.triton/dump/791dcf81763c6dee467e1d2c436fd6cf/triton_.ttir

module {
  tt.func public @triton__0d1d2d3d4de5(%arg0: !tt.ptr<bf16, 1> {tt.divisibility = 16 : i32}, %arg1: !tt.ptr<f32, 1> {tt.divisibility = 16 : i32}, %arg2: !tt.ptr<f32, 1> {tt.divisibility = 16 : i32}, %arg3: !tt.ptr<bf16, 1> {tt.divisibility = 16 : i32}, %arg4: i64 {tt.divisibility = 16 : i32, tt.max_divisibility = 16 : i32}, %arg5: i64) attributes {noinline = false} {
    %cst = arith.constant dense<0.000000e+00> : tensor<1x2048xbf16>
    %c50257_i64 = arith.constant 50257 : i64
    %cst_0 = arith.constant dense<true> : tensor<1x2048xi1>
    %c50257_i32 = arith.constant 50257 : i32
    %c2048_i32 = arith.constant 2048 : i32
    %c0_i32 = arith.constant 0 : i32
    %cst_1 = arith.constant dense<50257> : tensor<1x2048xi64>
    %cst_2 = arith.constant dense<0.000000e+00> : tensor<1x2048xf32>
    %cst_3 = arith.constant dense<0xFF800000> : tensor<1x2048xf32>
    %0 = tt.get_program_id x : i32
    %1 = arith.extsi %0 : i32 to i64
    %2 = tt.make_range {end = 2048 : i32, start = 0 : i32} : tensor<2048xi32>
    %3 = tt.expand_dims %2 {axis = 0 : i32} : (tensor<2048xi32>) -> tensor<1x2048xi32>
    %4 = arith.extsi %3 : tensor<1x2048xi32> to tensor<1x2048xi64>
    %5 = arith.muli %1, %c50257_i64 : i64
    %6 = tt.splat %5 : (i64) -> tensor<1x2048xi64>
    %7 = tt.splat %arg0 : (!tt.ptr<bf16, 1>) -> tensor<1x2048x!tt.ptr<bf16, 1>>
    %8 = scf.for %arg6 = %c0_i32 to %c50257_i32 step %c2048_i32 iter_args(%arg7 = %cst_3) -> (tensor<1x2048xf32>)  : i32 {
      %29 = arith.extsi %arg6 : i32 to i64
      %30 = tt.splat %29 : (i64) -> tensor<1x2048xi64>
      %31 = arith.addi %30, %4 : tensor<1x2048xi64>
      %32 = arith.cmpi slt, %31, %cst_1 : tensor<1x2048xi64>
      %33 = arith.addi %31, %6 : tensor<1x2048xi64>
      %34 = tt.addptr %7, %33 : tensor<1x2048x!tt.ptr<bf16, 1>>, tensor<1x2048xi64>
      %35 = tt.load %34, %32, %cst {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<1x2048xbf16>
      %36 = arith.extf %35 : tensor<1x2048xbf16> to tensor<1x2048xf32>
      %37 = arith.cmpf ogt, %arg7, %36 : tensor<1x2048xf32>
      %38 = arith.cmpf une, %arg7, %arg7 : tensor<1x2048xf32>
      %39 = arith.ori %37, %38 : tensor<1x2048xi1>
      %40 = arith.xori %39, %cst_0 : tensor<1x2048xi1>
      %41 = arith.andi %32, %40 : tensor<1x2048xi1>
      %42 = arith.select %41, %36, %arg7 : tensor<1x2048xi1>, tensor<1x2048xf32>
      scf.yield %42 : tensor<1x2048xf32>
    }
    %9 = "tt.reduce"(%8) <{axis = 1 : i32}> ({
    ^bb0(%arg6: f32, %arg7: f32):
      %29 = arith.cmpf ogt, %arg6, %arg7 : f32
      %30 = arith.cmpf une, %arg6, %arg6 : f32
      %31 = arith.ori %29, %30 : i1
      %32 = arith.select %31, %arg6, %arg7 : f32
      tt.reduce.return %32 : f32
    }) : (tensor<1x2048xf32>) -> tensor<1xf32>
    %10 = tt.expand_dims %9 {axis = 1 : i32} : (tensor<1xf32>) -> tensor<1x1xf32>
    %11 = tt.addptr %arg1, %1 : !tt.ptr<f32, 1>, i64
    %12 = tt.splat %11 : (!tt.ptr<f32, 1>) -> tensor<1x1x!tt.ptr<f32, 1>>
    tt.store %12, %10 {cache = 1 : i32, evict = 1 : i32} : tensor<1x1xf32>
