GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-05-03 01:04:23 +08:00
Code Issues Packages Projects Releases Wiki Activity

Reworked the dispatch of optimized cuda reduction kernels to workaround a nvcc bug that prevented the code from compiling in optimized mode in some cases

Browse Source
This commit is contained in:
Benoit Steiner 2016-01-08 16:28:53 -08:00
parent 53749ff415
commit 3358dfd5dd
1 changed files with 5 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
View File

@ -506,7 +506,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType; typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
typedef typename internal::remove_const<typename XprType::PacketReturnType>::type PacketReturnType; typedef typename internal::remove_const<typename XprType::PacketReturnType>::type PacketReturnType;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType* data) { EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType* data) {
m_impl.evalSubExprsIfNeeded(NULL); m_impl.evalSubExprsIfNeeded(NULL);
// Use the FullReducer if possible. // Use the FullReducer if possible.
@ -527,7 +527,6 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
} }
// Attempt to use an optimized reduction. // Attempt to use an optimized reduction.
#if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
else if (RunningOnGPU && data && (m_device.majorDeviceVersion() >= 3)) { else if (RunningOnGPU && data && (m_device.majorDeviceVersion() >= 3)) {
bool reducing_inner_dims = true; bool reducing_inner_dims = true;
for (int i = 0; i < NumReducedDims; ++i) { for (int i = 0; i < NumReducedDims; ++i) {
@ -537,12 +536,12 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
reducing_inner_dims &= m_reducedDims[NumInputDims - 1 - i]; reducing_inner_dims &= m_reducedDims[NumInputDims - 1 - i];
} }
} }
if (internal::InnerReducer<Self, Op, GpuDevice>::HasOptimizedImplementation && if (internal::InnerReducer<Self, Op, Device>::HasOptimizedImplementation &&
(reducing_inner_dims || ReducingInnerMostDims)) { (reducing_inner_dims || ReducingInnerMostDims)) {
const Index num_values_to_reduce = internal::array_prod(m_reducedDims); const Index num_values_to_reduce = internal::array_prod(m_reducedDims);
const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions); const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions);
Op reducer(m_reducer); Op reducer(m_reducer);
internal::InnerReducer<Self, Op, GpuDevice>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve); internal::InnerReducer<Self, Op, Device>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve);
return false; return false;
} }
@ -554,16 +553,15 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
preserving_inner_dims &= m_reducedDims[i]; preserving_inner_dims &= m_reducedDims[i];
} }
} }
if (internal::OuterReducer<Self, Op, GpuDevice>::HasOptimizedImplementation && if (internal::OuterReducer<Self, Op, Device>::HasOptimizedImplementation &&
preserving_inner_dims) { preserving_inner_dims) {
const Index num_values_to_reduce = internal::array_prod(m_reducedDims); const Index num_values_to_reduce = internal::array_prod(m_reducedDims);
const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions); const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions);
Op reducer(m_reducer); Op reducer(m_reducer);
internal::OuterReducer<Self, Op, GpuDevice>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve); internal::OuterReducer<Self, Op, Device>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve);
return false; return false;
} }
} }
#endif
return true; return true;
} }