Fix unittest of if and while with dynamic shape

[zyfncg]

PR types

Others

PR changes

Others

Description

Pcard-67164

修复动态shape下 if 和 while 子图单测。

主要工作：

1. 前端Pass执行流程重构，支持 If 和 While 等复杂嵌套Op的处理
2. 前端支持将元素计算个数较小的Op fallback 到 PHI 体系执行

测试脚本：

FLAGS_pd_unittest_use_cinn=1 ctest -R test_if_dy -V

FLAGS_pd_unittest_use_cinn=1 ctest -R test_while_dy -V

if 子图：
{
    (%0) = "pd_op.data" () {dtype:(pd_op.DataType)float32,name:"x",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[-1,-1],stop_gradient:[false],sym_shape_str:"shape[S0, S1], data[NULL]"} : () -> pd_op.tensor<-1x-1xf32>
    (%1) = "pd_op.shape" (%0) {stop_gradient:[true],sym_shape_str:"shape[2], data[S0, S1]"} : (pd_op.tensor<-1x-1xf32>) -> pd_op.tensor<2xi32>
    (%2) = cinn_op.fusion () -> pd_op.tensor<i32> {
    (%3) = "cinn_op.slice" (%1) {axes:[(Int64)0],decrease_axis:[(Int64)0],ends:[(Int64)2],infer_flags:[(Int64)1],starts:[(Int64)1],stop_gradient:[true],sym_shape_str:"shape[1], data[S1]"} : (pd_op.tensor<2xi32>) -> pd_op.tensor<i32>
    () = "cf.yield" (%3) {} : (pd_op.tensor<i32>) ->  
 }
    (%4) = "cinn_op.broadcast" (%2) {broadcast_axes:[],out_shape:[(Int64)1],stop_gradient:[true],sym_shape_str:"shape[1], data[NULL]"} : (pd_op.tensor<i32>) -> pd_op.tensor<1xi32>
    (%5) = cinn_op.fusion () -> pd_op.tensor<b> {
    (%6) = "pd_op.full" () {dtype:(pd_op.DataType)int32,place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[1],stop_gradient:[true],sym_shape_str:"shape[1], data[1]",value:(Float)1} : () -> pd_op.tensor<1xi32>
    (%7) = "cinn_op.broadcast" (%6) {broadcast_axes:[(Int64)0],out_shape:[(Int64)-1],stop_gradient:[true]} : (pd_op.tensor<1xi32>) -> pd_op.tensor<-1xi32>
    (%8) = "pd_op.greater_than" (%4, %7) {stop_gradient:[true],sym_shape_str:"shape[1], data[NULL]"} : (pd_op.tensor<1xi32>, pd_op.tensor<-1xi32>) -> pd_op.tensor<b>
    () = "cf.yield" (%8) {} : (pd_op.tensor<b>) ->  
 }
    (%9) = pd_op.if (%5) {stop_gradient:[false],sym_shape_str:"shape[S0, S1], data[NULL]"} -> pd_op.tensor<-1x-1xf32>{
        (%10) = "pd_op.exp" (%0) {stop_gradient:[false],sym_shape_str:"shape[S0, S1], data[NULL]"} : (pd_op.tensor<-1x-1xf32>) -> pd_op.tensor<-1x-1xf32>
        (%11) = "pd_op.subtract" (%10, %0) {stop_gradient:[false],sym_shape_str:"shape[S0, S1], data[NULL]"} : (pd_op.tensor<-1x-1xf32>, pd_op.tensor<-1x-1xf32>) -> pd_op.tensor<-1x-1xf32>
        () = "cf.yield" (%11) {} : (pd_op.tensor<-1x-1xf32>) -> 
    } else {
        () = "cf.yield" (%0) {} : (pd_op.tensor<-1x-1xf32>) -> 
    }
    (%12) = cinn_op.fusion () -> pd_op.tensor<-1x-1xf32> {
    (%13) = "pd_op.relu" (%9) {stop_gradient:[false],sym_shape_str:"shape[S0, S1], data[NULL]"} : (pd_op.tensor<-1x-1xf32>) -> pd_op.tensor<-1x-1xf32>
    () = "cf.yield" (%13) {} : (pd_op.tensor<-1x-1xf32>) ->  
 }
    () = "builtin.shadow_output" (%12) {output_name:"output_0",sym_shape_str:"shape[S0, S1], data[NULL]"} : (pd_op.tensor<-1x-1xf32>) -> 
}

