src/reduce.cu
Line 44 to : shuff = __shfl_down_sync(0xFFFFFFFF, x, offset);
include/contrib/moderngpu/include/device/intrinsics.cuh
Line 115 to : var = __shfl_up_sync(0xFFFFFFFF, var, delta, width);
Line 125 to : p.x = __shfl_up_sync(0xFFFFFFFF, p.x, delta, width);
Line 126 to : p.y = __shfl_up_sync(0xFFFFFFFF, p.y, delta, width);
Line 143 to : "shfl.up.sync.b32 r0|p, %1, %2, %3, %4;"
Line 158 to : "shfl.up.sync.b32 r0|p, %1, %2, %3, %4;"