!6 fix inefficient auto tiling for axis with tail and remove duplicated check

Merge pull request !6 from yangsijia/fix-issue-I1LEYY

src/poly/tiling_analyzer.cc

@@ -444,12 +444,10 @@ int64_t TileCandidate::CalActualTile(const CalAlignInfo *align_info) {
 void TileCandidate::UpdateMemoryAfterBuffer(const BufferEntry *buf, MemInferInfo *mem_infer_info) {
   CHECK(buf);
   CHECK(mem_infer_info);
-  CHECK(buf);
   const auto fix_size = buf->shape.as<IntImm>();
   if (fix_size == nullptr) {
     std::stringstream ss;
     ss << "Buffer " << buf->name << " contains dynamic shape " << buf->shape << ", skip.";
-
     analyzer_->logger_.AppendLog(DO_TILING, ss);
     return;
   }

src/poly/tiling_solver.cc

@@ -936,16 +936,17 @@ bool TraverseSolver::DoTiling(const TileInfo *info) {
 
     if (!mem_ok) continue;
     success = true;
-    if (dst % t == 0) {
+    auto tail = dst % t;
+    if (tail == 0) {
       if (deviation > best_no_iso_devs) continue;
       ss << "factor " << t << " has " << deviation << " deviation, update to no isolate factor";
       best_no_iso_val = t;
       best_no_iso_devs = deviation;
     } else {
       if (deviation > best_devs) continue;
-      if (analyzer_.scop_->pragma_allow_tail_tiling_) {
-        auto tail = dst - dst / t * t;
-        if (tail % GetMaxAlignBytes(axis->data_size) != 0) continue;
+      if (analyzer_.scop_->pragma_allow_tail_tiling_ && tail < GetMaxAlignBytes(axis->data_size)) {
+        ss << "factor " << t << " has " << tail << " tail that may disable multicore, skip.";
+        continue;
       }
       ss << "factor " << t << " has " << deviation << " deviation, update to isolate factor";
       best_val = t;