{
    (%0) = "data(phi_kernel)" () {dtype:(pd_op.DataType)bool,is_persisable:[false],kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"data",name:"_jst.0.x_v.0",op_name:"pd_op.data",place:(pd_op.Place)Place(gpu:0),shape:(pd_op.IntArray)[],stop_gradient:[true]} : () -> gpu_tensor<10x16xf32>
    (%1) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,is_persisable:[false],kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[16],stop_gradient:[true],value:(Float)1} : () -> gpu_tensor<16xf32>
    (%2) = "shape(phi_kernel)" (%0) {is_persisable:[false],kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"shape",op_name:"pd_op.shape",stop_gradient:[true]} : (gpu_tensor<10x16xf32>) -> cpu_tensor<2xi32>
    (%3) = "full_int_array(phi_kernel)" () {dtype:(pd_op.DataType)int64,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:int64>,kernel_name:"full_int_array",op_name:"pd_op.full_int_array",place:(pd_op.Place)Place(cpu),stop_gradient:[true],value:[(Int64)0]} : () -> cpu_tensor<1xi64>
    (%4) = "full_int_array(phi_kernel)" () {dtype:(pd_op.DataType)int64,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:int64>,kernel_name:"full_int_array",op_name:"pd_op.full_int_array",place:(pd_op.Place)Place(cpu),stop_gradient:[true],value:[(Int64)1]} : () -> cpu_tensor<1xi64>
    (%5) = "slice(phi_kernel)" (%2, %3, %4) {axes:[(Int64)0],decrease_axis:[(Int64)0],infer_flags:[(Int64)1],is_persisable:[false],kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:int32>,kernel_name:"slice",op_name:"pd_op.slice",stop_gradient:[true]} : (cpu_tensor<2xi32>, cpu_tensor<1xi64>, cpu_tensor<1xi64>) -> cpu_tensor<i32>
    (%6) = "mean(phi_kernel)" (%0) {axis:(pd_op.IntArray)[],is_persisable:[false],keepdim:false,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"mean",op_name:"pd_op.mean",stop_gradient:[true]} : (gpu_tensor<10x16xf32>) -> gpu_tensor<f32>
    (%7) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,is_persisable:[false],kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[],stop_gradient:[true],value:(Float)0} : () -> gpu_tensor<f32>
    (%8) = "less_than(phi_kernel)" (%6, %7) {is_persisable:[false],kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"less_than",op_name:"pd_op.less_than",stop_gradient:[true]} : (gpu_tensor<f32>, gpu_tensor<f32>) -> gpu_tensor<b>
    (%9, %10, %11) = pd_op.if (%8) -> gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>, gpu_tensor<16xf32>{
        (%12) = "add(phi_kernel)" (%0, %1) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"add",op_name:"pd_op.add"} : (gpu_tensor<10x16xf32>, gpu_tensor<16xf32>) -> gpu_tensor<10x16xf32>
        (%13) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[16],value:(Float)10} : () -> gpu_tensor<16xf32>
        (%14) = "full_int_array(phi_kernel)" () {dtype:(pd_op.DataType)int64,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:int64>,kernel_name:"full_int_array",op_name:"pd_op.full_int_array",place:(pd_op.Place)Place(cpu),stop_gradient:[true],value:[(Int64)0]} : () -> cpu_tensor<1xi64>
        (%15) = "full_int_array(phi_kernel)" () {dtype:(pd_op.DataType)int64,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:int64>,kernel_name:"full_int_array",op_name:"pd_op.full_int_array",place:(pd_op.Place)Place(cpu),stop_gradient:[true],value:[(Int64)1]} : () -> cpu_tensor<1xi64>
        (%16) = "slice(phi_kernel)" (%12, %14, %15) {axes:[(Int64)0],decrease_axis:[(Int64)0],infer_flags:[(Int64)1],kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"slice",op_name:"pd_op.slice"} : (gpu_tensor<10x16xf32>, cpu_tensor<1xi64>, cpu_tensor<1xi64>) -> gpu_tensor<16xf32>
        (%17) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[],value:(Float)10} : () -> gpu_tensor<f32>
        (%18) = "less_than(phi_kernel)" (%16, %17) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"less_than",op_name:"pd_op.less_than",stop_gradient:[true]} : (gpu_tensor<16xf32>, gpu_tensor<f32>) -> gpu_tensor<16xb>
        (%19, %20) = pd_op.if (%18) -> gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>{
            (%21) = "multiply(phi_kernel)" (%12, %13) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"multiply",op_name:"pd_op.multiply"} : (gpu_tensor<10x16xf32>, gpu_tensor<16xf32>) -> gpu_tensor<10x16xf32>
            (%22) = "relu(phi_kernel)" (%21) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"relu",op_name:"pd_op.relu"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
            (%23) = "mean(phi_kernel)" (%22) {axis:(pd_op.IntArray)[],keepdim:false,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"mean",op_name:"pd_op.mean"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<f32>
            (%24) = "cast(phi_kernel)" (%5) {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:int32>,kernel_name:"cast",op_name:"pd_op.cast"} : (cpu_tensor<i32>) -> cpu_tensor<f32>
            (%25) = "memcpy_h2d(phi_kernel)" (%24) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<f32>) -> gpu_tensor<f32>
            (%26) = "less_than(phi_kernel)" (%23, %25) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"less_than",op_name:"pd_op.less_than",stop_gradient:[true]} : (gpu_tensor<f32>, gpu_tensor<f32>) -> gpu_tensor<b>
            (%27, %28) = pd_op.if (%26) -> gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>{
