rebase to the latest llvm

intel · Jun 27, 2024 · 79b277d · 79b277d
1 parent 6f1445c
commit 79b277d
Show file tree

Hide file tree

Showing 4 changed files with 14 additions and 11 deletions.
diff --git a/cmake/llvm-version.txt b/cmake/llvm-version.txt
@@ -1 +1 @@
-763b96c86d81d51d0db430791a61fd1e8a406bce
+891ec2af45c02718c65f539cb6dad1758f079e73
diff --git a/include/gc/Dialect/Arith/Utils/EasyBuild.h b/include/gc/Dialect/Arith/Utils/EasyBuild.h
@@ -361,8 +361,9 @@ inline EBUnsigned extend(Type type, const EBUnsigned &a) {
 }
 
 inline EBFloatPoint extend(Type type, const EBFloatPoint &a) {
+  arith::FastMathFlagsAttr fastMathAttr;
   return OperatorHandlers::create<EBFloatPoint, arith::ExtFOp>(a.builder, type,
-                                                               a);
+                                                               a, fastMathAttr);
 }
 
 inline EBSigned trunc(Type type, const EBSigned &a) {

diff --git a/lib/gc/Transforms/DeepTileContractionNamedOp.cpp b/lib/gc/Transforms/DeepTileContractionNamedOp.cpp
@@ -497,10 +497,12 @@ generateOuterLoop(RewriterBase &b, linalg::LinalgOp linalgOp,
         auto tilingResult = linalgX::tileAllUsingForall(
             b, cast<PartialReductionOpInterface>(currentOp.getOperation()), {},
             tileSizes, newParallelDims, std::nullopt);
-        if (failed(tilingResult))
+        if (failed(tilingResult) &&
+            tilingResult->parallelTiledOps.size() == 1UL)
           return failure();
-        currentOp = dyn_cast<linalg::LinalgOp>(tilingResult->parallelTiledOp);
-        if (tilingResult->mergeOp) {
+        currentOp =
+            dyn_cast<linalg::LinalgOp>(tilingResult->parallelTiledOps.back());
+        if (!tilingResult->mergeOps.empty()) {
           for (const auto &fn : option.finalReduceCallBacks) {
             auto result = fn(b, currentOp.getLoc(), *tilingResult);
             if (succeeded(result)) {
@@ -672,7 +674,7 @@ struct deepTileMatmul : public OpInterfaceRewritePattern<linalg::LinalgOp> {
                                  initValue[0].getDefiningOp())
                                  .getDpsInits()[0]);
         }
-        return dyn_cast<linalg::LinalgOp>(result.parallelTiledOp);
+        return dyn_cast<linalg::LinalgOp>(result.parallelTiledOps.back());
       };
       option.finalReduceCallBacks.push_back(removeReduncantFill);
     }

diff --git a/lib/gc/Transforms/Tiling.cpp b/lib/gc/Transforms/Tiling.cpp
@@ -284,7 +284,7 @@ static void calculateTileOffsetsAndSizes(
   OpBuilder::InsertionGuard g(b);
   b.setInsertionPointToStart(forallOp.getBody(0));
 
-  ValueRange threadIds = forallOp.getInductionVars();
+  SmallVector<Value> threadIds = forallOp.getInductionVars();
   SmallVector<OpFoldResult> nonZeroNumThreads =
       llvm::to_vector(llvm::make_filter_range(numThreads, [](OpFoldResult ofr) {
         return !isConstantIntValue(ofr, 0);
@@ -755,8 +755,8 @@ FailureOr<linalg::ForallReductionTilingResult> tileReductionUsingForall(
   ForallReductionTilingResult results;
   results.initialValues = initTensors;
   results.loops = forallOp;
-  results.parallelTiledOp = tiledOp;
-  results.mergeOp = mergeOp;
+  results.parallelTiledOps = {tiledOp};
+  results.mergeOps = {mergeOp};
   return results;
 }
 
@@ -1069,8 +1069,8 @@ FailureOr<linalg::ForallReductionTilingResult> tileAllUsingForall(
   ForallReductionTilingResult results;
   results.initialValues = initTensors;
   results.loops = forallOp;
-  results.parallelTiledOp = tiledOp;
-  results.mergeOp = mergeOp;
+  results.parallelTiledOps = SmallVector<Operation *>{tiledOp};
+  results.mergeOps = SmallVector<Operation *>{mergeOp};
   return results;
 }