Enable unsigned quantized infererence using XNNPACK.
PiperOrigin-RevId: 488179656
This commit is contained in:
parent
da36468409
commit
bb6fa85703
|
@ -151,6 +151,8 @@ InferenceCalculatorCpuImpl::MaybeCreateDelegate(CalculatorContext* cc) {
|
|||
auto xnnpack_opts = TfLiteXNNPackDelegateOptionsDefault();
|
||||
xnnpack_opts.num_threads =
|
||||
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),
|
||||
&TfLiteXNNPackDelegateDelete);
|
||||
}
|
||||
|
|
|
@ -114,6 +114,8 @@ InferenceCalculatorXnnpackImpl::CreateDelegate(CalculatorContext* cc) {
|
|||
auto xnnpack_opts = TfLiteXNNPackDelegateOptionsDefault();
|
||||
xnnpack_opts.num_threads =
|
||||
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),
|
||||
&TfLiteXNNPackDelegateDelete);
|
||||
}
|
||||
|
|
|
@ -941,6 +941,8 @@ absl::Status TfLiteInferenceCalculator::LoadDelegate(CalculatorContext* cc) {
|
|||
if (use_xnnpack) {
|
||||
auto xnnpack_opts = TfLiteXNNPackDelegateOptionsDefault();
|
||||
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),
|
||||
&TfLiteXNNPackDelegateDelete);
|
||||
RET_CHECK_EQ(interpreter_->ModifyGraphWithDelegate(delegate_.get()),
|
||||
|
|
Loading…
Reference in New Issue
Block a user