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

[XLA:CPU] Don't create more shards than the max allowed parallelism.

Browse Source 
The max allowed parallelism might e.g. be 1, in which case creating shards is
unnecessary and unhelpful.

PiperOrigin-RevId: 248941468
This commit is contained in:
Justin Lebar 2019-05-19 08:30:28 -07:00 committed by TensorFlower Gardener
parent 64aee049de
commit 810b454169
1 changed files with 3 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
tensorflow/compiler/xla/service/cpu/parallel_task_assignment.cc
View File

@ -74,8 +74,9 @@ class DefaultCostModel : public ParallelCostModel {
// Limit max parallelism for I/O bound instructions by assuming a
// sub-linear scaling function (fit based on empirical benchmark results).
// TODO(b/29630486) Develop system bandwidth model.
max_parallelism =
std::ceil(std::sqrt(tensorflow::port::NumSchedulableCPUs()));
max_parallelism = std::min<int64>(
max_parallelism_,
std::ceil(std::sqrt(tensorflow::port::NumSchedulableCPUs())));
// Use shape size instruction cost and L2 cache size min per-thread cost.
instruction_cost = shape_size_(instruction->shape());
min_cost_per_thread = 256LL << 10; // 256KB L2 Cache size.