MatrixFreeCG is not working with 1.60

[fangde]

MatrixFree is not working
by following the blog (https://mp.weixin.qq.com/s/BrgY5U4WBjqPo1smf4DkbA) of MatrixFreeCG in taichi . we test the matrix free cg solver.
none of api(taichi_cg_solver or MatrixFreeCG) is working.

To Reproduce
we use the sample code in the blog

# matrix freee cg solver

import taichi as ti
arch = ti.cuda# or ti.cuda
ti.init(arch=arch)
n = 4
x = ti.field(dtype=ti.f32, shape=n) # x 为未知数向量
Ax = ti.field(dtype=ti.f32, shape=n) # Ax 为矩阵向量点乘的结果
@ti.kernel
def compute_matrix_vector(v:ti.template(), mv:ti.template()):
    for i in v:
        mv[i] = 2 * v[i]
        
compute_matrix_vector(x, Ax)
A = ti.linalg.LinearOperator(compute_matrix_vector)
b = ti.field(ti.f32, shape=n)
@ti.kernel
def init_b():
    for i in b:
        b[i] = 1.0
init_b()
print(b)

ti.linalg.taichi_cg_solver(A, b, x)

the output is

[Taichi] Starting on arch=cuda
[1. 1. 1. 1.]
---------------------------------------------------------------------------
TaichiIndexError                          Traceback (most recent call last)
Cell In[4], line 22
     19 init_b()
     20 print(b)
---> 22 ti.linalg.taichi_cg_solver(A, b, x)

File [~/anaconda2/envs/taichi16/lib/python3.10/site-packages/taichi/linalg/taichi_cg.py:100](https://vscode-remote+ssh-002dremote-002bdlhome-002elocal.vscode-resource.vscode-cdn.net/home/fliu/taichisparse_test/~/anaconda2/envs/taichi16/lib/python3.10/site-packages/taichi/linalg/taichi_cg.py:100), in taichi_cg_solver(A, b, x, tol, maxiter, quiet)
     97         if not quiet:
     98             print(f">>> Iter = {i+1:4}, Residual = {sqrt(new_rTr):e}")
--> 100 solve()
    101 vector_fields_snode_tree.destroy()
    102 scalar_snode_tree.destroy()

File [~/anaconda2/envs/taichi16/lib/python3.10/site-packages/taichi/linalg/taichi_cg.py:75](https://vscode-remote+ssh-002dremote-002bdlhome-002elocal.vscode-resource.vscode-cdn.net/home/fliu/taichisparse_test/~/anaconda2/envs/taichi16/lib/python3.10/site-packages/taichi/linalg/taichi_cg.py:75), in taichi_cg_solver..solve()
     74 def solve():
---> 75     init()
     76     initial_rTr = reduce(r, r)
     77     if not quiet:

File [~/anaconda2/envs/taichi16/lib/python3.10/site-packages/taichi/lang/kernel_impl.py:976](https://vscode-remote+ssh-002dremote-002bdlhome-002elocal.vscode-resource.vscode-cdn.net/home/fliu/taichisparse_test/~/anaconda2/envs/taichi16/lib/python3.10/site-packages/taichi/lang/kernel_impl.py:976), in _kernel_impl..wrapped(*args, **kwargs)
    974     return primal(*args, **kwargs)
    975 except (TaichiCompilationError, TaichiRuntimeError) as e:
--> 976     raise type(e)("\n" + str(e)) from None

TaichiIndexError: 
File "/home/fliu/anaconda2/envs/taichi16/lib/python3.10/site-packages/taichi/linalg/taichi_cg.py", line 48, in init:
            r[I] = b[I]
            ^^^^
Field with dim 2 accessed with indices of dim 1

we also test the new name api as

ti.linalg.MatrixFreeCG(A,x)

MatrixFree is not named,output as follows

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[7], line 1
----> 1 ti.linalg.MatrixFreeCG(A,x)

AttributeError: module 'taichi.linalg' has no attribute 'MatrixFreeCG'

Additional comments
If possible, please also consider attaching the output of command ti diagnose. This produces the detailed environment information and hopefully helps us diagnose faster.

If you have local commits (e.g. compile fixes before you reproduce the bug), please make sure you first make a PR to fix the build errors and then report the bug.

[houkensjtu]

Hi @fangde ! Thank you for trying out our MatrixFreeCG. The issue was two-fold:

1. MatrixFreeCG does not exist in 1.6.0 and will be released in 1.7.0. That's why you saw the following error message

AttributeError: module 'taichi.linalg' has no attribute 'MatrixFreeCG'

2. The RuntimeError is caused by a previous (known) bug in our CG solver, fortunately it's solved by PR#8070([bug] Fix MatrixFreeCG so it can handle multiple input sizes. #8070). Again, this bug fix will come with 1.7.0. For the moment, please try to use shape=(n,1) instead of shape=n for one-dimension vectors.

Let me know if you have further questions!