Index: mlir/test/Dialect/GPU/multiple-all-reduce.mlir
===================================================================
--- /dev/null
+++ mlir/test/Dialect/GPU/multiple-all-reduce.mlir
@@ -0,0 +1,25 @@
+// RUN: mlir-opt --gpu-kernel-outlining --convert-gpu-to-nvvm %s | FileCheck %s
+
+func @main() {
+  %data = alloc() : memref<2x6xf32>
+  %sum = alloc() : memref<2xf32>
+  %mul = alloc() : memref<2xf32>
+  %c1 = constant 1 : index
+
+  // ADD + MUL
+  gpu.launch blocks(%bx, %by, %bz) in (%grid_x = %c1, %grid_y = %c1, %grid_z = %c1)
+             threads(%tx, %ty, %tz) in (%block_x = %c1, %block_y = %c1, %block_z = %c1) {
+    %val = load %data[%bx, %tx] : memref<2x6xf32>
+    %reduced0 = "gpu.all_reduce"(%val) ({}) { op = "add" } : (f32) -> (f32)
+    store %reduced0, %sum[%bx] : memref<2xf32>
+    %reduced1 = "gpu.all_reduce"(%val) ({}) { op = "mul" } : (f32) -> (f32)
+    store %reduced1, %mul[%bx] : memref<2xf32>
+    gpu.terminator
+  }
+
+// CHECK:      gpu.module @main_kernel {
+// CHECK-NEXT:   llvm.mlir.global internal @reduce_buffer_0() {addr_space = 3 : i32} : !llvm<"[32 x float]">
+// CHECK-NEXT:   llvm.mlir.global internal @reduce_buffer() {addr_space = 3 : i32} : !llvm<"[32 x float]">
+
+  return
+}
Index: mlir/test/mlir-cuda-runner/multiple-all-reduce.mlir
===================================================================
--- /dev/null
+++ mlir/test/mlir-cuda-runner/multiple-all-reduce.mlir
@@ -0,0 +1,64 @@
+// RUN: mlir-cuda-runner %s --shared-libs=%cuda_wrapper_library_dir/libcuda-runtime-wrappers%shlibext,%linalg_test_lib_dir/libmlir_runner_utils%shlibext --entry-point-result=void | FileCheck %s
+
+func @main() {
+  %data = alloc() : memref<2x6xf32>
+  %sum = alloc() : memref<2xf32>
+  %mul = alloc() : memref<2xf32>
+  %cst0 = constant 0.0 : f32
+  %cst1 = constant 1.0 : f32
+  %cst2 = constant 2.0 : f32
+  %cst4 = constant 4.0 : f32
+  %cst8 = constant 8.0 : f32
+  %cst16 = constant 16.0 : f32
+
+  %cst3 = constant 3.0 : f32
+  %cst6 = constant 6.0 : f32
+  %cst7 = constant 7.0 : f32
+  %cst10 = constant 10.0 : f32
+  %cst11 = constant 11.0 : f32
+
+  %c0 = constant 0 : index
+  %c1 = constant 1 : index
+  %c2 = constant 2 : index
+  %c3 = constant 3 : index
+  %c4 = constant 4 : index
+  %c5 = constant 5 : index
+  %c6 = constant 6 : index
+    
+  store %cst0, %data[%c0, %c0] : memref<2x6xf32>
+  store %cst1, %data[%c0, %c1] : memref<2x6xf32>
+  store %cst2, %data[%c0, %c2] : memref<2x6xf32>
+  store %cst4, %data[%c0, %c3] : memref<2x6xf32>
+  store %cst8, %data[%c0, %c4] : memref<2x6xf32>
+  store %cst16, %data[%c0, %c5] : memref<2x6xf32>
+
+  store %cst2, %data[%c1, %c0] : memref<2x6xf32>
+  store %cst3, %data[%c1, %c1] : memref<2x6xf32>
+  store %cst6, %data[%c1, %c2] : memref<2x6xf32>
+  store %cst7, %data[%c1, %c3] : memref<2x6xf32>
+  store %cst10, %data[%c1, %c4] : memref<2x6xf32>
+  store %cst11, %data[%c1, %c5] : memref<2x6xf32>
+
+  // ADD + MUL
+  gpu.launch blocks(%bx, %by, %bz) in (%grid_x = %c2, %grid_y = %c1, %grid_z = %c1)
+             threads(%tx, %ty, %tz) in (%block_x = %c6, %block_y = %c1, %block_z = %c1) {
+    %val = load %data[%bx, %tx] : memref<2x6xf32>
+    %reduced0 = "gpu.all_reduce"(%val) ({}) { op = "add" } : (f32) -> (f32)
+    store %reduced0, %sum[%bx] : memref<2xf32>
+    %reduced1 = "gpu.all_reduce"(%val) ({}) { op = "mul" } : (f32) -> (f32)
+    store %reduced1, %mul[%bx] : memref<2xf32>
+    gpu.terminator
+  }
+
+  %ptr_sum = memref_cast %sum : memref<2xf32> to memref<*xf32>
+  call @print_memref_f32(%ptr_sum) : (memref<*xf32>) -> ()
+  // CHECK: [31, 39]
+
+  %ptr_mul = memref_cast %mul : memref<2xf32> to memref<*xf32>
+  call @print_memref_f32(%ptr_mul) : (memref<*xf32>) -> ()
+  // CHECK: [0, 27720]
+
+  return
+}
+
+func @print_memref_f32(memref<*xf32>)