while 子图：
{
    (%0) = "pd_op.data" () {dtype:(pd_op.DataType)float32,name:"x",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[1,-1,768],stop_gradient:[false],sym_shape_str:"shape[1, S0, 768], data[NULL]"} : () -> pd_op.tensor<1x-1x768xf32>
    (%1) = cinn_op.fusion () -> pd_op.tensor<f32> {
    (%2) = "cinn_op.reduce_sum" (%0) {dim:[],keep_dim:false,stop_gradient:[false],sym_shape_str:"shape[], data[NULL]"} : (pd_op.tensor<1x-1x768xf32>) -> pd_op.tensor<f32>
    () = "cf.yield" (%2) {} : (pd_op.tensor<f32>) ->  
 }
    (%3) = "cinn_op.broadcast" (%1) {broadcast_axes:[],out_shape:[(Int64)1],stop_gradient:[false],sym_shape_str:"shape[1], data[NULL]"} : (pd_op.tensor<f32>) -> pd_op.tensor<1xf32>
    (%4, %5) = cinn_op.fusion () -> pd_op.tensor<1xf32>, pd_op.tensor<1xb> {
    (%6) = "pd_op.full" () {dtype:(pd_op.DataType)int32,place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[1],stop_gradient:[true],sym_shape_str:"shape[1], data[1]",value:(Float)1} : () -> pd_op.tensor<1xi32>
    (%7) = "pd_op.cast" (%6) {dtype:(pd_op.DataType)float32,stop_gradient:[true],sym_shape_str:"shape[1], data[1]"} : (pd_op.tensor<1xi32>) -> pd_op.tensor<1xf32>
    (%8) = "pd_op.full" () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[1],stop_gradient:[true],sym_shape_str:"shape[1], data[0]",value:(Float)0} : () -> pd_op.tensor<1xf32>
    (%9) = "cinn_op.broadcast" (%8) {broadcast_axes:[(Int64)0],out_shape:[(Int64)-1],stop_gradient:[true]} : (pd_op.tensor<1xf32>) -> pd_op.tensor<-1xf32>
    (%10) = "pd_op.full" () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[1],stop_gradient:[true],sym_shape_str:"shape[1], data[0]",value:(Float)0} : () -> pd_op.tensor<1xf32>
    (%11) = "cinn_op.broadcast" (%7) {broadcast_axes:[(Int64)0],out_shape:[(Int64)-1],stop_gradient:[true]} : (pd_op.tensor<1xf32>) -> pd_op.tensor<-1xf32>
    (%12) = "cinn_op.broadcast" (%10) {broadcast_axes:[(Int64)0],out_shape:[(Int64)-1],stop_gradient:[true]} : (pd_op.tensor<1xf32>) -> pd_op.tensor<-1xf32>
    (%13) = "pd_op.greater_than" (%3, %9) {stop_gradient:[false],sym_shape_str:"shape[1], data[NULL]"} : (pd_op.tensor<1xf32>, pd_op.tensor<-1xf32>) -> pd_op.tensor<1xb>
    (%14) = "pd_op.less_than" (%12, %11) {stop_gradient:[true],sym_shape_str:"shape[1], data[NULL]"} : (pd_op.tensor<-1xf32>, pd_op.tensor<-1xf32>) -> pd_op.tensor<1xb>
    (%15) = "pd_op.logical_and" (%13, %14) {stop_gradient:[true],sym_shape_str:"shape[1], data[NULL]"} : (pd_op.tensor<1xb>, pd_op.tensor<1xb>) -> pd_op.tensor<1xb>
    () = "cf.yield" (%10, %15) {} : (pd_op.tensor<1xf32>, pd_op.tensor<1xb>) ->  
 }
    (%16, %17) = "pd_op.while"(cond=%5, inputs=%4, %0) { 
    ^%arg0, %arg1