                (%29) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[10,16],value:(Float)-1} : () -> gpu_tensor<10x16xf32>
                (%30) = "abs(phi_kernel)" (%22) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"abs",op_name:"pd_op.abs"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                (%31) = "assign(phi_kernel)" (%30) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                (%32) = "assign(phi_kernel)" (%29) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                () = "cf.yield" (%31, %32) {} : (gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> 
            } else {
                (%33) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[10,16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<10x16xf32>
                (%34) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[10,16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<10x16xf32>
                (%35) = "memcpy_h2d(phi_kernel)" (%34) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                (%36) = "memcpy_h2d(phi_kernel)" (%33) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                () = "cf.yield" (%35, %36) {} : (gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> 
            }
            (%37) = "logical_not(phi_kernel)" (%26) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:bool>,kernel_name:"logical_not",op_name:"pd_op.logical_not",stop_gradient:[true]} : (gpu_tensor<b>) -> gpu_tensor<b>
            (%38, %39) = pd_op.if (%37) -> gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>{
                (%40) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[10,16],value:(Float)-1} : () -> gpu_tensor<10x16xf32>
                (%41) = "subtract(phi_kernel)" (%22, %40) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"subtract",op_name:"pd_op.subtract"} : (gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                (%42) = "assign(phi_kernel)" (%41) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                (%43) = "assign(phi_kernel)" (%40) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                () = "cf.yield" (%43, %42) {} : (gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> 
            } else {
                (%44) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[10,16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<10x16xf32>
                (%45) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[10,16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<10x16xf32>
                (%46) = "memcpy_h2d(phi_kernel)" (%45) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                (%47) = "memcpy_h2d(phi_kernel)" (%44) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                () = "cf.yield" (%46, %47) {} : (gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> 
            }
            (%48) = "cast(phi_kernel)" (%26) {dtype:(pd_op.DataType)int32,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:bool>,kernel_name:"cast",op_name:"pd_op.cast"} : (gpu_tensor<b>) -> gpu_tensor<i32>
            (%49) = "pd_op.select_input" (%48, %39, %27) {} : (gpu_tensor<i32>, gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
            (%50) = "pd_op.select_input" (%48, %38, %28) {} : (gpu_tensor<i32>, gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
            (%51) = "assign(phi_kernel)" (%49) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
            (%52) = "assign(phi_kernel)" (%50) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
            () = "cf.yield" (%52, %51) {} : (gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> 
        } else {
            (%53) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[10,16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<10x16xf32>
            (%54) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[10,16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<10x16xf32>
            (%55) = "memcpy_h2d(phi_kernel)" (%54) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
            (%56) = "memcpy_h2d(phi_kernel)" (%53) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
            () = "cf.yield" (%55, %56) {} : (gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> 
        }
        (%57) = "logical_not(phi_kernel)" (%18) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:bool>,kernel_name:"logical_not",op_name:"pd_op.logical_not",stop_gradient:[true]} : (gpu_tensor<16xb>) -> gpu_tensor<16xb>
        (%58, %59) = pd_op.if (%57) -> gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>{
            (%60) = "multiply(phi_kernel)" (%12, %13) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"multiply",op_name:"pd_op.multiply"} : (gpu_tensor<10x16xf32>, gpu_tensor<16xf32>) -> gpu_tensor<10x16xf32>
            (%61) = "relu(phi_kernel)" (%60) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"relu",op_name:"pd_op.relu"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
            (%62) = "mean(phi_kernel)" (%61) {axis:(pd_op.IntArray)[],keepdim:false,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"mean",op_name:"pd_op.mean"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<f32>
