merrymercy commented on a change in pull request #6877:
URL: https://github.com/apache/incubator-tvm/pull/6877#discussion_r521989104
##########
File path: src/auto_scheduler/search_policy/sketch_policy.cc
##########
@@ -116,8 +116,8 @@ SketchPolicy::SketchPolicy(SearchTask task, CostModel
program_cost_model,
} else if (IsGPUTask(node->search_task)) {
// Sketch Generation Rules
node->sketch_rules.push_back(&rule_add_cache_read_stage);
- node->sketch_rules.push_back(&rule_always_inline);
node->sketch_rules.push_back(&rule_special_compute_location_gpu);
+ node->sketch_rules.push_back(&rule_always_inline);
Review comment:
The old order generates wrong sketches for winograd.
For some stages, we have to apply `rule_special_compute_location_gpu`. But
the older order will just inline them.
----------------------------------------------------------------
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:
us...@infra.apache.org
