guocju opened a new issue, #18684:
URL: https://github.com/apache/tvm/issues/18684
A small test:
``` python
import tvm
from tvm.script import ir as I
from tvm.script import tir as T
@tvm.register_global_func("tir_print")
def tir_print(A):
print(A)
@I.ir_module
class TestCeilDiv:
@T.prim_func
def main(n: T.int32) -> None:
T.func_attr({"target": T.target("llvm")})
res : T.int32 = T.ceildiv(n, T.int32(10))
T.call_packed("tir_print", res)
mod = TestCeilDiv
exe = tvm.build(mod)
exe["main"](24)
```
### Expected behavior
Expected result of ceildiv(24,10) is 3.
### Actual behavior
The result of this test is 4.
### debug
The direct cause is incorrect optimization of TIR by the LowerIntrin pass.
after LowerIntrin pass, tir became:
`T.Cast("int64", T.Div(n - 2147483631, 10) - -214748364)`
This is obviously wrong.
In src/tir/transforms/lower_intrin.cc:156
`PrimExpr ceildiv = truncdiv(-min + (op->b - 1), op->b);`
In my test case, min is -2147483648+9.
b-1 is 9.
`-min + (op->b - 1)` = 2147483648, after const fold in
src/arith/const_fold.h:80, the return value is -2147483648, because 2147483648
exceeds the upper limit of int32.
This resulted in a non-equivalent transformation.
### Possible fix method
To avoid overflow of '(op->b - 1) - min' , we can assert (b-1-a_min) <
MAX_VALUE+1.
That is (b_max-a_min) < MAX_VALUE + 2.
We can avoid this issue by changing the condition of the if statement in
L122.
```cpp
if (const_int_bound_a->min_value < 0 &&
const_int_bound_a->min_value >
(const_int_bound_b->max_value -
(Downcast<IntImm>(tvm::max_value(op->a->dtype.element_of()))->value) - 2)) {
```
Of course, this may not be the best fix.
Because the floormod node uses similar optimization，it also has the same
problem, in src/tir/transforms/lower_intrin.cc:217.
* needs-triage
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
