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

Improved the performance of reductions on CUDA devices

Browse Source
This commit is contained in:
Benoit Steiner
2016-01-04 17:25:00 -08:00
parent 515dee0baf
commit cfff40b1d4
2 changed files with 128 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
View File

@@ -337,6 +337,16 @@ struct FullReducer<Self, Op, ThreadPoolDevice, true> {
#endif
// Default inner reducer
template <typename Self, typename Op, typename Device>
struct InnerReducer {
static const bool HasOptimizedImplementation = false;
static EIGEN_DEVICE_FUNC void run(const Self&, Op&, const Device&, typename Self::CoeffReturnType*, typename Self::Index, typename Self::Index) {
assert(false && "Not implemented");
}
};
// Default outer reducer
template <typename Self, typename Op, typename Device>
struct OuterReducer {
@@ -352,6 +362,9 @@ struct OuterReducer {
template <int B, int N, typename S, typename R, typename I>
__global__ void FullReductionKernel(R, const S, I, typename S::CoeffReturnType*);
template <int NPT, typename S, typename R, typename I>
__global__ void InnerReductionKernel(R, const S, I, I, typename S::CoeffReturnType*);
template <int NPT, typename S, typename R, typename I>
__global__ void OuterReductionKernel(R, const S, I, I, typename S::CoeffReturnType*);
#endif
@@ -516,6 +529,23 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
// Attempt to use an optimized reduction.
#if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
else if (RunningOnGPU && data && (m_device.majorDeviceVersion() >= 3)) {
bool reducing_inner_dims = true;
for (int i = 0; i < NumReducedDims; ++i) {
if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
reducing_inner_dims &= m_reducedDims[i];
} else {
reducing_inner_dims &= m_reducedDims[NumInputDims - 1 - i];
}
}
if (internal::InnerReducer<Self, Op, GpuDevice>::HasOptimizedImplementation &&
(reducing_inner_dims || ReducingInnerMostDims)) {
const Index num_values_to_reduce = internal::array_prod(m_reducedDims);
const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions);
Op reducer(m_reducer);
internal::InnerReducer<Self, Op, GpuDevice>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve);
return false;
}
bool preserving_inner_dims = true;
for (int i = 0; i < NumReducedDims; ++i) {
if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
@@ -615,6 +645,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
#endif
#if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
template <int B, int N, typename S, typename R, typename I> friend void internal::FullReductionKernel(R, const S, I, typename S::CoeffReturnType*);
template <int NPT, typename S, typename R, typename I> friend void internal::InnerReductionKernel(R, const S, I, I, typename S::CoeffReturnType*);
template <int NPT, typename S, typename R, typename I> friend void internal::OuterReductionKernel(R, const S, I, I, typename S::CoeffReturnType*);
#endif