#pragma unroll
  for(nvfuser_index_t i317 = 0; i317 < 2; ++i317) {
    // Allocate global tensor T20
    // Allocate global tensor T21
    // Allocate global tensor T22
    T5_reduction.reduceGroup(
      RefTuple<float>(T5[i317]),
      ConstRefTuple<float>(T19[i317]),
      VolatilePtrTuple<float>(&T20[0]),
      LocalTuple<float>(0),
      [](float &a, float b) { a = a + b; },
      RefTuple<float>(T9[i317]),
      ConstRefTuple<float>(T18[i317]),
      VolatilePtrTuple<float>(&T21[0]),
      LocalTuple<float>(0),
      [](float &a, float b) { a = a + b; },
      &T22[0],
      shared_mem,
      true,
      true);
  }

  {
    constexpr int i317 = 0;
    T5_reduction.reduceGroup(
        RefTuple<float>(T5[i317]),
        ConstRefTuple<float>(T19[i317]),
        VolatilePtrTuple<float>(&T20[0]),
        LocalTuple<float>(0),
        [](float &a, float b) { a = a + b; },
        RefTuple<float>(T9[i317]),
        ConstRefTuple<float>(T18[i317]),
        VolatilePtrTuple<float>(&T21[0]),
        LocalTuple<float>(0),
        [](float &a, float b) { a = a + b; },

        RefTuple<float>(T5[i317 + 1]),
        ConstRefTuple<float>(T19[i317 + 1]),
        VolatilePtrTuple<float>(&T20[3136]),
        LocalTuple<float>(0),
        [](float &a, float b) { a = a + b; },

        RefTuple<float>(T9[i317 + 1]),
        ConstRefTuple<float>(T18[i317 + 1]),
        VolatilePtrTuple<float>(&T21[3136]),
        LocalTuple<float>(0),
        [](float &a, float b) { a = a + b; },

        &T22[0],
        shared_mem,
        true,
        true,
        true,
        true);
  }

Launch Parameters: BlockDim.x = 16, BlockDim.y = 8, BlockDim.z = -1, GridDim.x = 1, GridDim.y = 196, GridDim.z = -1, Smem Size = 512
Original: 46 us
Two-way horizontal grouping: 36 us
Two-way horizontal and two-way cross-iteration grouping: 29 us