arch dx11 the atomic_add not working

[mgttt]

here is the testing code

from time import time as now

import taichi as ti
#ti.init(arch=ti.gpu) # ok for normal gpu but failed for directx
ti.init(arch=ti.dx11) # not ok
#ti.init(arch=ti.cpu) #ok
#ti.init()

ext_arr = ti.types.ndarray

@ti.kernel
def ti_test1(a:ext_arr(),b:ext_arr()):
  for I in ti.grouped(a): ti.atomic_add(b[I.sum()],a[I])
  
a = np.random.rand(3,4,5).astype(np.float32)
print('a',a)
b = np.zeros((sum(a.shape),),dtype=np.float32)
print('b',b)
print(a[1,1,1])
t1=now()
ti_test1(a,b)
ti.sync()
t2=now()
#print(a[1,1,1])
print('t2-t1',t2-t1)
print(b)
print(a.sum(),b.sum())

# should pass:
assert abs(a.sum()-b.sum())<0.001