[Mlir-commits] [mlir] [mlir][PartialReductionTilingInterface] Generalize implementation of `tileUsingSCF` for `ReductionTilingStrategy::PartialOuterReduction`. (PR #143467)

[Kunwar Grover]

================
@@ -165,28 +166,46 @@ static void checkSafeToTileToForall(TilingInterface op,
   assert((numThreads.empty() || (numThreads.size() == iterators.size())) &&
          "when specified, expected number of threads to use for each loop");
 
+  bool isParallelTiling = false, isReductionTiling = false;
   for (auto [index, iterator, tileSize] :
        llvm::enumerate(iterators, tileSizes)) {
-    // If num threads is specified, check that it is greater than one only for
-    // parallel dimensions.
-    if (!numThreads.empty()) {
-      if (std::optional<int64_t> constNumThreads =
-              getConstantIntValue(numThreads[index])) {
-        if (constNumThreads.value() > 1 &&
+    if (!isConstantIntValue(tileSize, 0)) {
+      isParallelTiling |= iterator == utils::IteratorType::parallel;
+      isReductionTiling |= iterator == utils::IteratorType::reduction;
+    }
+
+    if (loopType == scf::SCFTilingOptions::LoopType::ForallOp &&
+        reductionStrategy == ReductionTilingStrategy::FullReduction) {
+      // If num threads is specified, check that it is greater than one only for
+      // parallel dimensions.
+      if (!numThreads.empty()) {
+        if (std::optional<int64_t> constNumThreads =
+                getConstantIntValue(numThreads[index])) {
+          if (constNumThreads.value() > 1 &&
+              iterator != utils::IteratorType::parallel) {
+            op.emitWarning() << "tiling is not thread safe at axis #" << index;
----------------
Groverkss wrote:

scf.reduce is only needed in a parallel loop context, so it we don't emit it in FullReduction, so it does fit there (FullReduction does not have a parallel loop context).

https://github.com/llvm/llvm-project/pull/143467