Fill unexpectedly crashes with CUDA backend

[turbo0628]

import taichi as ti

ti.init(arch=ti.cuda, device_memory_GB=2) # if we give 2.1, it works well.

arr = ti.ndarray(ti.f32, 128 * 1024 * 1024) # This should only use 1GB of GPU memory
arr.fill(1.2345)

fails with CUDA backend, err msg:

[E 12/21/22 22:03:04.230 65462] [cuda_driver.h:operator()@88] CUDA Error CUDA_ERROR_INVALID_VALUE: invalid argument while calling memsetd32 (cuMemsetD32_v2)

The regression happened between v1.0.2 and v1.0.3

[turbo0628]

device memory has to be 4x larger than ndarray to work normally, minimal reprod:

import taichi as ti

ti.init(arch=ti.cuda, print_ir=True, device_memory_GB=1.1) # crashes until we raise device memory to 4.1GB

n = 1024 * 256 * 1024 # 1GB

arr = ti.ndarray(ti.f32, n)

arr.fill(0.125)

[turbo0628]

Root cause: The fill bug is due to the change in ndarray element size definitions.

The pass-in size argument is expected to be used for i32 addressing, but it now counts bytes. In the above example the size has value of 1024 * 1024 * 1024 but expected 1024 * 256 * 1024.