    %13 = arith.muli %1, %c50257_i64 : i64
    %14 = tt.splat %13 : (i64) -> tensor<1x2048xi64>
    %15 = tt.splat %arg0 : (!tt.ptr<bf16, 1>) -> tensor<1x2048x!tt.ptr<bf16, 1>>
    %16 = tt.broadcast %10 : (tensor<1x1xf32>) -> tensor<1x2048xf32>
    %17 = scf.for %arg6 = %c0_i32 to %c50257_i32 step %c2048_i32 iter_args(%arg7 = %cst_2) -> (tensor<1x2048xf32>)  : i32 {
      %29 = arith.extsi %arg6 : i32 to i64
      %30 = tt.splat %29 : (i64) -> tensor<1x2048xi64>
      %31 = arith.addi %30, %4 : tensor<1x2048xi64>
      %32 = arith.cmpi slt, %31, %cst_1 : tensor<1x2048xi64>
      %33 = arith.addi %31, %14 : tensor<1x2048xi64>
      %34 = tt.addptr %15, %33 : tensor<1x2048x!tt.ptr<bf16, 1>>, tensor<1x2048xi64>
      %35 = tt.load %34, %32, %cst {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<1x2048xbf16>
      %36 = arith.extf %35 : tensor<1x2048xbf16> to tensor<1x2048xf32>
      %37 = arith.subf %36, %16 : tensor<1x2048xf32>
      %38 = math.exp %37 : tensor<1x2048xf32>
      %39 = arith.addf %arg7, %38 : tensor<1x2048xf32>
      %40 = arith.select %32, %39, %arg7 : tensor<1x2048xi1>, tensor<1x2048xf32>
      scf.yield %40 : tensor<1x2048xf32>
    }
    %18 = "tt.reduce"(%17) <{axis = 1 : i32}> ({
    ^bb0(%arg6: f32, %arg7: f32):
      %29 = arith.addf %arg6, %arg7 : f32
      tt.reduce.return %29 : f32
    }) : (tensor<1x2048xf32>) -> tensor<1xf32>
    %19 = tt.expand_dims %18 {axis = 1 : i32} : (tensor<1xf32>) -> tensor<1x1xf32>
    %20 = tt.addptr %arg2, %1 : !tt.ptr<f32, 1>, i64
    %21 = tt.splat %20 : (!tt.ptr<f32, 1>) -> tensor<1x1x!tt.ptr<f32, 1>>
    tt.store %21, %19 {cache = 1 : i32, evict = 1 : i32} : tensor<1x1xf32>
    %22 = arith.muli %1, %c50257_i64 : i64
    %23 = tt.splat %22 : (i64) -> tensor<1x2048xi64>
    %24 = tt.splat %arg0 : (!tt.ptr<bf16, 1>) -> tensor<1x2048x!tt.ptr<bf16, 1>>
    %25 = tt.broadcast %10 : (tensor<1x1xf32>) -> tensor<1x2048xf32>
    %26 = math.log %19 : tensor<1x1xf32>
    %27 = tt.broadcast %26 : (tensor<1x1xf32>) -> tensor<1x2048xf32>
    %28 = tt.splat %arg3 : (!tt.ptr<bf16, 1>) -> tensor<1x2048x!tt.ptr<bf16, 1>>
    scf.for %arg6 = %c0_i32 to %c50257_i32 step %c2048_i32  : i32 {
      %29 = arith.extsi %arg6 : i32 to i64
      %30 = tt.splat %29 : (i64) -> tensor<1x2048xi64>
      %31 = arith.addi %30, %4 : tensor<1x2048xi64>
      %32 = arith.cmpi slt, %31, %cst_1 : tensor<1x2048xi64>
      %33 = arith.addi %31, %23 : tensor<1x2048xi64>
      %34 = tt.addptr %24, %33 : tensor<1x2048x!tt.ptr<bf16, 1>>, tensor<1x2048xi64>
      %35 = tt.load %34, %32, %cst {cache = 1 : i32, evict = 2 : i32, isVolatile = false} : tensor<1x2048xbf16>
      %36 = arith.extf %35 : tensor<1x2048xbf16> to tensor<1x2048xf32>
      %37 = arith.subf %36, %25 : tensor<1x2048xf32>
      %38 = arith.subf %37, %27 : tensor<1x2048xf32>
      %39 = tt.addptr %28, %33 : tensor<1x2048x!tt.ptr<bf16, 1>>, tensor<1x2048xi64>
      %40 = arith.truncf %38 : tensor<1x2048xf32> to tensor<1x2048xbf16>
      tt.store %39, %40, %32 {cache = 1 : i32, evict = 1 : i32} : tensor<1x2048xbf16>
    }
    tt.return
  }
}