antinucleon commented on a change in pull request #7132:
URL: https://github.com/apache/tvm/pull/7132#discussion_r547936669
##########
File path: src/auto_scheduler/search_policy/sketch_policy.cc
##########
@@ -115,21 +115,35 @@ SketchPolicy::SketchPolicy(SearchTask task, CostModel
program_cost_model,
node->mutation_rules.push_back(std::make_shared<MutateParallel>(0.01));
} 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_special_compute_location_gpu);
- node->sketch_rules.push_back(&rule_always_inline);
- node->sketch_rules.push_back(&rule_simplify_compute_with_const_tensor);
- node->sketch_rules.push_back(&rule_cross_thread_reduction);
- node->sketch_rules.push_back(&rule_add_cache_write_stage);
- node->sketch_rules.push_back(&rule_multi_level_tiling_with_fusion);
- node->sketch_rules.push_back(&rule_multi_level_tiling);
- node->sketch_rules.push_back(&rule_skip_stage);
+ if (node->search_task->target->GetAttr<String>("device", "") == "mali") {
+ node->sketch_rules.push_back(&rule_always_inline);
Review comment:
@FrozenGene @merrymercy I think it worth to export sketch API to Python,
make it more configurable. For a given workload, we may want to try multiple
sketches pattern, exporting will enable these experiments without rebuild.
----------------------------------------------------------------
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]
