Lunderberg commented on code in PR #16523:
URL: https://github.com/apache/tvm/pull/16523#discussion_r1479902583
##########
src/arith/pattern_match.h:
##########
@@ -628,10 +628,11 @@ inline PRampExpr<TBase, TStride, TLanes> ramp(const
Pattern<TBase>& base,
}
template <typename TBase>
-inline PRampExpr<TBase, PConstWithTypeLike<TBase>, PConst<int>> ramp(const
Pattern<TBase>& base,
- int
stride, int lanes) {
- return PRampExpr<TBase, PConstWithTypeLike<TBase>, PConst<int>>(
- base.derived(), PConstWithTypeLike<TBase>(base.derived(), stride),
PConst<int>(lanes));
+inline PRampExpr<TBase, PConstWithTypeLike<TBase>, PConstWithTypeLike<TBase>>
ramp(
+ const Pattern<TBase>& base, int stride, int lanes) {
+ return PRampExpr<TBase, PConstWithTypeLike<TBase>,
PConstWithTypeLike<TBase>>(
+ base.derived(), PConstWithTypeLike<TBase>(base.derived(), stride),
Review Comment:
Thank you for drawing attention to it, and I think this overload looks
reasonable with the `int lanes` argument. The purpose of this overload is to
define patterns that have a specific value, but whose dtype doesn't matter for
the purpose of matching. So, `ramp(pvar_base, 1, 4)` would match vectorized
groups of 4 values, regardless of whether the `4` was stored as
`IntImm(DataType::Int(32), 4)` or `IntImm(DataType::Int(64), 4)`.
--
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]
