Memory leak

[jiajun-c]

When variables are re-declared each time through the loop like below， the memory can not be released.

        z_temp = ti.Vector.field(3, dtype=ti.f32, shape=1000000) # the particle position
        zz_temp = ti.Vector.field(3, dtype=ti.f32, shape=1000000)

This code is well.

if __name__ == "__main__":
    while True:
        z_temp = ti.Vector.field(3, dtype=ti.f32, shape=1000000) # the particle position
        zz_temp = ti.Vector.field(3, dtype=ti.f32, shape=1000000)

This code is well too.

if __name__ == "__main__":
    with open("./data/scenes/demo_fluid_and_rigid.json", "r") as f:
        simulation_config = json.load(f)
    points, indices = getlines(simulation_config['configuration'])
    print(points)
    config = simulation_config['configuration']
    ps = particle_systemv4.ParticleSystemV4(simulation_config)
    # add fluid and rigid
    # ps.add_fluid_and_rigid()
    wcsph = WCSPHV2(ps)
    # wcsph.step()
    # wcsph.step()
    while True:
        for i in range(5):
            wcsph.step()

Whenever you merge these two codes, you will get a memory leak.

[jiajun-c]

For each code, the memory usage of gpu is lower than 500MB, after merge them, the memory usage will be bigger than 24GB after some time steps.

[jiajun-c]

Full code: git@github.com:jiajun-c/Ti-SPH.git

Branch: benchmark

command: python3 main_fluid_rigid.py

[jiajun-c]

minimal code

import taichi as ti
ti.init(ti.cuda)

@ti.data_oriented
class Leaker:
    def __init__(self):
        pass
    
    @ti.kernel
    def foo(self):
        pass

@ti.kernel
def bar():
    pass

if __name__ == "__main__":
    # leaker = Leaker()
    while True:
        # leaker.foo()
        bar()
        print("11")
        # bugger = ti.Vector.field(3, dtype=ti.f32, shape=1000000) # the particle position
        bugger1 = ti.field(dtype=ti.i32, shape=1000000)