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

TFLM: Save arena by reusing flatbuffer tensor shape.

Browse Source 
This will save up to several hundreds of bytes (sizeof(int) * #dims * #tensors). Not significant, but still a nice incremental improvement.

PiperOrigin-RevId: 286507276
Change-Id: I1c25f88d24fdf264b6676275da1d9fe8a0d45732
This commit is contained in:
Tiezhen WANG 2019-12-19 19:27:39 -08:00 committed by TensorFlower Gardener
parent 77fad4a3ff
commit 67d1f027b1
1 changed files with 8 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
tensorflow/lite/micro/micro_allocator.cc
View File

@ -418,16 +418,14 @@ TfLiteStatus MicroAllocator::InitializeRuntimeTensor(
size_t type_size;
TF_LITE_ENSURE_STATUS(BytesRequiredForTensor(
flatbuffer_tensor, &result->bytes, &type_size, error_reporter));
// Copy the shape of the tensor from the serialized data into the runtime
// form. We have to allocate memory for this.
result->dims =
reinterpret_cast<TfLiteIntArray*>(memory_allocator_.AllocateFromTail(
TfLiteIntArrayGetSizeInBytes(flatbuffer_tensor.shape()->Length()),
alignof(TfLiteIntArray)));
result->dims->size = flatbuffer_tensor.shape()->Length();
for (size_t n = 0; n < flatbuffer_tensor.shape()->Length(); ++n) {
result->dims->data[n] = flatbuffer_tensor.shape()->Get(n);
}
// TFLM doesn't allow reshaping the tensor which requires dynamic memory
// allocation so it is safe to drop the const qualifier. In the future, if we
// really want to update the tensor shape, we can always pass in a new
// TfLiteIntArray - especially we have to do so if the dimension is changed.
result->dims = const_cast<TfLiteIntArray*>(
reinterpret_cast<const TfLiteIntArray*>(flatbuffer_tensor.shape()));
// Copy the quantization information from the serialized data.
const auto* src_quantization = flatbuffer_tensor.quantization();
if (src_quantization && src_quantization->scale() &&