Hi! Since I’m a newbie to Taichi, I thought it could have been appropriate to discuss my problem here before opening an issue on Github, so to be sure I’m not doing anything fundamentally wrong.
I wrote this simple program for measuring performance:
import taichi as ti
import timeit
@ti.data_oriented
class Laplacian2d:
@ti.kernel
def run(self, in_phi: ti.template(), out_lap: ti.template()):
for i, j, k in out_lap:
out_lap[i, j, k] = (
-4 * in_phi[i, j, k]
+ in_phi[i - 1, j, k]
+ in_phi[i + 1, j, k]
+ in_phi[i, j - 1, k]
+ in_phi[i, j + 1, k]
)
def main(arch):
ti.init(arch=arch)
shape = (128, 128, 128)
nt = 1000
phi = ti.field(float, shape=shape)
lap = ti.field(float, shape=shape)
laplacian = Laplacian2d()
start = timeit.default_timer()
for _ in range(nt):
laplacian.run(phi, lap)
stop = timeit.default_timer()
print(f"Elapsed time: {stop - start} s")
if __name__ == "__main__":
main(ti.gpu)
The kernel compiles and runs fine, but when it comes to releasing device memory (I guess), it throws [cuda_driver.h:operator()@80] CUDA Error CUDA_ERROR_CONTEXT_IS_DESTROYED: context is destroyed while calling stream_synchronize (cuStreamSynchronize). Any idea?
I’m using Python3.8 and CUDA 11.0 on a Linux machine.