talksik/mediapipe
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Enable unsigned quantized infererence using XNNPACK.

Browse Source 
PiperOrigin-RevId: 488179656
This commit is contained in:
Alan Kelly
2022-11-13 07:10:04 -08:00
committed by Copybara-Service
parent da36468409
commit bb6fa85703
3 changed files with 6 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
mediapipe/calculators/tensor/inference_calculator_cpu.cc
+2
View File
@@ -151,6 +151,8 @@ InferenceCalculatorCpuImpl::MaybeCreateDelegate(CalculatorContext* cc) {
auto xnnpack_opts = TfLiteXNNPackDelegateOptionsDefault(); auto xnnpack_opts = TfLiteXNNPackDelegateOptionsDefault();
xnnpack_opts.num_threads = xnnpack_opts.num_threads =
GetXnnpackNumThreads(opts_has_delegate, opts_delegate); GetXnnpackNumThreads(opts_has_delegate, opts_delegate);
// TODO Remove once XNNPACK is enabled by default.
xnnpack_opts.flags |= TFLITE_XNNPACK_DELEGATE_FLAG_QU8;
return TfLiteDelegatePtr(TfLiteXNNPackDelegateCreate(&xnnpack_opts), return TfLiteDelegatePtr(TfLiteXNNPackDelegateCreate(&xnnpack_opts),
&TfLiteXNNPackDelegateDelete); &TfLiteXNNPackDelegateDelete);
} }
mediapipe/calculators/tensor/inference_calculator_xnnpack.cc
+2
View File
@@ -114,6 +114,8 @@ InferenceCalculatorXnnpackImpl::CreateDelegate(CalculatorContext* cc) {
auto xnnpack_opts = TfLiteXNNPackDelegateOptionsDefault(); auto xnnpack_opts = TfLiteXNNPackDelegateOptionsDefault();
xnnpack_opts.num_threads = xnnpack_opts.num_threads =
GetXnnpackNumThreads(opts_has_delegate, opts_delegate); GetXnnpackNumThreads(opts_has_delegate, opts_delegate);
// TODO Remove once XNNPACK is enabled by default.
xnnpack_opts.flags |= TFLITE_XNNPACK_DELEGATE_FLAG_QU8;
return TfLiteDelegatePtr(TfLiteXNNPackDelegateCreate(&xnnpack_opts), return TfLiteDelegatePtr(TfLiteXNNPackDelegateCreate(&xnnpack_opts),
&TfLiteXNNPackDelegateDelete); &TfLiteXNNPackDelegateDelete);
} }
mediapipe/calculators/tflite/tflite_inference_calculator.cc
+2
View File
@@ -941,6 +941,8 @@ absl::Status TfLiteInferenceCalculator::LoadDelegate(CalculatorContext* cc) {
if (use_xnnpack) { if (use_xnnpack) {
auto xnnpack_opts = TfLiteXNNPackDelegateOptionsDefault(); auto xnnpack_opts = TfLiteXNNPackDelegateOptionsDefault();
xnnpack_opts.num_threads = GetXnnpackNumThreads(calculator_opts); xnnpack_opts.num_threads = GetXnnpackNumThreads(calculator_opts);
// TODO Remove once XNNPACK is enabled by default.
xnnpack_opts.flags |= TFLITE_XNNPACK_DELEGATE_FLAG_QU8;
delegate_ = TfLiteDelegatePtr(TfLiteXNNPackDelegateCreate(&xnnpack_opts), delegate_ = TfLiteDelegatePtr(TfLiteXNNPackDelegateCreate(&xnnpack_opts),
&TfLiteXNNPackDelegateDelete); &TfLiteXNNPackDelegateDelete);
RET_CHECK_EQ(interpreter_->ModifyGraphWithDelegate(delegate_.get()), RET_CHECK_EQ(interpreter_->ModifyGraphWithDelegate(delegate_.get()),