Nearest neighbor failure

[ChenhanYu]

## problem size
#n=65536
#n=147456
#n=262144
n=100000
#n=102400
#n=1048576
#n=2097152
## maximum leaf node size
m=64
## maximum off-diagonal ranks
s=2048
## number of neighbors
k=128
## number of right hand sides
nrhs=128
## user tolerance
stol=1E-3
## user computation budget
budget=0.03
## distance type (geometry, kernel, angle)
distance="angle"
## spdmatrix type (testsuit, dense, ooc, mlp, kernel, userdefine)
matrixtype="kernel"
## kernelmatrix type (gaussian, laplace)
kerneltype="gaussian"
#kerneltype="laplace"
## hidden layer configuration (512-512-512)
hiddenlayer="512-512-512"
# ======= Do not change anything below this line ========
mpiexec="ibrun tacc_affinity"
executable="./test_gofmm"

[ChenhanYu]

This bug is not able to be reproduced using test_mpigofmm. As a result, this is the problem solely for the shared memory version.

[ChenhanYu]

This problem has been resolved in the following commit on the develop branch. Close the issue now.

commit 51ba3d7ce16d34aec06f302ea941c905ac827d4a Author: Chenhan Yu ych@login3.stampede2.tacc.utexas.edu Date: Tue Jan 22 18:32:43 2019 -0600

Bug fixed for the leaf node size. Will deprecate members in
tree:Steup since they are duplicated with Configuration.

Instead of using a fixed range to decide the middle
of the median split, we now gradually increase the
left and right quantiles.

The left and right quantiles increase 10% every time
if they ended up being the same as the median.