[Layout] Introduce Flexible Parallel to Support T.serial and local buffers...

[Layout] Introduce Flexible Parallel to Support T.serial and local buffers inside T.Parallel loop (#844)

* Support T.serial and local buffers inside T.Parallel loop.

* Fix reducer layout in T.Parallel nested inside other loops

* Debug output with LOG(INFO)

* Add disable option for WGMMA.

* fix

* Use DLOG; fix missing registration for new pass config

* bug fix

* lint fix

* Enhance GEMM instruction set with UTCMMA and improve local buffer handling in casting example

* Update format.sh shebang, improve logging in layout inference, and enhance buffer store wrapper with detailed comments

* Enhance GEMM instantiation logic and improve layout inference for local buffer detection

- Updated the GEMM instantiation logic to include a check for WGMMA compatibility, ensuring that the conditions for using WGMMA are more robust.
- Refined the layout inference process to better identify when loops manipulate only local buffers, improving the accuracy of thread binding decisions in parallel loops.

---------
Co-authored-by: Huanqi Cao <caohuanqi@deepseek.com>

.clang-tidy

@@ -42,6 +42,7 @@ Checks: >
   -cppcoreguidelines-pro-type-static-cast-downcast,
   -performance-unnecessary-value-param,
   -performance-enum-size,
+  -clang-analyzer-deadcode.DeadStores,
 
 WarningsAsErrors: '*'
 

examples/cast/example_group_per_split_token_cast_to_fp8.py

@@ -29,7 +29,7 @@ def group_per_split_token_cast_to_fp8(M, M_max, N, BG, blk_m):
             y_s_local = T.alloc_fragment((blk_m,), accum_dtype)
             y_q_local = T.alloc_fragment((blk_m, group_size), accum_dtype)
             y_q_local_fp8 = T.alloc_fragment((blk_m, group_size), "float8_e4m3")
-            row_offset = T.alloc_local((1,), "int32")
+            row_offset = T.alloc_fragment((1,), "int32")
 
             T.annotate_layout({
                 y_local:

src/op/builtin.cc

@@ -29,6 +29,7 @@ TVM_REGISTER_PASS_CONFIG_OPTION(kDisableFastMath, Bool);
 TVM_REGISTER_PASS_CONFIG_OPTION(kEnableFastMath, Bool);
 TVM_REGISTER_PASS_CONFIG_OPTION(kPtxasRegisterUsageLevel, Integer);
 TVM_REGISTER_PASS_CONFIG_OPTION(kEnablePTXASVerboseOutput, Bool);
+TVM_REGISTER_PASS_CONFIG_OPTION(kDisableWGMMA, Bool);
 TVM_REGISTER_PASS_CONFIG_OPTION(kDisableShuffleElect, Bool);
 
 DataType cuTensorMapType() { return DataType::UInt(8, 128); }

src/op/builtin.h

@@ -45,6 +45,7 @@ static constexpr const char *kPtxasRegisterUsageLevel =
     "tl.ptxas_register_usage_level";
 static constexpr const char *kEnablePTXASVerboseOutput =
     "tl.enable_ptxas_verbose_output";
+static constexpr const char *kDisableWGMMA = "tl.disable_wgmma";
 static constexpr const char *kDisableShuffleElect = "tl.disable_shuffle_elect";
 /*!
  * \brief Whether to disable dynamic tail split

src/op/gemm.cc

@@ -92,10 +92,14 @@ TileOperator GemmNode::Clone() const {
 }
 
 GemmNode::GemmInst GemmNode::GetGemmInst(int block_size, Target target) const {
+  tvm::transform::PassContext ctxt = tvm::transform::PassContext::Current();
+
   int warp_size = TargetGetWarpSize(target);
   int num_warps = block_size / warp_size;
-  bool allow_wgmma = TargetIsHopper(target) && (this->M >= 64) &&
-                     (num_warps % 4 == 0) && CheckWGMMA();
+  bool allow_wgmma =
+      !ctxt->GetConfig(kDisableWGMMA, Optional<Bool>()).value_or(false) &&
+      TargetIsHopper(target) && (this->M >= 64) && (num_warps % 4 == 0) &&
+      CheckWGMMA();
   if (allow_wgmma) {
     return GemmInst::kWGMMA;
   } else if (TargetIsCDNA(target)) {

