The same code works on Nvidia CUDA but doesn't work on AMD Vulkan

[ustcfdm]

Describe the bug When I run a Taichi kernel with Nvidia CUDA, it always works well. However, when I try it with AMD Vulkan, it will fail.

To Reproduce Here is a sample code.

import time
import numpy as np
import taichi as ti

ti.init(arch=ti.gpu)

@ti.kernel      # Test kernel
def test_kernel(a: ti.types.ndarray(ndim=3), b: ti.types.ndarray(ndim=3)):
    nslices, nrow, ncols = a.shape
    for sli, row, col in b:
        for n in range(nslices):
            b[sli, row, col] += a[n, row, col]
        b[sli, row, col] /= nslices

m = 64      # Test data size
m = 512

a = np.random.random((m, m, m)).astype('float32')   
b = np.zeros_like(a)

t1 = time.time()
test_kernel(a, b)
t2 = time.time()
print(f'Time: {t2-t1}s')

Log/Screenshots For Nvidia CUDA, here is the result: When m = 64,

[Taichi] version 1.7.0, llvm 15.0.1, commit 2fd24490, win, python 3.10.10
[Taichi] Starting on arch=cuda
Time: 0.049997806549072266s

When m = 512,

[Taichi] version 1.7.0, llvm 15.0.1, commit 2fd24490, win, python 3.10.10
[Taichi] Starting on arch=cuda
Time: 1.6580820083618164s

For AMD Vulkan, here is the result: When m = 64, it works.

[Taichi] version 1.7.1, llvm 15.0.1, commit 0f143b2f, win, python 3.10.6
[Taichi] Starting on arch=vulkan
Time: 0.040009260177612305s

However, when m = 512, it fails.

'C:\x5cUsers\x5cMango\x5cDesktop\x5cWFH\x5ctest.py' ;949194d8-2dc7-4409-bbea-16e3abdba9d4[Taichi] version 1.7.1, llvm 15.0.1, comm[W 07/27/24 19:56:04.137 17488] [cuda_driver.cpp:taichi::lang::CUDADriverBase::load_lib@36] nvcuda.dll lib not found.
RHI Error: (0) Vulkan device might be lost (vkQueueSubmit failed)
Assertion failed: false && "Error without return code", file C:\Users\buildbot\actions-runner\_work\taichi\taichi\taichi\rhi\vulkan\vulkan_device.cpp, line 2038

If I try it again, my screen goes black. I have to reboot my computer.

Additional comments Nvidia version: GeForce RTX 2080 SUPER 8GB AMD version: Radeon RX 7700 XT 12 GB