strange behavior using atomic operation with snode on cuda

[yhesper]

Describe the bug
Using atomic_min when looping through snode gives wrong result on cuda.

To Reproduce

import taichi as ti

ti.init(arch=ti.gpu)

x = ti.field(ti.f32)
block = ti.root.pointer(ti.ij, (1000,1))
pixel = block.dense(ti.ij, (1,1))
pixel.place(x)

@ti.kernel
def test():
    hit_time = 1000.0
    x[100, 0] = 1.0
    x[200, 0] = 2.0
    for i,j in x:
        # print(hit_time, i, j, x[i,j])
        # hit_time += 0.0
        ti.atomic_min(hit_time, x[i,j])
    print(hit_time)

test()

Log/Screenshots
The above program should print 1.0. However, it will print 0.0 if we don't uncomment one of the commented lines. This seems to only happen when arch is cuda.