src/op/parallel.cc

@@ -128,9 +128,13 @@ private:
  * visitor's reducer_info_map_. Continues traversal into the loop body.
  */
 void ParallelLoopNestVisitor::VisitStmt_(const ForNode *op) {
-  ICHECK(op->kind == ForKind::kParallel);
-  p->loop_vars_.push_back(
-      IterVar(Range(op->min, op->extent), op->loop_var, IterVarType::kDataPar));
+  if (op->kind == ForKind::kParallel)
+    p->loop_vars_.push_back(IterVar(Range(op->min, op->extent), op->loop_var,
+                                    IterVarType::kDataPar));
+  else
+    p->inner_vars_.Set(op->loop_var,
+                       IterVar(Range(op->min, op->extent), op->loop_var,
+                               IterVarType::kOrdered));
   p->analyzer_.Bind(op->loop_var, Range::FromMinExtent(op->min, op->extent));
   auto reducer_info_map =
       op->annotations.Get(attr::kReducerInfo)->as<Map<Var, ReducerInfo>>();
@@ -244,17 +248,33 @@ LayoutMap ParallelOpNode::InferLayout(const LayoutInferArgs &T,
   }
   auto compute_loop_layout_from_buffer = [&](const Buffer &buffer) {
     Fragment src_layout = T.layout_map[buffer].as<Fragment>().value();
+    DLOG(INFO) << "[compute_loop_layout_from_buffer] infer from buffer `"
+               << buffer << "` of layout " << src_layout->DebugOutput() << '\n';
+    Fragment result;
     if (IsCommonAccessIndice(buffer)) {
-      return src_layout;
+      result = src_layout;
     } else {
       Var rep;
       auto rep_iter = IterVar({0, src_layout->ReplicateExtent()}, rep,
                               IterVarType::kDataPar);
       PrimExpr loop_var_to_thread =
           src_layout->ForwardThread(indice_map_[buffer], rep);
-      return Fragment(loop_vars_, {}, loop_var_to_thread, rep_iter)
+      loop_var_to_thread = analyzer_.Simplify(loop_var_to_thread);
+      PostOrderVisit(loop_var_to_thread, [&](const ObjectRef &objref) {
+        if (auto opt_var = objref.as<Var>();
+            opt_var && inner_vars_.count(*opt_var)) {
+          std::ostringstream oss;
+          oss << "loop_var_to_thread = " << loop_var_to_thread
+              << "contains inner var" << *opt_var;
+          throw LayoutConflictException(oss.str());
+        }
+      });
+      result = Fragment(loop_vars_, {}, loop_var_to_thread, rep_iter)
                    ->BindThreadRange(T.thread_bounds);
     }
+    DLOG(INFO) << "[compute_loop_layout_from_buffer] ... and get "
+               << result->DebugOutput() << '\n';
+    return result;
   };
   if (source_buffer.defined()) {
     loop_layout_ = compute_loop_layout_from_buffer(source_buffer);
@@ -317,15 +337,21 @@ LayoutMap ParallelOpNode::InferLayout(const LayoutInferArgs &T,
           IfBufferRemapLoopGenerator::run(root_, T.buffer_remap, T.layout_map);
       int vector_size = GetVectorizeSize(maybe_remapped_root_);
 
+      DLOG(INFO) << "[PlanLoopPartition] vector_size = " << vector_size << '\n';
+
       PrimExpr loop_total_size = 1;
       for (Stmt l = root_; l.as<For>().has_value();
            l = l.as<For>().value()->body)
         loop_total_size = loop_total_size * l.as<For>().value()->extent;
+      DLOG(INFO) << "[PlanLoopPartition] loop_total_size = " << loop_total_size
+                 << '\n';
       while (!analyzer_.CanProve(
                  floormod(loop_total_size,
                           T.thread_bounds->extent * vector_size) == 0) &&
              vector_size > 1)
         vector_size /= 2;
+      DLOG(INFO) << "[PlanLoopPartition] after adjust: vector_size = "
+                 << vector_size << '\n';
 
