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

Rewrite a useless std::memory_order_release as std::memory_order_relaxed.

Browse Source 
It's OK because these stores happen before calling Execute, and these values
will only be consumed in other threads after Execute has communicated with
these threads... which will involve a release-store anyway.

This might matter because this suggests that at least on ARM32, each release-store may involve a __sync_synchronize and the compiler may not know to keep only the last memory barrier in a loop that only performs stores:
https://godbolt.org/z/HYijeI

PiperOrigin-RevId: 259982740
This commit is contained in:
Benoit Jacob 2019-07-25 11:09:09 -07:00 committed by TensorFlower Gardener
parent 64105e1c41
commit 625f9aad07
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
tensorflow/lite/experimental/ruy/trmul.cc
View File

@ -237,7 +237,7 @@ void TrMul(TrMulParams* params, Context* context) {
const int size = NumBlocksPerSide(side, block_map); const int size = NumBlocksPerSide(side, block_map);
allocator->Allocate(size, &packed[side]); allocator->Allocate(size, &packed[side]);
for (int i = 0; i < size; i++) { for (int i = 0; i < size; i++) {
packed[side][i].store(false, std::memory_order_release); packed[side][i].store(false, std::memory_order_relaxed);
} }
} }
} }