[GitHub] [tvm] masahi commented on a diff in pull request #13621: [TIR][TOPI][x86][CI] Support skylake avx512

[GitBox <gi...@apache.org>]

masahi commented on code in PR #13621:
URL: https://github.com/apache/tvm/pull/13621#discussion_r1071618299


##########
src/meta_schedule/space_generator/space_generator.cc:
##########
@@ -93,6 +103,10 @@ void SpaceGeneratorNode::InitializeWithTuneContext(const TuneContext& context) {
       default_sch_rules = ScheduleRule::DefaultVNNI();
       default_postprocs = Postproc::DefaultVNNI();
       default_mutator_probs = Mutator::DefaultVNNI();
+    } else if (kind == "avx512") {
+      default_sch_rules = ScheduleRule::DefaultAVX512();
+      default_postprocs = Postproc::DefaultVNNI();

Review Comment:
   Rename `DefaultVNNI` to `DefaultCPUTensorization`



##########
src/meta_schedule/schedule_rule/schedule_rule.cc:
##########
@@ -130,6 +130,51 @@ Array<ScheduleRule> ScheduleRule::DefaultVNNI() {
   };
 }
 
+Array<ScheduleRule> ScheduleRule::DefaultAVX512() {

Review Comment:
   Unify `DefaultVNNI` and this function.



##########
tests/python/unittest/test_meta_schedule_trace_apply.py:
##########
@@ -25,6 +25,9 @@
 from tvm.target import Target

Review Comment:
   Please remove change from this file, since VNNI vs AVX512 difference doesn't matter for this unittest.



##########
tests/python/unittest/test_tir_schedule_analysis.py:
##########
@@ -41,6 +41,71 @@
 from tvm.te import create_prim_func

Review Comment:
   Also remove change from this file



##########
src/meta_schedule/space_generator/space_generator.cc:
##########
@@ -93,6 +103,10 @@ void SpaceGeneratorNode::InitializeWithTuneContext(const TuneContext& context) {
       default_sch_rules = ScheduleRule::DefaultVNNI();
       default_postprocs = Postproc::DefaultVNNI();
       default_mutator_probs = Mutator::DefaultVNNI();
+    } else if (kind == "avx512") {
+      default_sch_rules = ScheduleRule::DefaultAVX512();
+      default_postprocs = Postproc::DefaultVNNI();
+      default_mutator_probs = Mutator::DefaultVNNI();

Review Comment:
   Remove `DefaultVNNI` and use `DefaultLLVM` (they are the same) for both VNNI and AVX512 



##########
src/meta_schedule/space_generator/space_generator.cc:
##########
@@ -73,6 +81,8 @@ void SpaceGeneratorNode::InitializeWithTuneContext(const TuneContext& context) {
     Array<ScheduleRule> default_sch_rules;
     Array<Postproc> default_postprocs;
     Map<Mutator, FloatImm> default_mutator_probs;
+    // TODO(vvchernov): check if need separated ScheduleRule, Postproc, Mutator

Review Comment:
   seperated -> dedicated



##########
tests/python/unittest/test_meta_schedule_16x4_integration.py:
##########
@@ -28,6 +28,7 @@
 from tvm.tir.schedule import BlockRV, Schedule

Review Comment:
   Let's rename it to `test_meta_schedule_cpu_dot_product.py`



-- 
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: commits-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org