GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-04-29 07:14:12 +08:00
Code Issues Packages Projects Releases Wiki Activity

Do not reduce parallelism too much in contractions with small number of threads

Browse Source
This commit is contained in:
Eugene Zhulenev 2019-02-04 12:59:33 -08:00
parent eb21bab769
commit 8491127082
1 changed files with 12 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
View File

@ -339,10 +339,19 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
// If there is enough available parallelism in sharding dimension we can
// call kernels in sync mode and use thread local memory for packed data.
const Index sharding_dim_tasks = shard_by_col ? nn : nm;
if (!parallel_pack_ && sharding_dim_tasks >= device_.numThreadsInPool()) {
parallelize_by_sharding_dim_only_ = true;
int num_worker_threads = device_.numThreadsInPool();
const int num_worker_threads = device_.numThreadsInPool();
// With small number of threads we want to make sure that we do not reduce
// parallelism too much.
const int oversharding_factor =
num_worker_threads <= 4 ? 8 :
num_worker_threads <= 8 ? 4 :
num_worker_threads <= 16 ? 2 : 1;
if (!parallel_pack_ &&
sharding_dim_tasks >= oversharding_factor * num_worker_threads) {
parallelize_by_sharding_dim_only_ = true;
if (shard_by_col) {
can_use_thread_local_packed_ = new std::atomic<bool>[nn_];