to_numpy() broken on macos 12.6 and vulkan

[joshu-vp]

to_numpy() broken after upgrade to macos 12.6, on M1 chip

prints x.to_numpy() as [0, 0, 0, ..., 0]

import taichi as ti
import numpy as np

ti.init(arch=ti.vulkan)
num_particles = 1000

x = ti.Vector.field(2, dtype=ti.f32, shape=num_particles)

@ti.kernel
def substep():
    for i in range(num_particles):
        x[i] += x[i]

def main():
    for i in range(num_particles):
        x[i] = ti.Vector(np.random.rand(2), dt=ti.f32)
    substep()
    print(x.to_numpy())

main()

[Linyou]

The methods from_numpy() and from_torch() are also broken too. Currently, all methods involving external arrays are not working, but they work fine if I switch to metal backend. The only way to transfer these external arrays to taichi is via the python scope.

import taichi as ti
import numpy as np
import torch

ti.init(arch=ti.vulkan)


a = ti.field(dtype=ti.f32, shape=(10,))

float32 = torch.ones(10, dtype=torch.float32)
print("float32: ", float32)

a.from_torch(float32)

for i in range(10):
    print("a: ", a[i])

@ti.kernel
def from_numpy(dst: ti.template(), src: ti.types.ndarray()):
    for I in ti.grouped(src):
        dst[I] = src[I]

def from_numpy_python(dst, src):
    for i in range(dst.shape[0]):
        dst[i] = src[i]

float32 = np.array([1.0]*10).astype(np.float32)

from_numpy(a, float32)

for i in range(10):
    print("a: ", a[i])

from_numpy_python(a, float32)

for i in range(10):
    print("a: ", a[i])

[bobcao3]

Sounds like an synchronization issue... Can you try to add a ti.sync() before and after kernels involving external arrays

[Linyou]

I have tried ti.sync() but the code is still not working as expected @bobcao3

[strongoier]

I'll close this issue because #6295 is essentially talking about the same problem and we'll update our progress there.