lhutton1 commented on PR #16782:
URL: https://github.com/apache/tvm/pull/16782#issuecomment-2045202993
Thanks @ekalda @Lunderberg @cbalint13! Let's investigate adding the target
dependent error message in a follow-up
--
This is an automated message from the Apache Git Service.
To respond
lhutton1 merged PR #16782:
URL: https://github.com/apache/tvm/pull/16782
--
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:
ekalda commented on PR #16782:
URL: https://github.com/apache/tvm/pull/16782#issuecomment-2037036256
Thanks for all the reviews and discussion! The latest version now includes
changes as a response to @lhutton1 review. I also looked into using the target
info in the LoopVectorizer and
ekalda commented on code in PR #16782:
URL: https://github.com/apache/tvm/pull/16782#discussion_r1551562798
##
src/tir/transforms/vectorize_loop.cc:
##
@@ -635,19 +701,22 @@ class Vectorizer : public StmtMutator, public
ExprFunctora) && b.same_as(op->b)) {
return
ekalda commented on code in PR #16782:
URL: https://github.com/apache/tvm/pull/16782#discussion_r1551562489
##
src/tir/transforms/vectorize_loop.cc:
##
@@ -433,20 +488,27 @@ class Vectorizer : public StmtMutator, public
ExprFunctorVisitExpr(op->value);
if
ekalda commented on code in PR #16782:
URL: https://github.com/apache/tvm/pull/16782#discussion_r1551562114
##
src/tir/transforms/vectorize_loop.cc:
##
@@ -182,21 +199,29 @@ class Vectorizer : public StmtMutator, public
ExprFunctora) && b.same_as(op->b)) {
return
ekalda commented on code in PR #16782:
URL: https://github.com/apache/tvm/pull/16782#discussion_r1551560120
##
tests/python/tir-transform/test_tir_transform_vectorize.py:
##
@@ -64,28 +61,86 @@ def test_vectorize_vector():
assert isinstance(stmt.body.value,
ekalda commented on code in PR #16782:
URL: https://github.com/apache/tvm/pull/16782#discussion_r1551559396
##
src/tir/ir/expr.cc:
##
@@ -196,7 +196,9 @@ TVM_REGISTER_NODE_TYPE(StringImmNode);
// Cast
Cast::Cast(DataType t, PrimExpr value, Span span) {
Lunderberg commented on PR #16782:
URL: https://github.com/apache/tvm/pull/16782#issuecomment-2028775777
That's a good point. This only applies for cases that would currently
produce an error, where a `ForKind::kVectorized` is applied to a loop with
dynamic extent.
I'm convinced,
cbalint13 commented on PR #16782:
URL: https://github.com/apache/tvm/pull/16782#issuecomment-2027792331
Trying to reason loudly on this,
> When it comes to targets that don't support SVE, I'd expect these targets
to not trigger the creation of scalable vectors. In the current
lhutton1 commented on code in PR #16782:
URL: https://github.com/apache/tvm/pull/16782#discussion_r1542695305
##
src/tir/transforms/vectorize_loop.cc:
##
@@ -37,19 +37,36 @@
namespace tvm {
namespace tir {
-// TODO(ekalda): P5 in https://github.com/apache/tvm/issues/16455
ekalda commented on PR #16782:
URL: https://github.com/apache/tvm/pull/16782#issuecomment-2020983646
Thank you for your feedback @Lunderberg, much appreciated!
> The implementation looks reasonable, though I have one main question for
it: What is the behavior of the updated pass for
lhutton1 commented on code in PR #16782:
URL: https://github.com/apache/tvm/pull/16782#discussion_r1539232691
##
src/tir/ir/expr.cc:
##
@@ -196,7 +196,9 @@ TVM_REGISTER_NODE_TYPE(StringImmNode);
// Cast
Cast::Cast(DataType t, PrimExpr value, Span span) {
Lunderberg commented on PR #16782:
URL: https://github.com/apache/tvm/pull/16782#issuecomment-2020540959
The implementation looks reasonable, though I have one main question for it:
What is the behavior of the updated pass for a target that doesn't support SVE?
Prior SVE-commits enabled
ekalda commented on PR #16782:
URL: https://github.com/apache/tvm/pull/16782#issuecomment-2018282951
cc the usual suspects @lhutton1 @leandron @Anndrey24 @tqchen @Lunderberg
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub
ekalda opened a new pull request, #16782:
URL: https://github.com/apache/tvm/pull/16782
This patch add support for turning loops marked for vectorizing into
scalable vectors if the extent of the loop is a vscale dependent expression in
a correct form.
The testing for both scalable
16 matches
Mail list logo