       // Check if coalesced_width is defined
       if (auto coalesced_width =
@@ -342,7 +368,12 @@ LayoutMap ParallelOpNode::InferLayout(const LayoutInferArgs &T,
           LOG(FATAL) << "coalesced_width should be an IntImmNode.";
         }
       }
+      DLOG(INFO) << "[PlanLoopPartition] root_ = " << root_
+                 << " ############# vector_size = " << vector_size
+                 << ", thread_bounds = " << T.thread_bounds << '\n';
       loop_layout_ = PlanLoopPartition(root_, vector_size, T.thread_bounds);
+      DLOG(INFO) << "[PlanLoopPartition] loop_layout_ = "
+                 << loop_layout_->DebugOutput() << '\n';
     }
   } else {
     return {};

src/op/parallel.h

@@ -128,6 +128,7 @@ private:
   void AddPredicate(const PrimExpr &expr) const {
     predicate_ = predicate_.defined() ? And(expr, predicate_.value()) : expr;
   }
+
   // Allow ParallelLoopNestVisitor to access private members.
   friend class ParallelLoopNestVisitor;
 
@@ -139,6 +140,8 @@ private:
   std::unordered_set<Buffer, ObjectPtrHash, ObjectPtrEqual> buffer_is_write_;
   // The loop variables for the parallel loop nest.
   Array<IterVar> loop_vars_;
+  // The inner_vars_
+  Map<Var, IterVar> inner_vars_;
   // Analyzer for simplifying and analyzing expressions, mutable for lazy use.
   mutable arith::Analyzer analyzer_;
   // Mapping from buffer to reducer info.

src/transform/layout_inference.cc

@@ -105,13 +105,16 @@ public:
            "required for layout inference.";
 
     // Run InferLayout
+    DLOG(INFO) << "[RunInferStep] working on " << cur_infer_id << '\n';
     auto updates =
         next->InferLayout(LayoutInferArgs{target_, thread_bounds, layout_map,
                                           &analyzer_, buffer_oob},
                           level);
-
     // Process the returned updates
     for (const auto &[buffer, layout] : updates) {
+      DLOG(INFO) << "    consider update " << buffer << " as "
+                 << layout->DebugOutput() << '\n';
+
       // Basic validity checks
       ICHECK(buffer.defined()) << "InferLayout returned an undefined buffer.";
       ICHECK(layout.defined()) << "InferLayout returned an undefined layout.";
@@ -140,6 +143,8 @@ public:
           if (ProveFragmentContains(src_layout, dst_layout, indices, indices,
                                     inner_analyzer)) {
             layout_map.Set(buffer, layout);
+            DLOG(INFO) << "    layout broadcast from "
+                       << src_layout->DebugOutput() << ", accepted" << '\n';
             continue;
           }
         }
@@ -151,6 +156,7 @@ public:
       } else {
         // Otherwise, update map
         layout_map.Set(buffer, layout);
+        DLOG(INFO) << "    new layout accepted" << '\n';
         if (!update_queue)
           continue;
 
@@ -210,6 +216,11 @@ public:
         << "Size mismatch: buffer_oob_vec_ and infer_list_ must match in "
            "length.";
 
+    DLOG(INFO) << "[InferLayout] all participating operators:" << '\n';
+    for (int i = 0; i < infer_list_stmt_.size(); ++i) {
+      DLOG(INFO) << "    op " << i << ":" << infer_list_stmt_[i] << '\n';
+    }
+
     // If needed, you can also check that annotated_layout_map_ is not empty, or
     // anything else relevant to your setup.
 
