>From merged pr https://github.com/apache/tvm/pull/17278/files . We can do code
>generation for `LetStmt`. For example, if we disable let-inlining in the
>Simplify Pass:
```python
iter_var: T.int32() = T.ceildiv(K, block_K)
for ko in T.serial(iter_var):
...
```
The generated CUDA code becomes:
```c
int32_t iter_var = K / Block_K;
for (int32_t k = 0; k < iter_var; k++){
...
}
```
This is powerful, but I'd like to discuss feasible solutions for cleanly
defining local variables. For instance, when using single-element buffers in
TVM:
```python
a = T.alloc_buffer([1], dtype="int32", scope="local")
a[0] += 1
a[0] -= 0
```
We encounter inconveniences:
1. Numerous `[0]` indices clutter the code (non-essential and inelegant)
2. Generated code becomes verbose:
```c
int32_t a[1];
a[0] = a[0] + 1;
a[0] = a[0] - 1;
```
**Possible Solutions:**
**Approach 1: Extend LetStmt**
```python
a: T.int32() = 0 # Define
a += 1 # Reassign (currently causes redefinition error)
```
Would generate cleaner C:
```c
int32_t a = 0;
a = a + 1;
```
*Concern:* This might violate `Let` semantics (immutable bindings).
**Approach 2: Introduce Local Variables via Buffer Extension**
This would require AST modifications to support first-class local variables
instead of single-element buffers.
Looking forward to community thoughts on these approaches or alternative
solutions.
---
[Visit
Topic](https://discuss.tvm.apache.org/t/how-to-design-a-local-variable/18490/1)
to respond.
You are receiving this because you enabled mailing list mode.
To unsubscribe from these emails, [click
here](https://discuss.tvm.apache.org/email/unsubscribe/7d282001f937071556497b5f3cf2604d7f3f98726526a7d1d1c709cabd51e80c).
