Refine print log for symbolic shape

[zyfncg]

PR types

Others

PR changes

Others

Description

Pcard-73448

引入PrintHook优化Program中符号维度的信息打印。打印效果如下：

优化前：

(%0) = "pd_op.data" () {dtype:(pd_op.DataType)float32,name:"x",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[-1,128],stop_gradient:[false],sym_shape_str:"shape[S0, 128], data[NULL]",symbolic_shape:(shape_data)[S0,128]_[nullopt]} : () -> pd_op.tensor<-1x128xf32>
(%1) = "pd_op.data" () {dtype:(pd_op.DataType)float32,name:"y",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[-1,128],stop_gradient:[false],sym_shape_str:"shape[S1, 128], data[NULL]",symbolic_shape:(shape_data)[S1,128]_[nullopt]} : () -> pd_op.tensor<-1x128xf32>
(%2) = "pd_op.exp" (%0) {stop_gradient:[false],sym_shape_str:"shape[S0, 128], data[NULL]",symbolic_shape:(shape_data)[S0,128]_[nullopt]} : (pd_op.tensor<-1x128xf32>) -> pd_op.tensor<-1x128xf32>
(%3) = "pd_op.shape" (%2) {stop_gradient:[false],sym_shape_str:"shape[2], data[S0, 128]",symbolic_shape:(shape_data)[2]_[S0,128]} : (pd_op.tensor<-1x128xf32>) -> pd_op.tensor<2xi32>
(%4) = "pd_op.shape" (%1) {stop_gradient:[false],sym_shape_str:"shape[2], data[S1, 128]",symbolic_shape:(shape_data)[2]_[S1,128]} : (pd_op.tensor<-1x128xf32>) -> pd_op.tensor<2xi32>
(%5) = "pd_op.shape_broadcast" (%3, %4) {stop_gradient:[false],sym_shape_str:"shape[2], data[Broadcast(S1, S0), 128]"} : (pd_op.tensor<2xi32>, pd_op.tensor<2xi32>) -> pd_op.tensor<2xi32>
(%6) = "pd_op.expand" (%1, %5) {stop_gradient:[false],sym_shape_str:"shape[Broadcast(S1, S0), 128], data[NULL]"} : (pd_op.tensor<-1x128xf32>, pd_op.tensor<2xi32>) -> pd_op.tensor<-1x-1xf32>
(%7) = "pd_op.shape" (%2) {stop_gradient:[false],sym_shape_str:"shape[2], data[S0, 128]",symbolic_shape:(shape_data)[2]_[S0,128]} : (pd_op.tensor<-1x128xf32>) -> pd_op.tensor<2xi32>
(%8) = "pd_op.shape" (%6) {stop_gradient:[false],sym_shape_str:"shape[2], data[Broadcast(S1, S0), 128]",symbolic_shape:(shape_data)[2]_[Broadcast(S1, S0),128]} : (pd_op.tensor<-1x-1xf32>) -> pd_op.tensor<2xi32>
(%9) = "pd_op.shape_broadcast" (%7, %8) {stop_gradient:[false],sym_shape_str:"shape[2], data[Broadcast(Broadcast(S1, S0), S0), 128]"} : (pd_op.tensor<2xi32>, pd_op.tensor<2xi32>) -> pd_op.tensor<2xi32>
(%10) = "pd_op.subtract" (%2, %6) {stop_gradient:[false],sym_shape_str:"shape[S0, 128], data[NULL]",symbolic_shape:(shape_data)[S0,128]_[nullopt]} : (pd_op.tensor<-1x128xf32>, pd_op.tensor<-1x-1xf32>) -> pd_op.tensor<-1x128xf32>
() = "builtin.shadow_output" (%10) {output_name:"output_0",sym_shape_str:"shape[S0, 128], data[NULL]"} : (pd_op.tensor<-1x128xf32>) -> 

优化后：(移除了symbolic_shape在op属性上的打印信息，并将该信息放置在 output value 对应位置)

