Fix issue #3264, accuracy issues on k80 GPUs. (#3293)

This commit is contained in:
Thejaswi 2018-05-04 06:44:08 +05:30 committed by Rory Mitchell
parent e1f57b4417
commit c80d51ccb3

View File

@ -37,9 +37,9 @@ __device__ GradientPairSumT ReduceFeature(const GradientPairSumT* begin,
bool thread_active = itr + threadIdx.x < end;
// Scan histogram
GradientPairSumT bin = thread_active ? *(itr + threadIdx.x) : GradientPairSumT();
local_sum += ReduceT(temp_storage->sum_reduce).Reduce(bin, cub::Sum());
local_sum += bin;
}
local_sum = ReduceT(temp_storage->sum_reduce).Reduce(local_sum, cub::Sum());
if (threadIdx.x == 0) {
shared_sum = local_sum;