masahi commented on code in PR #15439:
URL: https://github.com/apache/tvm/pull/15439#discussion_r1279109603
##########
src/tir/transforms/lower_thread_allreduce.cc:
##########
@@ -719,12 +719,16 @@ class ThreadAllreduceBuilder final : public
StmtExprMutator {
// Also, the warp/wavefront size differs (64 on rocm, 32 on cuda and metal).
bool IsWarpReduction(const std::vector<DataType>& types, int group_extent,
int reduce_extent,
int contiguous_reduce_extent) {
- if ((target_->kind->name != "cuda") && (target_->kind->name != "rocm") &&
- (target_->kind->name != "metal")) {
+ if (target_->kind->name == "vulkan") {
+ if (target_->GetAttr<Integer>("supported_subgroup_operations") == 0) {
Review Comment:
Here, ideally we should check the availability of each subgroup feature by
bit mask such as `VK_SUBGROUP_FEATURE_SHUFFLE_BIT`, and
`VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT`. But we cannot include a vulkan
header in this file, so for simplicity I'm assuming that non-zero
`supported_subgroup_operations` implies most common shuffle operations are
supported.
--
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]
