[MetaSchedule] Swap the order of RewriteTensorize and VerifyGPUCode t…

…o reduce tuning time (apache#13259) * [MetaSchedule] Swap the order of RewriteTensorize and VerifyGPUCode to reduce tuning time * add comment
driazati · Nov 2, 2022 · 7536068 · 7536068
1 parent 87f52af
commit 7536068
Showing 1 changed file with 3 additions and 1 deletion.
diff --git a/src/meta_schedule/postproc/postproc.cc b/src/meta_schedule/postproc/postproc.cc
@@ -77,8 +77,10 @@ Array<Postproc> Postproc::DefaultCUDATensorCore() {
       Postproc::RewriteUnboundBlock(/*max_threadblocks=*/256),
       Postproc::RewriteParallelVectorizeUnroll(),
       Postproc::RewriteReductionBlock(),
-      Postproc::RewriteTensorize(/*vectorize_init_loop=*/false),
       Postproc::VerifyGPUCode(),
+      // RewriteTensorize is relatively expensive and it doesn't affect the validity of a sample, so
+      // run it only on samples that have passed VerifyGPUCode.
+      Postproc::RewriteTensorize(/*vectorize_init_loop=*/false),
   };
 }