在使用稀疏数据结构和自动微分实现一个仿真算法时遇到了难以解决的报错，在此提供最小可复现的代码，希望有大佬解答
import taichi as ti
ti.init(arch=ti.cuda)
global_table = {}
tmp = ti.Vector.field(3, float, shape=(4, 4))
tmp.fill((0.5, 0.5, 0.5))
def init_global_table():
global_table["debug"] = tmp
init_global_table()
space_blocks = ti.field(ti.i32)
S1 = ti.root.pointer(ti.ijk, (2, 2, 2))
S2 = S1.bitmasked(ti.ijk, (2, 2, 2))
array = S2.dynamic(ti.l, 4, chunk_size=32)
array.place(space_blocks)
@ti.kernel
def shard():
for i, j, k, l in ti.ndrange(4, 4, 4, 4):
space_blocks[i, j, k].append(0)
shard()
p = ti.Vector.field(3, dtype=float, shape=(), needs_grad=True)
p[None] = (0.7, 0.7, 0.7)
loss = ti.field(float, shape = (), needs_grad=True)
loss.grad[None] = 1
loss[None] = 0
label = ti.Vector.field(3, dtype=float, shape=(), needs_grad=False)
label[None] = (0.1, 0.2, 0.5)
@ti.kernel
def loss_func():
# 或者 for i in ti.group(x) # x是一个field
for _ in range(4):
x = int(p[None][0] // 1)
y = int(p[None][1] // 1)
z = int(p[None][2] // 1)
for index in range(space_blocks[x, y, z].length()):
target = global_table["debug"][space_blocks[x, y, z, index], space_blocks[x, y, z, index]]
loss[None] -= (target.dot(p[None]) - label[None].norm()) * (target.dot(p[None]) - label[None].norm())
'''
# 这样不会报错,但程序逻辑不同
x = int(p[None][0] // 1)
y = int(p[None][1] // 1)
z = int(p[None][2] // 1)
@ti.kernel
def loss_func():
for index in range(space_blocks[x, y, z].length()):
target = global_table["debug"][space_blocks[x, y, z, index], space_blocks[x, y, z, index]]
loss[None] -= (target.dot(p[None]) - label[None].norm()) * (target.dot(p[None]) - label[None].norm())
'''
with ti.ad.Tape(loss=loss):
loss_func()
print('grad: ',p.grad[None])