(%0) = "pd_op.data" () {dtype:(pd_op.DataType)float32,name:"x",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[-1,128],stop_gradient:[false],sym_shape_str:"shape[S0, 128], data[NULL]",symbolic_shape:} : () -> pd_op.tensor<-1x128xf32> { (shape[S0, 128], data[NULL]) }
(%1) = "pd_op.data" () {dtype:(pd_op.DataType)float32,name:"y",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[-1,128],stop_gradient:[false],sym_shape_str:"shape[S1, 128], data[NULL]",symbolic_shape:} : () -> pd_op.tensor<-1x128xf32> { (shape[S1, 128], data[NULL]) }
(%2) = "pd_op.exp" (%0) {stop_gradient:[false],sym_shape_str:"shape[S0, 128], data[NULL]",symbolic_shape:} : (pd_op.tensor<-1x128xf32>) -> pd_op.tensor<-1x128xf32> { (shape[S0, 128], data[NULL]) }
(%3) = "pd_op.shape" (%2) {stop_gradient:[false],sym_shape_str:"shape[2], data[S0, 128]",symbolic_shape:} : (pd_op.tensor<-1x128xf32>) -> pd_op.tensor<2xi32> { (shape[2], data[S0, 128]) }
(%4) = "pd_op.shape" (%1) {stop_gradient:[false],sym_shape_str:"shape[2], data[S1, 128]",symbolic_shape:} : (pd_op.tensor<-1x128xf32>) -> pd_op.tensor<2xi32> { (shape[2], data[S1, 128]) }
(%5) = "pd_op.shape_broadcast" (%3, %4) {stop_gradient:[false],sym_shape_str:"shape[2], data[Broadcast(S1, S0), 128]"} : (pd_op.tensor<2xi32>, pd_op.tensor<2xi32>) -> pd_op.tensor<2xi32> { (shape[2], data[Broadcast(S1, S0), 128]) }
(%6) = "pd_op.expand" (%1, %5) {stop_gradient:[false],sym_shape_str:"shape[Broadcast(S1, S0), 128], data[NULL]"} : (pd_op.tensor<-1x128xf32>, pd_op.tensor<2xi32>) -> pd_op.tensor<-1x-1xf32> { (shape[Broadcast(S1, S0), 128], data[NULL]) }
(%7) = "pd_op.shape" (%2) {stop_gradient:[false],sym_shape_str:"shape[2], data[S0, 128]",symbolic_shape:} : (pd_op.tensor<-1x128xf32>) -> pd_op.tensor<2xi32> { (shape[2], data[S0, 128]) }
(%8) = "pd_op.shape" (%6) {stop_gradient:[false],sym_shape_str:"shape[2], data[Broadcast(S1, S0), 128]",symbolic_shape:} : (pd_op.tensor<-1x-1xf32>) -> pd_op.tensor<2xi32> { (shape[2], data[Broadcast(S1, S0), 128]) }
(%9) = "pd_op.shape_broadcast" (%7, %8) {stop_gradient:[false],sym_shape_str:"shape[2], data[Broadcast(Broadcast(S1, S0), S0), 128]"} : (pd_op.tensor<2xi32>, pd_op.tensor<2xi32>) -> pd_op.tensor<2xi32> { (shape[2], data[Broadcast(Broadcast(S1, S0), S0), 128]) }
(%10) = "pd_op.subtract" (%2, %6) {stop_gradient:[false],sym_shape_str:"shape[S0, 128], data[NULL]",symbolic_shape:} : (pd_op.tensor<-1x128xf32>, pd_op.tensor<-1x-1xf32>) -> pd_op.tensor<-1x128xf32> { (shape[S0, 128], data[NULL]) }
() = "builtin.shadow_output" (%10) {output_name:"output_0",sym_shape_str:"shape[S0, 128], data[NULL]"} : (pd_op.tensor<-1x128xf32>) ->  {  }

[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.