Skip to content

Optimization in LSTM for batch > 1 cases on HiFi. #3244

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

Open
wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Optimization in LSTM for batch > 1 cases on HiFi. #3244

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c122105
Optimization in LSTM for batch > 1 cases on HiFi.
pramods-cad Oct 27, 2025
c98ec28
Addressed review comments.
pramods-cad Oct 31, 2025
f5d86dd
Fixed code style errors.
pramods-cad Nov 9, 2025
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
4 changes: 2 additions & 2 deletions tensorflow/lite/micro/kernels/xtensa/lstm_eval.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -473,7 +473,7 @@ void LstmStepManager::UpdateBatch() {
// Multi-batch for time_major input
RuntimeShape LstmStepManager::InputShape() const {
int batch_size = 1;
if (size_info_.time_major) {
if (size_info_.time_major || ((size_info_.batch_size > 1 && size_info_.time_steps == 1))) {
batch_size = size_info_.batch_size;
}
const int dims[2] = {batch_size, size_info_.input_dimension};
Expand All @@ -485,7 +485,7 @@ RuntimeShape LstmStepManager::InputShape() const {
// Multi-batch for time_major input
RuntimeShape LstmStepManager::StateShape() const {
int batch_size = 1;
if (size_info_.time_major) {
if (size_info_.time_major || (size_info_.batch_size > 1 && size_info_.time_steps == 1)) {
batch_size = size_info_.batch_size;
}
const int dims[2] = {batch_size, size_info_.state_dimension};
Expand Down
29 changes: 20 additions & 9 deletions tensorflow/lite/micro/kernels/xtensa/lstm_eval.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -666,6 +666,11 @@ void LstmStep(const LstmStepManager& step_info, const OpDataLSTM& op_data,
int input_dimension = step_info.input_dimension();
int state_dimension = step_info.state_dimension();

const auto& size_info = op_data.size_info;
if(size_info.batch_size > 1 && size_info.time_steps == 1) {
num_batches = size_info.batch_size;
}

// Check offset validity to avoid memory overflow
TFLITE_DCHECK_LE(step_info.InputOffset() + num_batches * input_dimension,
tflite::micro::GetTensorShape(input).FlatSize());
Expand Down Expand Up @@ -805,16 +810,22 @@ TfLiteStatus EvalLstm(const OpDataLSTM& op_data,
}
} else {
// batch first, unable to size the input data. single batch inference
for (int b = 0; b < size_info.batch_size; b++) {
for (int t = 0; t < size_info.time_steps; t++) {
lstm_internal::LstmStep<ActivationType, WeightType, CellType, BiasType>(
step_info, op_data, kernel_content, buffers);
// prepare for the next time step
step_info.UpdateTime();
if(size_info.batch_size > 1 && size_info.time_steps == 1) {
// Ramesh
lstm_internal::LstmStep<ActivationType, WeightType, CellType, BiasType>(
step_info, op_data, kernel_content, buffers);
} else {
for (int b = 0; b < size_info.batch_size; b++) {
for (int t = 0; t < size_info.time_steps; t++) {
lstm_internal::LstmStep<ActivationType, WeightType, CellType, BiasType>(
step_info, op_data, kernel_content, buffers);
// prepare for the next time step
step_info.UpdateTime();
}
// prepare for the next batch
step_info.UpdateBatch();
step_info.ResetTime();
}
// prepare for the next batch
step_info.UpdateBatch();
step_info.ResetTime();
}
}
return kTfLiteOk;
Expand Down