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

Add block evaluation V2 to TensorAsyncExecutor.

Browse Source 
Add async evaluation to a number of ops.
This commit is contained in:
Rasmus Munk Larsen
2019-10-22 12:42:44 -07:00
parent 668ab3fc47
commit 97c0c5d485
9 changed files with 226 additions and 66 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
View File

@@ -689,15 +689,14 @@ struct TensorReductionEvaluatorBase<const TensorReductionOp<Op, Dims, ArgType, M
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
EIGEN_STRONG_INLINE
#if !defined(EIGEN_HIPCC)
// Marking this as EIGEN_DEVICE_FUNC for HIPCC requires also doing the same for all the functions
// being called within here, which then leads to proliferation of EIGEN_DEVICE_FUNC markings, one
// of which will eventually result in an NVCC error
EIGEN_DEVICE_FUNC
#endif
bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
m_impl.evalSubExprsIfNeeded(NULL);
#if !defined(EIGEN_HIPCC)
// Marking this as EIGEN_DEVICE_FUNC for HIPCC requires also doing the same
// for all the functions being called within here, which then leads to
// proliferation of EIGEN_DEVICE_FUNC markings, one of which will eventually
// result in an NVCC error
EIGEN_DEVICE_FUNC
#endif
bool evalSubExprsIfNeededCommon(EvaluatorPointerType data) {
// Use the FullReducer if possible.
if ((RunningFullReduction && RunningOnSycl) ||(RunningFullReduction &&
internal::FullReducer<Self, Op, Device>::HasOptimizedImplementation &&
@@ -802,6 +801,34 @@ struct TensorReductionEvaluatorBase<const TensorReductionOp<Op, Dims, ArgType, M
return true;
}
#ifdef EIGEN_USE_THREADS
template <typename EvalSubExprsCallback>
EIGEN_STRONG_INLINE
#if !defined(EIGEN_HIPCC)
EIGEN_DEVICE_FUNC
#endif
void
evalSubExprsIfNeededAsync(EvaluatorPointerType data,
EvalSubExprsCallback done) {
m_impl.evalSubExprsIfNeededAsync(NULL, [this, data, done](bool) {
done(evalSubExprsIfNeededCommon(data));
});
}
#endif
EIGEN_STRONG_INLINE
#if !defined(EIGEN_HIPCC)
// Marking this as EIGEN_DEVICE_FUNC for HIPCC requires also doing the same
// for all the functions being called within here, which then leads to
// proliferation of EIGEN_DEVICE_FUNC markings, one of which will eventually
// result in an NVCC error
EIGEN_DEVICE_FUNC
#endif
bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
m_impl.evalSubExprsIfNeeded(NULL);
return evalSubExprsIfNeededCommon(data);
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
m_impl.cleanup();
if (m_result) {