devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Speedup trivial tensor broadcasting on GPU by enforcing unaligned loads. See PR 437.

Browse Source
This commit is contained in:
Gael Guennebaud
2018-07-31 10:10:14 +02:00
parent 723856dec1
commit 679eece876
1 changed files with 11 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
View File

@@ -284,7 +284,13 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
if (isCopy) {
#ifdef EIGEN_GPU_COMPILE_PHASE
// See PR 437: on NVIDIA P100 and K20m we observed a x3-4 speed up by enforcing
// unaligned loads here. The reason is unclear though.
return m_impl.template packet<Unaligned>(index);
#else
return m_impl.template packet<LoadMode>(index);
#endif
} else if (oneByN && !nByOne) {
return packetNByOne<LoadMode>(index);
} else if (!oneByN && nByOne) {
@@ -296,7 +302,12 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
}
} else {
if (isCopy) {
#ifdef EIGEN_GPU_COMPILE_PHASE
// See above.
return m_impl.template packet<Unaligned>(index);
#else
return m_impl.template packet<LoadMode>(index);
#endif
} else if (oneByN && !nByOne) {
return packetOneByN<LoadMode>(index);
} else if (!oneByN && nByOne) {