[mhlo] fix ScatterOp::fold for batching dims

PiperOrigin-RevId: 675727628
openxla · Sep 17, 2024 · 891d972 · 891d972
1 parent ee266ed
commit 891d972
Show file tree

Hide file tree

Showing 2 changed files with 31 additions and 4 deletions.
diff --git a/xla/mlir_hlo/mhlo/IR/hlo_ops.cc b/xla/mlir_hlo/mhlo/IR/hlo_ops.cc
@@ -5884,16 +5884,24 @@ LogicalResult ScatterOp::fold(
     // value.
     indexIndex.clear();
     if (indexVectorDim == 0) indexIndex.push_back(0);
+    auto updateWindowDims = getScatterDimensionNumbers().getUpdateWindowDims();
     for (int64_t i = 0; i < static_cast<int64_t>(updateIndex.size()); ++i) {
-      if (llvm::count(getScatterDimensionNumbers().getUpdateWindowDims(), i) ==
-          0)
+      if (!llvm::is_contained(updateWindowDims, i))
         indexIndex.push_back(updateIndex[i]);
       if (static_cast<int64_t>(indexIndex.size()) == indexVectorDim)
         indexIndex.push_back(0);
     }
 
     // Compute the index for the given update value in the base tensor.
     baseIndex.assign(baseType.getRank(), 0);
+    auto inputBatchingDims =
+        getScatterDimensionNumbers().getInputBatchingDims();
+    auto scatterIndicesBatchingDims =
+        getScatterDimensionNumbers().getScatterIndicesBatchingDims();
+    for (auto [operandDim, indicesDim] :
+         llvm::zip_equal(inputBatchingDims, scatterIndicesBatchingDims)) {
+      baseIndex[operandDim] = indexIndex[indicesDim];
+    }
     uint64_t indexCount = indexType.getShape()[indexVectorDim];
     for (uint64_t i = 0; i < indexCount; ++i) {
       uint64_t operandDim =
@@ -5905,9 +5913,10 @@ LogicalResult ScatterOp::fold(
     uint64_t updateWindowDimIndex = 0;
     auto insertedWindowDims =
         getScatterDimensionNumbers().getInsertedWindowDims();
-    auto updateWindowDims = getScatterDimensionNumbers().getUpdateWindowDims();
     for (uint64_t i = 0; i < baseIndex.size(); ++i) {
-      if (llvm::count(insertedWindowDims, i)) continue;
+      if (llvm::is_contained(insertedWindowDims, i) ||
+          llvm::is_contained(inputBatchingDims, i))
+        continue;
       baseIndex[i] += updateIndex[updateWindowDims[updateWindowDimIndex]];
       updateWindowDimIndex++;
     }

diff --git a/xla/mlir_hlo/tests/Dialect/mhlo/canonicalize/scatter.mlir b/xla/mlir_hlo/tests/Dialect/mhlo/canonicalize/scatter.mlir
@@ -53,6 +53,24 @@ func.func @scatter_full_overwrite() ->  tensor<512x1x6400x6400xf32> {
 
 // -----
 
+// Verify that a full overwrite of the "base" with a batched scatter is
+// correctly folded.
+func.func @scatter_batching_dims_full_overwrite() ->  tensor<3x1x6400x6400xf32> {
+  %base = mhlo.constant dense<0.000000e+00> : tensor<3x1x6400x6400xf32>
+  %index = mhlo.constant dense<0> : tensor<3x1xi32>
+  %update = mhlo.constant dense<1.000000e+00> : tensor<3x1x6400x6400xf32>
+  %scatter = "mhlo.scatter"(%base, %index, %update) ({
+    ^bb0(%arg5: tensor<f32>, %arg6: tensor<f32>):
+      "mhlo.return"(%arg6) : (tensor<f32>) -> ()
+  }) {indices_are_sorted = true, scatter_dimension_numbers = #mhlo.scatter<update_window_dims = [1, 2, 3], input_batching_dims = [0], scatter_indices_batching_dims = [0], scatter_dims_to_operand_dims = [3], index_vector_dim = 1>, unique_indices = true} : (tensor<3x1x6400x6400xf32>, tensor<3x1xi32>, tensor<3x1x6400x6400xf32>) -> tensor<3x1x6400x6400xf32>
+
+  // CHECK: %[[FOLD:.*]] = mhlo.constant dense<1.000000e+00> : tensor<3x1x6400x6400xf32>
+  // CHECK: return %[[FOLD]]
+  func.return %scatter :  tensor<3x1x6400x6400xf32>
+}
+
+// -----
+
 // Verify that a full overwrite of the "base" with a scatter is correctly folded
 // even if the base and update are not constant values.
 func.func @scatter_full_overwrite_non_const(