[CINN]Add remove_redundant_full_ops_pass

[leon062112]

PR Category

CINN

PR Types

Performance

Description

Pcard-67164

1. 添加remove_redundant_full_ops_pass, 将重复出现的full_int_array以及full op移除

Pass前后IR Printing

398: ===---------------------------------------------------------------------------------===
398:         IRPrinting on builtin.module before remove_redundant_full_ops_pass pass
398: ===---------------------------------------------------------------------------------===
398: {
398:     (%0) = "pd_op.data" [id:88] () {dtype:float32,name:"var_0",place:Place(undefined:0),shape:[-1,64,-1,-1],stop_gradient:[false]} : () -> tensor<-1x64x-1x-1xf32>
398:     (%1) = "pd_op.data" [id:89] () {dtype:float32,name:"var_1",place:Place(undefined:0),shape:[-1,64,-1,-1],stop_gradient:[false]} : () -> tensor<-1x64x-1x-1xf32>
398:     (%2) = "pd_op.full" [id:90] () {dtype:float64,place:Place(cpu),shape:[1],stop_gradient:[true],value:1} : () -> tensor<1xf64>
398:     (%3) = "pd_op.full" [id:91] () {dtype:float64,place:Place(cpu),shape:[1],stop_gradient:[true],value:11} : () -> tensor<1xf64>
398:     (%4) = "pd_op.full" [id:92] () {dtype:float64,place:Place(cpu),shape:[1],stop_gradient:[true],value:1} : () -> tensor<1xf64>
398:     (%5) = "pd_op.arange" [id:93] (%2, %3, %4) {dtype:int64,place:Place(undefined:0),stop_gradient:[true]} : (tensor<1xf64>, tensor<1xf64>, tensor<1xf64>) -> tensor<10xi64>
398:     (%6, %7) = "cinn_runtime.jit_kernel" [id:165] (%5, %0) {kernel_info:(fn_scale_cast_reshape_gs_bc_gs_bc_mul_)} : (tensor<10xi64>, tensor<-1x64x-1x-1xf32>) -> tensor<10x1x1x1xf32>, tensor<10x64x-1x-1xf32>
398:     (%8) = "cinn_runtime.jit_kernel" [id:166] (%1, %6) {kernel_info:(fn_gs_bc_gs_bc_mul_)} : (tensor<-1x64x-1x-1xf32>, tensor<10x1x1x1xf32>) -> tensor<10x64x-1x-1xf32>
398:     (%9) = "cinn_runtime.jit_kernel" [id:167] (%0, %1, %7, %8) {kernel_info:(fn_gs_bc_gs_bc_add_)} : (tensor<-1x64x-1x-1xf32>, tensor<-1x64x-1x-1xf32>, tensor<10x64x-1x-1xf32>, tensor<10x64x-1x-1xf32>) -> tensor<10x64x-1x-1xf32>
398:     () = "builtin.shadow_output" [id:102] (%6) {output_name:"middle_0"} : (tensor<10x1x1x1xf32>) -> 
398:     () = "builtin.shadow_output" [id:103] (%7) {output_name:"middle_1"} : (tensor<10x64x-1x-1xf32>) -> 
398:     () = "builtin.shadow_output" [id:104] (%8) {output_name:"middle_2"} : (tensor<10x64x-1x-1xf32>) -> 
398:     () = "builtin.shadow_output" [id:105] (%9) {output_name:"output_0"} : (tensor<10x64x-1x-1xf32>) -> 
398: }

398: ===--------------------------------------------------------------------------------===
398:         IRPrinting on builtin.module after remove_redundant_full_ops_pass pass
398: ===--------------------------------------------------------------------------------===
398: {
398:     (%0) = "pd_op.data" [id:88] () {dtype:float32,name:"var_0",place:Place(undefined:0),shape:[-1,64,-1,-1],stop_gradient:[false]} : () -> tensor<-1x64x-1x-1xf32>
398:     (%1) = "pd_op.data" [id:89] () {dtype:float32,name:"var_1",place:Place(undefined:0),shape:[-1,64,-1,-1],stop_gradient:[false]} : () -> tensor<-1x64x-1x-1xf32>
398:     (%2) = "pd_op.full" [id:90] () {dtype:float64,place:Place(cpu),shape:[1],stop_gradient:[true],value:1} : () -> tensor<1xf64>
398:     (%3) = "pd_op.full" [id:91] () {dtype:float64,place:Place(cpu),shape:[1],stop_gradient:[true],value:11} : () -> tensor<1xf64>
398:     (%4) = "pd_op.arange" [id:93] (%2, %3, %2) {dtype:int64,place:Place(undefined:0),stop_gradient:[true]} : (tensor<1xf64>, tensor<1xf64>, tensor<1xf64>) -> tensor<10xi64>
398:     (%5, %6) = "cinn_runtime.jit_kernel" [id:165] (%4, %0) {kernel_info:(fn_scale_cast_reshape_gs_bc_gs_bc_mul_)} : (tensor<10xi64>, tensor<-1x64x-1x-1xf32>) -> tensor<10x1x1x1xf32>, tensor<10x64x-1x-1xf32>
398:     (%7) = "cinn_runtime.jit_kernel" [id:166] (%1, %5) {kernel_info:(fn_gs_bc_gs_bc_mul_)} : (tensor<-1x64x-1x-1xf32>, tensor<10x1x1x1xf32>) -> tensor<10x64x-1x-1xf32>
398:     (%8) = "cinn_runtime.jit_kernel" [id:167] (%0, %1, %6, %7) {kernel_info:(fn_gs_bc_gs_bc_add_)} : (tensor<-1x64x-1x-1xf32>, tensor<-1x64x-1x-1xf32>, tensor<10x64x-1x-1xf32>, tensor<10x64x-1x-1xf32>) -> tensor<10x64x-1x-1xf32>
398:     () = "builtin.shadow_output" [id:102] (%5) {output_name:"middle_0"} : (tensor<10x1x1x1xf32>) -> 
398:     () = "builtin.shadow_output" [id:103] (%6) {output_name:"middle_1"} : (tensor<10x64x-1x-1xf32>) -> 
398:     () = "builtin.shadow_output" [id:104] (%7) {output_name:"middle_2"} : (tensor<10x64x-1x-1xf32>) -> 
398:     () = "builtin.shadow_output" [id:105] (%8) {output_name:"output_0"} : (tensor<10x64x-1x-1xf32>) -> 
398: }

[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.