@@ -470,6 +481,13 @@ private:
 
   void InferInFreeMode(LayoutMap &layout_map,
                        const LayoutMap &strict_layout_map) {
+
+    DLOG(INFO) << "Enforced layout maps:" << '\n';
+    for (auto &&[k, v] : layout_map) {
+      DLOG(INFO) << "    " << k << ": " << v->DebugOutput() << '\n';
+    }
+    DLOG(INFO) << '\n';
+
     // Group operators into connected components
     UnionFind<int> uf;
     for (int i = 0; i < infer_list_.size(); i++) {
@@ -505,52 +523,53 @@ private:
     std::vector<bool> in_queue(infer_list_.size(), false);
 
     for (auto &&[root, members] : components) {
+      DLOG(INFO) << "======================= processing component " << root
+                 << '\n';
       decltype(infer_list_) best_infer_list;
       LayoutMap best_layout_map;
       int64_t min_reg_num = INT64_MAX;
+      int min_reg_num_infer_root = -1;
 
+      // Try each member as the root of inference for this component
       for (int attempt_infer_root : members) {
-        // backup infer_list_ in class member
+        DLOG(INFO) << "----------------------- try root " << attempt_infer_root
+                   << '\n';
+        // Backup the current infer_list_ state
         auto back_infer_list = BackupInferList();
-        // create temporarily used layout_map, new handle so that it copies on
-        // write
+        // Copy the current layout_map for temporary use
         LayoutMap tmp_layout_map = layout_map;
-        // infer from attempt_infer_root in free mode
         bool do_update = true;
         try {
+          // Run inference starting from attempt_infer_root
           RunInferStep(attempt_infer_root, InferLevel::kFree, true,
                        tmp_layout_map, strict_layout_map, q, in_queue);
           FinishInferQueue(InferLevel::kFree, tmp_layout_map, strict_layout_map,
                            q, in_queue);
-          // Silly workaround: we have no clue if single root will iterate over
-          // the entire component, since the InferLayout implementations have
-          // complicated conditioning inside and we know nothing about it.
-          // This would constantly result in incomplete layouts for buffers in
-          // this component. Instead of trying all combinations of root
-          // selection order, we simply go through all other loops in order
-          // after the first search from attempt_infer_root.
+
+          // After the first search, run inference for all other members in
+          // order
           for (int other_infer_root : members) {
             if (other_infer_root != attempt_infer_root) {
               RunInferStep(other_infer_root, InferLevel::kFree, true,
                            tmp_layout_map, strict_layout_map, q, in_queue);
-              // must also be kFree here to avoid conflicts.
               FinishInferQueue(InferLevel::kFree, tmp_layout_map,
                                strict_layout_map, q, in_queue);
             }
           }
-        } catch (LayoutConflictException e) {
-          // such an order fails, try others
+        } catch (const LayoutConflictException &e) {
           do_update = false;
-        } catch (NormalizeIterException e) {
-          // such an order encounters iterators that is not normalizable, try
-          // others e.g. i * 576 % 2048
+          DLOG(INFO) << "attempt failed due to LayoutConflictException "
+                     << e.what() << '\n';
+        } catch (const NormalizeIterException &e) {
           do_update = false;
+          DLOG(INFO) << "attempt failed due to NormalizeIterException "
+                     << e.what() << '\n';
         }
 
