merrymercy commented on a change in pull request #6042:
URL: https://github.com/apache/incubator-tvm/pull/6042#discussion_r453785929
##########
File path: src/tir/transforms/unroll_loop.cc
##########
@@ -57,6 +58,9 @@ struct UnrollLoopConfigNode : public
tvm::AttrsNode<UnrollLoopConfigNode> {
TVM_ATTR_FIELD(explicit_unroll)
.describe("Whether to explicitly unroll the loop instead of setting a
pragma")
.set_default(true);
+ TVM_ATTR_FIELD(explicit_unroll_max_extent)
+ .describe("The maximum extent of a loop that can be unrolled
explicitly (-1 for infinite)")
+ .set_default(32);
Review comment:
I think 32 is a very safe value. How do you know 32 is too large for ARM
CPU?
This value is just to avoid the llvm hanging issue, not for code
performance. (i.e. The schedule developers should choose a reasonable unroll
factor by themself.)
As a global configuration, this default value should be large enough to
avoid performance regression.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
[email protected]