            (%63) = "cast(phi_kernel)" (%5) {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:int32>,kernel_name:"cast",op_name:"pd_op.cast"} : (cpu_tensor<i32>) -> cpu_tensor<f32>
            (%64) = "memcpy_h2d(phi_kernel)" (%63) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<f32>) -> gpu_tensor<f32>
            (%65) = "less_than(phi_kernel)" (%62, %64) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"less_than",op_name:"pd_op.less_than",stop_gradient:[true]} : (gpu_tensor<f32>, gpu_tensor<f32>) -> gpu_tensor<b>
            (%66, %67) = pd_op.if (%65) -> gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>{
                (%68) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[10,16],value:(Float)-1} : () -> gpu_tensor<10x16xf32>
                (%69) = "abs(phi_kernel)" (%61) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"abs",op_name:"pd_op.abs"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                (%70) = "assign(phi_kernel)" (%69) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                (%71) = "assign(phi_kernel)" (%68) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                () = "cf.yield" (%71, %70) {} : (gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> 
            } else {
                (%72) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[10,16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<10x16xf32>
                (%73) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[10,16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<10x16xf32>
                (%74) = "memcpy_h2d(phi_kernel)" (%73) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                (%75) = "memcpy_h2d(phi_kernel)" (%72) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                () = "cf.yield" (%74, %75) {} : (gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> 
            }
            (%76) = "logical_not(phi_kernel)" (%65) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:bool>,kernel_name:"logical_not",op_name:"pd_op.logical_not",stop_gradient:[true]} : (gpu_tensor<b>) -> gpu_tensor<b>
            (%77, %78) = pd_op.if (%76) -> gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>{
                (%79) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[10,16],value:(Float)-1} : () -> gpu_tensor<10x16xf32>
                (%80) = "subtract(phi_kernel)" (%61, %79) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"subtract",op_name:"pd_op.subtract"} : (gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                (%81) = "assign(phi_kernel)" (%80) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                (%82) = "assign(phi_kernel)" (%79) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                () = "cf.yield" (%81, %82) {} : (gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> 
            } else {
                (%83) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[10,16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<10x16xf32>
                (%84) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[10,16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<10x16xf32>
                (%85) = "memcpy_h2d(phi_kernel)" (%84) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                (%86) = "memcpy_h2d(phi_kernel)" (%83) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
                () = "cf.yield" (%85, %86) {} : (gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> 
            }
            (%87) = "cast(phi_kernel)" (%65) {dtype:(pd_op.DataType)int32,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:bool>,kernel_name:"cast",op_name:"pd_op.cast"} : (gpu_tensor<b>) -> gpu_tensor<i32>
            (%88) = "pd_op.select_input" (%87, %77, %67) {} : (gpu_tensor<i32>, gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
            (%89) = "pd_op.select_input" (%87, %78, %66) {} : (gpu_tensor<i32>, gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
            (%90) = "assign(phi_kernel)" (%88) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
            (%91) = "assign(phi_kernel)" (%89) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
            () = "cf.yield" (%91, %90) {} : (gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> 
        } else {
            (%92) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[10,16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<10x16xf32>
            (%93) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[10,16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<10x16xf32>
            (%94) = "memcpy_h2d(phi_kernel)" (%93) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
            (%95) = "memcpy_h2d(phi_kernel)" (%92) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
            () = "cf.yield" (%94, %95) {} : (gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> 
        }
        (%96) = "cast(phi_kernel)" (%18) {dtype:(pd_op.DataType)int32,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:bool>,kernel_name:"cast",op_name:"pd_op.cast"} : (gpu_tensor<16xb>) -> gpu_tensor<16xi32>
        (%97) = "pd_op.select_input" (%96, %59, %20) {} : (gpu_tensor<16xi32>, gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
