mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2024-12-27 07:29:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

Disable Tree reduction for GPU.

Browse Source 
For moderately sized inputs, running the Tree reduction quickly
fills/overflows the GPU thread stack space, leading to memory errors.
This was happening in the `cxx11_tensor_complex_gpu` test, for example.
Disabling tree reduction on GPU fixes this.
This commit is contained in:
Antonio Sanchez 2021-10-08 11:38:13 -07:00 committed by Rasmus Munk Larsen
parent 360290fc42
commit 24ebb37f38
1 changed files with 17 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

29
unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
View File

@ -168,7 +168,11 @@ struct GenericDimReducer<-1, Self, Op> {
template <typename Self, typename Op, bool Vectorizable = (Self::InputPacketAccess && Self::ReducerTraits::PacketAccess), template <typename Self, typename Op, bool Vectorizable = (Self::InputPacketAccess && Self::ReducerTraits::PacketAccess),
bool UseTreeReduction = (!Self::ReducerTraits::IsStateful && bool UseTreeReduction = (!Self::ReducerTraits::IsStateful &&
!Self::ReducerTraits::IsExactlyAssociative)> !Self::ReducerTraits::IsExactlyAssociative &&
// GPU threads can quickly run out of stack space
// for moderately sized inputs.
!Self::RunningOnGPU
)>
struct InnerMostDimReducer { struct InnerMostDimReducer {
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Self::CoeffReturnType reduce(const Self& self, typename Self::Index firstIndex, typename Self::Index numValuesToReduce, Op& reducer) { static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Self::CoeffReturnType reduce(const Self& self, typename Self::Index firstIndex, typename Self::Index numValuesToReduce, Op& reducer) {
typename Self::CoeffReturnType accum = reducer.initialize(); typename Self::CoeffReturnType accum = reducer.initialize();
@ -568,6 +572,18 @@ struct TensorReductionEvaluatorBase<const TensorReductionOp<Op, Dims, ArgType, M
// Indexed by output dimensions. // Indexed by output dimensions.
static const int NumPreservedStrides = max_n_1<NumOutputDims>::size; static const int NumPreservedStrides = max_n_1<NumOutputDims>::size;
// For full reductions
#if defined(EIGEN_USE_GPU) && (defined(EIGEN_GPUCC))
static constexpr bool RunningOnGPU = internal::is_same<Device, Eigen::GpuDevice>::value;
static constexpr bool RunningOnSycl = false;
#elif defined(EIGEN_USE_SYCL)
static const bool RunningOnSycl = internal::is_same<typename internal::remove_all<Device>::type, Eigen::SyclDevice>::value;
static const bool RunningOnGPU = false;
#else
static constexpr bool RunningOnGPU = false;
static constexpr bool RunningOnSycl = false;
#endif
enum { enum {
IsAligned = false, IsAligned = false,
PacketAccess = Self::InputPacketAccess && ReducerTraits::PacketAccess, PacketAccess = Self::InputPacketAccess && ReducerTraits::PacketAccess,
@ -989,17 +1005,6 @@ struct TensorReductionEvaluatorBase<const TensorReductionOp<Op, Dims, ArgType, M
// Operation to apply for computing the reduction. // Operation to apply for computing the reduction.
Op m_reducer; Op m_reducer;
// For full reductions
#if defined(EIGEN_USE_GPU) && (defined(EIGEN_GPUCC))
static const bool RunningOnGPU = internal::is_same<Device, Eigen::GpuDevice>::value;
static const bool RunningOnSycl = false;
#elif defined(EIGEN_USE_SYCL)
static const bool RunningOnSycl = internal::is_same<typename internal::remove_all<Device>::type, Eigen::SyclDevice>::value;
static const bool RunningOnGPU = false;
#else
static const bool RunningOnGPU = false;
static const bool RunningOnSycl = false;
#endif
EvaluatorPointerType m_result; EvaluatorPointerType m_result;
const Device EIGEN_DEVICE_REF m_device; const Device EIGEN_DEVICE_REF m_device;