GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-05-22 12:37:35 +08:00
Code Issues Packages Projects Releases Wiki Activity

Doubled the speed of full reductions on GPUs.

Browse Source
This commit is contained in:
Benoit Steiner 2015-12-18 14:07:31 -08:00
parent 3abd8470ca
commit 75a7fa1919
1 changed files with 5 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
View File

@ -87,15 +87,15 @@ __global__ void FullReductionKernel(Reducer reducer, const Self input, Index num
}
typename Self::CoeffReturnType accum = reducer.initialize();
for (Index i = 0; i < NumPerThread; ++i) {
const Index index = first_index + i * BlockSize;
if (index >= num_coeffs) {
break;
}
Index max_iter = numext::mini<Index>(num_coeffs - first_index, NumPerThread*BlockSize);
for (Index i = 0; i < max_iter; i+=BlockSize) {
const Index index = first_index + i;
eigen_assert(index < num_coeffs);
typename Self::CoeffReturnType val = input.m_impl.coeff(index);
reducer.reduce(val, &accum);
}
#pragma unroll
for (int offset = warpSize/2; offset > 0; offset /= 2) {
reducer.reduce(__shfl_down(accum, offset), &accum);
}