         if (do_update) {
-          // compute total register number
+          // Compute the total register number for this layout
           int64_t reg_num = 0;
-          for (auto &&[buffer, layout] : tmp_layout_map) {
+          for (const auto &[buffer, layout] : tmp_layout_map) {
             if (auto frag = layout.as<Fragment>()) {
               int64_t frag_reg_num = 1;
               for (auto i : frag.value()->OutputShape()) {
@@ -561,21 +580,24 @@ private:
               reg_num += frag_reg_num;
             }
           }
-          // if it's any better, update the best_* storage
+          // Update the best plan if this one uses fewer registers
           if (reg_num < min_reg_num) {
-            best_infer_list = std::move(infer_list_);
+            best_infer_list =
+                BackupInferList(); // Use backup to avoid moving out infer_list_
             best_layout_map = tmp_layout_map;
             min_reg_num = reg_num;
+            min_reg_num_infer_root = attempt_infer_root;
           }
         }
-        // recover stateful infer_list_, head on next
+        // Restore infer_list_ state for the next attempt
         infer_list_ = std::move(back_infer_list);
       }
-      if (min_reg_num < INT64_MAX) {
-        // now apply the best plan for this component
+      ICHECK(min_reg_num < INT64_MAX) << "no available layout found" << '\n';
+      // Apply the best plan for this component
       infer_list_ = std::move(best_infer_list);
       layout_map = best_layout_map;
-      }
+      DLOG(INFO) << "[InferInFreeMode] Final selection is attempt_infer_root = "
+                 << min_reg_num_infer_root << '\n';
     }
   }
 };
@@ -682,20 +704,25 @@ private:
       // Here, A_local is a register-local buffer held independently by each
       // thread, so explicit thread binding is not required.
       //
-      // We use PostOrderVisit to detect whether the buffer store targets a
-      // "local" buffer, which indicates register usage and justifies skipping
+      // We use PostOrderVisit to detect whether the loop only manuplates
+      // "local" buffers, which indicates register usage and justifies skipping
       // thread binding.
-      bool is_register_store = false;
+      bool local_register_only = true;
       PostOrderVisit(root, [&](const ObjectRef &obj) {
         if (const auto *store = obj.as<BufferStoreNode>()) {
-          if (store->buffer.scope() == "local") {
-            is_register_store = true;
+          if (store->buffer.scope() != "local") {
+            local_register_only = false;
+          }
+        } else if (const auto *load = obj.as<BufferLoadNode>()) {
+          if (load->buffer.scope() != "local") {
+            local_register_only = false;
           }
         }
       });
 
       auto loop_layout = result_.for_map[root];
-      bool parallel_loop = !is_register_store && !skip_thread_partition_;
+      // FIXME: tell in-Parallel and out-of-Parallel `local`s apart
+      bool parallel_loop = !skip_thread_partition_ && !local_register_only;
 
       if (parallel_loop) {
         for_node =

src/transform/layout_reducer.cc

@@ -178,7 +178,8 @@ private:
   Stmt VisitStmt_(const ForNode *op) final {
     // only annotate the outermost loop
     bool should_annotate = false;
-    if (!inside_reducer_range_.empty() && !already_annotated_) {
+    if (!inside_reducer_range_.empty() && !already_annotated_ &&
+        op->kind == ForKind::kParallel) {
       should_annotate = true;
       already_annotated_ = true;
     }

src/transform/merge_shared_memory_allocations.cc

@@ -639,12 +639,12 @@ private:
   };
 
   void PlanAlignment(const Stmt &stmt) {
-    LOG(INFO) << "PlanAlignment";
+    DLOG(INFO) << "PlanAlignment";
     PostOrderVisit(stmt, [&](const ObjectRef &node) {
       if (const auto *call = node.as<CallNode>()) {
         if (call->op.same_as(tl::tl_gemm()) ||
             call->op.same_as(tl::tl_gemm_sp())) {
-          LOG(INFO) << "PostOrderVisit CallNode tl_gemm and tl_gemm_sp: "
+          DLOG(INFO) << "PostOrderVisit CallNode tl_gemm and tl_gemm_sp: "
                      << call->op;
         }
       }

src/transform/storage_rewrite.cc

@@ -1789,7 +1789,7 @@ public:
     PrimExpr last_extent = extents[extents.size() - 1];
     extents.Set(extents.size() - 1,
                 last_extent / make_const(last_extent.dtype(), info.factor()));
-    LOG(INFO) << "Allocate with " << new_buffer_var << " and "
+    DLOG(INFO) << "Allocate with " << new_buffer_var << " and "
                << info.new_element_dtype << " extents: " << extents;
     return Allocate(new_buffer_var, info.new_element_dtype, extents,
                     op->condition, op->body);

tilelang/transform/pass_config.py

@@ -45,6 +45,9 @@ class PassConfigKey(str, Enum):
     TL_DISABLE_SAFE_MEMORY_ACCESS = "tl.disable_safe_memory_legalize"
     """Disable safe memory access optimization. Default: False"""
 
+    TL_DISABLE_WGMMA = "tl.disable_wgmma"
+    """Disable usage of Hopper WGMMA. Default: False"""
+
     TL_DEBUG_MERGE_SHARED_MEMORY_ALLOCATIONS = "tl.debug_merge_shared_memory_allocations"
     """Enable debug information for merge shared memory allocations. Default: False"""