module {
  tt.func public @triton__0d1d2d3de4e(%arg0: !tt.ptr<f32, 1> {tt.divisibility = 16 : i32}, %arg1: !tt.ptr<i64, 1> {tt.divisibility = 16 : i32}, %arg2: !tt.ptr<f32, 1> {tt.divisibility = 16 : i32}, %arg3: i32 {tt.divisibility = 16 : i32, tt.max_divisibility = 16 : i32}, %arg4: i32 {tt.max_divisibility = 8 : i32}) attributes {noinline = false} {
    %cst = arith.constant dense<256> : tensor<64x1xi64>
    %cst_0 = arith.constant dense<0> : tensor<64x1xi64>
    %cst_1 = arith.constant dense<512> : tensor<64x1xi64>
    %c4_i32 = arith.constant 4 : i32
    %c120_i32 = arith.constant 120 : i32
    %c0_i32 = arith.constant 0 : i32
    %cst_2 = arith.constant dense<true> : tensor<64x1xi1>
    %cst_3 = arith.constant dense<256> : tensor<64x1xi32>
    %cst_4 = arith.constant dense<131072> : tensor<1x4xi32>
    %cst_5 = arith.constant dense<120> : tensor<1x4xi32>
    %cst_6 = arith.constant dense<0.000000e+00> : tensor<64x4xf32>
    %c64_i32 = arith.constant 64 : i32
    %0 = tt.get_program_id x : i32
    %1 = arith.muli %0, %c64_i32 : i32
    %2 = tt.make_range {end = 64 : i32, start = 0 : i32} : tensor<64xi32>
    %3 = tt.expand_dims %2 {axis = 1 : i32} : (tensor<64xi32>) -> tensor<64x1xi32>
    %4 = tt.splat %1 : (i32) -> tensor<64x1xi32>
    %5 = arith.addi %4, %3 : tensor<64x1xi32>
    %6 = tt.make_range {end = 4 : i32, start = 0 : i32} : tensor<4xi32>
    %7 = tt.expand_dims %6 {axis = 0 : i32} : (tensor<4xi32>) -> tensor<1x4xi32>
    %8 = tt.broadcast %5 : (tensor<64x1xi32>) -> tensor<64x4xi32>
    %9 = tt.splat %arg0 : (!tt.ptr<f32, 1>) -> tensor<64x4x!tt.ptr<f32, 1>>
    %10 = scf.for %arg5 = %c0_i32 to %c120_i32 step %c4_i32 iter_args(%arg6 = %cst_6) -> (tensor<64x4xf32>)  : i32 {
      %27 = tt.splat %arg5 : (i32) -> tensor<1x4xi32>
      %28 = arith.addi %27, %7 : tensor<1x4xi32>
      %29 = arith.cmpi slt, %28, %cst_5 : tensor<1x4xi32>
      %30 = arith.muli %28, %cst_4 : tensor<1x4xi32>
      %31 = tt.broadcast %30 : (tensor<1x4xi32>) -> tensor<64x4xi32>
      %32 = arith.addi %8, %31 : tensor<64x4xi32>
      %33 = tt.addptr %9, %32 : tensor<64x4x!tt.ptr<f32, 1>>, tensor<64x4xi32>
      %34 = tt.broadcast %29 : (tensor<1x4xi1>) -> tensor<64x4xi1>
      %35 = tt.load %33, %34, %cst_6 {cache = 1 : i32, evict = 2 : i32, isVolatile = false} : tensor<64x4xf32>
      %36 = arith.addf %arg6, %35 : tensor<64x4xf32>
      %37 = arith.select %34, %36, %arg6 : tensor<64x4xi1>, tensor<64x4xf32>
      scf.yield %37 : tensor<64x4xf32>
    }
    %11 = "tt.reduce"(%10) <{axis = 1 : i32}> ({
    ^bb0(%arg5: f32, %arg6: f32):
      %27 = arith.addf %arg5, %arg6 : f32
      tt.reduce.return %27 : f32
    }) : (tensor<64x4xf32>) -> tensor<64xf32>
    %12 = tt.expand_dims %11 {axis = 1 : i32} : (tensor<64xf32>) -> tensor<64x1xf32>
    %13 = arith.divsi %5, %cst_3 : tensor<64x1xi32>
    %14 = arith.remsi %5, %cst_3 : tensor<64x1xi32>
    %15 = tt.splat %arg1 : (!tt.ptr<i64, 1>) -> tensor<64x1x!tt.ptr<i64, 1>>
    %16 = tt.addptr %15, %13 : tensor<64x1x!tt.ptr<i64, 1>>, tensor<64x1xi32>
    %17 = tt.load %16 {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<64x1xi64>
    %18 = arith.addi %17, %cst_1 : tensor<64x1xi64>
    %19 = arith.cmpi slt, %17, %cst_0 : tensor<64x1xi64>
    %20 = arith.select %19, %18, %17 : tensor<64x1xi1>, tensor<64x1xi64>
    %21 = arith.muli %20, %cst : tensor<64x1xi64>
    %22 = arith.extsi %14 : tensor<64x1xi32> to tensor<64x1xi64>
    %23 = arith.addi %22, %21 : tensor<64x1xi64>
    %24 = tt.splat %arg2 : (!tt.ptr<f32, 1>) -> tensor<64x1x!tt.ptr<f32, 1>>
    %25 = tt.addptr %24, %23 : tensor<64x1x!tt.ptr<f32, 1>>, tensor<64x1xi64>
    %26 = "tt.atomic_rmw"(%25, %12, %cst_2) <{atomic_rmw_op = 5 : i32, scope = 1 : i32, sem = 4 : i32}> : (tensor<64x1x!tt.ptr<f32, 1>>, tensor<64x1xf32>, tensor<64x1xi1>) -> tensor<64x1xf32>
    tt.return
  }
}