[Mlir-commits] [mlir] [MLIR][XeGPU] Scattered ops sg-to-wi distribution (PR #154949)

[Artem Kroviakov]

================
@@ -811,6 +811,136 @@ struct GpuBarrierDistribution final : public gpu::WarpDistributionPattern {
   }
 };
 
+struct StoreDistribution final : public gpu::WarpDistributionPattern {
+  using gpu::WarpDistributionPattern::WarpDistributionPattern;
+  LogicalResult matchAndRewrite(gpu::WarpExecuteOnLane0Op warpOp,
+                                PatternRewriter &rewriter) const override {
+    auto yield = cast<gpu::YieldOp>(
+        warpOp.getBodyRegion().getBlocks().begin()->getTerminator());
+    Operation *lastNode = yield->getPrevNode();
+    auto storeScatterOp = dyn_cast_or_null<xegpu::StoreScatterOp>(lastNode);
+    if (!storeScatterOp)
+      return failure();
+    else if (!storeScatterOp.getOffsets())
+      return rewriter.notifyMatchFailure(storeScatterOp,
+                                         "Store op must have offsets argument");
+    else if (cast<VectorType>(storeScatterOp.getOffsets().getType())
+                 .getRank() != 1)
+      return rewriter.notifyMatchFailure(storeScatterOp,
+                                         "Expected 1D offsets vector");
+
+    VectorType storeVecTy =
+        cast<VectorType>(storeScatterOp.getValue().getType());
+    assert(storeVecTy.getRank() <= 2 &&
+           "Expected at most 2D result at SG level");
+    VectorType distStoreVecTy;
+    if (storeVecTy.getRank() == 2)
+      distStoreVecTy = VectorType::Builder(storeVecTy).dropDim(0);
+    else // rank 1
+      distStoreVecTy = VectorType::Builder(storeVecTy).setDim(0, 1);
+
+    SmallVector<size_t> newRetIndices;
+    SmallVector<Value> operands =
+        llvm::to_vector_of<Value>(storeScatterOp->getOperands());
+    SmallVector<Type> operandTypes =
----------------
akroviakov wrote:

Renamed to `operandTypesToYield`

https://github.com/llvm/llvm-project/pull/154949