        (%98) = "pd_op.select_input" (%96, %58, %19) {} : (gpu_tensor<16xi32>, gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
        (%99) = "assign(phi_kernel)" (%13) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<16xf32>) -> gpu_tensor<16xf32>
        (%100) = "assign(phi_kernel)" (%97) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
        (%101) = "assign(phi_kernel)" (%98) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
        () = "cf.yield" (%100, %101, %99) {} : (gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>, gpu_tensor<16xf32>) -> 
    } else {
        (%102) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<16xf32>
        (%103) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[10,16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<10x16xf32>
        (%104) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[10,16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<10x16xf32>
        (%105) = "memcpy_h2d(phi_kernel)" (%104) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
        (%106) = "memcpy_h2d(phi_kernel)" (%103) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
        (%107) = "memcpy_h2d(phi_kernel)" (%102) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<16xf32>) -> gpu_tensor<16xf32>
        () = "cf.yield" (%105, %106, %107) {} : (gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>, gpu_tensor<16xf32>) -> 
    }
    (%108) = "logical_not(phi_kernel)" (%8) {is_persisable:[false],kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:bool>,kernel_name:"logical_not",op_name:"pd_op.logical_not",stop_gradient:[true]} : (gpu_tensor<b>) -> gpu_tensor<b>
    (%109, %110, %111) = pd_op.if (%108) -> gpu_tensor<16xf32>, gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>{
        (%112) = "subtract(phi_kernel)" (%0, %1) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"subtract",op_name:"pd_op.subtract"} : (gpu_tensor<10x16xf32>, gpu_tensor<16xf32>) -> gpu_tensor<10x16xf32>
        (%113) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[16],value:(Float)10} : () -> gpu_tensor<16xf32>
        (%114) = "multiply(phi_kernel)" (%112, %113) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"multiply",op_name:"pd_op.multiply"} : (gpu_tensor<10x16xf32>, gpu_tensor<16xf32>) -> gpu_tensor<10x16xf32>
        (%115) = "relu(phi_kernel)" (%114) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"relu",op_name:"pd_op.relu"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
        (%116) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[10,16],value:(Float)-1} : () -> gpu_tensor<10x16xf32>
        (%117) = "abs(phi_kernel)" (%115) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"abs",op_name:"pd_op.abs"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
        (%118) = "assign(phi_kernel)" (%113) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<16xf32>) -> gpu_tensor<16xf32>
        (%119) = "assign(phi_kernel)" (%117) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
        (%120) = "assign(phi_kernel)" (%116) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"assign",op_name:"pd_op.assign"} : (gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
        () = "cf.yield" (%118, %119, %120) {} : (gpu_tensor<16xf32>, gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> 
    } else {
        (%121) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[10,16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<10x16xf32>
        (%122) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[10,16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<10x16xf32>
        (%123) = "full(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"full",op_name:"pd_op.full",place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[16],stop_gradient:[true],value:(Float)0} : () -> cpu_tensor<16xf32>
        (%124) = "memcpy_h2d(phi_kernel)" (%123) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<16xf32>) -> gpu_tensor<16xf32>
        (%125) = "memcpy_h2d(phi_kernel)" (%122) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
        (%126) = "memcpy_h2d(phi_kernel)" (%121) {dst_place_type:(Int32)1,kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"memcpy_h2d",op_name:"pd_op.memcpy_h2d"} : (cpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
        () = "cf.yield" (%124, %125, %126) {} : (gpu_tensor<16xf32>, gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> 
    }
    (%127) = "cast(phi_kernel)" (%8) {dtype:(pd_op.DataType)int32,is_persisable:[false],kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:bool>,kernel_name:"cast",op_name:"pd_op.cast",stop_gradient:[true]} : (gpu_tensor<b>) -> gpu_tensor<i32>
    (%128) = "pd_op.select_input" (%127, %109, %11) {is_persisable:[false],stop_gradient:[true]} : (gpu_tensor<i32>, gpu_tensor<16xf32>, gpu_tensor<16xf32>) -> gpu_tensor<16xf32>
    (%129) = "pd_op.select_input" (%127, %110, %9) {is_persisable:[false],stop_gradient:[true]} : (gpu_tensor<i32>, gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
    (%130) = "pd_op.select_input" (%127, %111, %10) {is_persisable:[false],stop_gradient:[true]} : (gpu_tensor<i32>, gpu_tensor<10x16xf32>, gpu_tensor<10x16xf32>) -> gpu_tensor<10x16xf32>
    () = "builtin.shadow_output" (%129) {output_name:"_generated_var_33"} : (gpu_tensor<10x16xf32>) -> 
}