[Refactor] Optimize debug message for parallel inference (#1096)

src/layout/utils.cc

@@ -93,7 +93,9 @@ Array<IterSplitExpr> get_unused_iters(const IterMark &mark,
     if (j == splits.size()) {
       ICHECK(lowest != splits.size());
       ICHECK(CanProveDivisible(splits[lowest]->lower_factor,
-                               expected_lower_factor));
+                               expected_lower_factor))
+          << " Cannot prove divisible for " << splits[lowest]->lower_factor
+          << " and " << expected_lower_factor;
       results.emplace_back(
           mark, expected_lower_factor,
           FloorDiv(splits[lowest]->lower_factor, expected_lower_factor), 1);

src/op/parallel.cc

@@ -373,8 +373,21 @@ LayoutMap ParallelOpNode::InferLayout(const LayoutInferArgs &T,
           throw LayoutConflictException(oss.str());
         }
       });
-      result = Fragment(loop_vars_, {}, loop_var_to_thread, rep_iter)
-                   ->BindThreadRange(T.thread_bounds);
+
+      try {
+        result = Fragment(loop_vars_, {}, loop_var_to_thread, rep_iter)
+                     ->BindThreadRange(T.thread_bounds);
+      } catch (const tvm::runtime::Error &err) {
+        std::ostringstream msg;
+        msg << "Layout inference for buffer `" << buffer->name
+            << "` failed inside `T.parallel` loop.";
+
+        msg << "\nUnderlying TVM error: " << err.what();
+        msg << "\nProblematic loop AST:\n " << root_;
+        msg << "\nHint: ensure the loop extent divides the thread binding or "
+               "adjust the fragment mapping.";
+        LOG(FATAL) << msg.str();
+      }
     }
     DLOG(INFO) << "[compute_loop_layout_from_buffer] ... and get "
                << result->DebugOutput() << '\n';