_weight_0 = i * stride;
for (int j = 0; j < targets_; j++) //for each target
{

30
onnxruntime/core/providers/cpu/ml/svmclassifier.cc
View file

@ -124,14 +124,22 @@ Status SVMClassifier<T>::Compute(OpKernelContext* ctx) const {
const T* x_data = X->template Data<T>();
int64_t zindex = 0;
std::vector<float> scores;
std::vector<float> kernels;
std::vector<int64_t> votes;
std::vector<float> probsp2;
const int64_t class_count_squared = class_count_ * class_count_;
probsp2.reserve(class_count_squared);
scores.reserve(class_count_squared);
for (int64_t n = 0; n < N; n++) //for each example
{
scores.clear();
kernels.clear();
int64_t current_weight_0 = n * stride;
int64_t maxclass = -1;
std::vector<float> decisions;
std::vector<float> scores;
std::vector<float> kernels;
std::vector<int64_t> votes;
if (vector_count_ == 0 && mode_ == SVM_TYPE::SVM_LINEAR) {
for (int64_t j = 0; j < class_count_; j++) { //for each class
@ -150,7 +158,8 @@ Status SVMClassifier<T>::Compute(OpKernelContext* ctx) const {
feature_count_, get_kernel_type());
kernels.push_back(val);
}
votes.resize(class_count_, 0);
votes.assign(class_count_, 0);
for (int64_t i = 0; i < class_count_; i++) { // for each class
for (int64_t j = i + 1; j < class_count_; j++) { // for each class
double sum = 0;
@ -182,9 +191,7 @@ Status SVMClassifier<T>::Compute(OpKernelContext* ctx) const {
if (proba_.size() > 0 && mode_ == SVM_TYPE::SVM_SVC) {
//compute probabilities from the scores
int64_t num = class_count_ * class_count_;
std::vector<float> probsp2(num, 0.f);
std::vector<float> estimates(class_count_, 0.f);
probsp2.assign(class_count_squared, 0.f);
int64_t index = 0;
for (int64_t i = 0; i < class_count_; ++i) {
int64_t p1 = i * class_count_ + i + 1;
@ -199,10 +206,9 @@ Status SVMClassifier<T>::Compute(OpKernelContext* ctx) const {
p2 += class_count_;
}
}
multiclass_probability(class_count_, probsp2, estimates);
// copy probabilities back into scores
scores.resize(estimates.size());
std::copy(estimates.begin(), estimates.end(), scores.begin());
scores.assign(class_count_, 0.f);
multiclass_probability(class_count_, probsp2, scores);
}
float max_weight = 0;