experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Fix edge case bug in handling FP16 weights in XNNPACK delegate

Browse Source 
Quasi-static tensors may become subgraph outputs after partitioning; we need to
explicitly exclude them from outputs and treat as static tensors.

PiperOrigin-RevId: 313522428
Change-Id: I621cb575b52caa59910a281078c3c505e796880f
This commit is contained in:
Marat Dukhan 2020-05-27 21:41:37 -07:00 committed by TensorFlower Gardener
parent f150cad3cb
commit dedac5053f
1 changed files with 9 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc
View File

@ -101,9 +101,15 @@ class Subgraph {
const std::unordered_set<int> inputs(
&params->input_tensors->data[0],
&params->input_tensors->data[params->input_tensors->size]);
const std::unordered_set<int> outputs(
&params->output_tensors->data[0],
&params->output_tensors->data[params->output_tensors->size]);
std::unordered_set<int> outputs;
for (int o = 0; o < params->output_tensors->size; o++) {
const int output_tensor_idx = params->output_tensors->data[o];
// Exclude quasi-static tensors which may have become subgraph outputs
// after partitioning.
if (delegate->static_unpacked_data_map_.count(output_tensor_idx) == 0) {
outputs.insert(output_tensor_idx);
}
}
std::unordered_set<int> externals(outputs);
TfLiteIntArray* execution_plan;