        (%18) = "pd_op.exp" (%arg1) {stop_gradient:[false],sym_shape_str:"shape[], data[NULL]"} : (pd_op.tensor<1x-1x768xf32>) -> pd_op.tensor<1x-1x768xf32>
        (%19) = "pd_op.subtract" (%18, %arg1) {stop_gradient:[false],sym_shape_str:"shape[], data[NULL]"} : (pd_op.tensor<1x-1x768xf32>, pd_op.tensor<1x-1x768xf32>) -> pd_op.tensor<1x-1x768xf32>
        (%20) = "pd_op.full" () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[1],stop_gradient:[true],sym_shape_str:"shape[1], data[1]",value:(Float)1} : () -> pd_op.tensor<1xf32>
        (%21) = "pd_op.scale" (%arg0, %20) {bias:(Float)1,bias_after_scale:true,stop_gradient:[true],sym_shape_str:"shape[], data[NULL]"} : (pd_op.tensor<1xf32>, pd_op.tensor<1xf32>) -> pd_op.tensor<1xf32>
        (%22) = "pd_op.full_int_array" () {dtype:(pd_op.DataType)int64,place:(pd_op.Place)Place(cpu),stop_gradient:[true],sym_shape_str:"shape[0], data[]",value:[]} : () -> pd_op.tensor<0xi64>
        (%23) = "pd_op.sum" (%19, %22) {dtype:(pd_op.DataType)Undefined,keepdim:false,stop_gradient:[false],sym_shape_str:"shape[], data[NULL]"} : (pd_op.tensor<1x-1x768xf32>, pd_op.tensor<0xi64>) -> pd_op.tensor<f32>
        (%24) = "pd_op.full" () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[1],stop_gradient:[true],sym_shape_str:"shape[1], data[0]",value:(Float)0} : () -> pd_op.tensor<1xf32>
        (%25) = "pd_op.greater_than" (%23, %24) {stop_gradient:[false],sym_shape_str:"shape[1], data[NULL]"} : (pd_op.tensor<f32>, pd_op.tensor<1xf32>) -> pd_op.tensor<1xb>
        (%26) = "pd_op.full" () {dtype:(pd_op.DataType)int32,place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[1],stop_gradient:[true],sym_shape_str:"shape[1], data[1]",value:(Float)1} : () -> pd_op.tensor<1xi32>
        (%27) = "pd_op.cast" (%26) {dtype:(pd_op.DataType)float32,stop_gradient:[true],sym_shape_str:"shape[1], data[1]"} : (pd_op.tensor<1xi32>) -> pd_op.tensor<1xf32>
        (%28) = "pd_op.less_than" (%21, %27) {stop_gradient:[true],sym_shape_str:"shape[1], data[NULL]"} : (pd_op.tensor<1xf32>, pd_op.tensor<1xf32>) -> pd_op.tensor<1xb>
        (%29) = "pd_op.logical_and" (%25, %28) {stop_gradient:[true],sym_shape_str:"shape[1], data[NULL]"} : (pd_op.tensor<1xb>, pd_op.tensor<1xb>) -> pd_op.tensor<1xb>
        () = "cf.yield" (%29, %21, %19) {} : (pd_op.tensor<1xb>, pd_op.tensor<1xf32>, pd_op.tensor<1x-1x768xf32>) -> 
    }
    () = "builtin.shadow_output" (%17) {output_name:"output_0",sym_shape_str:"shape[], data[NULL]"} : (pd_op.tensor<1x-1x768xf32>) -> 
}

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[tc20042008]

重构成

const auto XXX = [&](int i) {
  ...
};
for (...) {
  if (XXX(i)) continue;
}

[zyfncg]

好的, 在后续PR里修改

[XiaoguangHu01]

LGTM