[PIR+CINN]Support all modules of PIR+CINN with Python unittest

[Aurelius84]

PR types

New features

PR changes

Others

Description

Pcard-67164

支持了子图 「exp+sub」 的端到端 PIR+CINN 执行。

调试命令：FLAG_logbuflevel=-1 FLAGS_print_ir=1 FLAGS_call_stack_level=2 GLOG_v=6 FLAGS_enable_pir_api=1 python test_cinn_sub_graph.py > err.log 2>&1

Step 1: Python 端原生 Program

ForwardProgram is :
{
 (%0) = "pd_op.data" () {dtype:(pd_op.DataType)float32,name:"_jst.0.x.0",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[64,128],stop_gradient:[false]} : () -> pd_op.tensor<64x128xf32>
 (%1) = "pd_op.exp" (%0) {stop_gradient:[false]} : (pd_op.tensor<64x128xf32>) -> pd_op.tensor<64x128xf32>
 (%2) = "pd_op.subtract" (%1, %0) {stop_gradient:[false]} : (pd_op.tensor<64x128xf32>, pd_op.tensor<64x128xf32>) -> pd_op.tensor<64x128xf32>
 () = "builtin.set_parameter" (%2) {parameter_name:"output_0"} : (pd_op.tensor<64x128xf32>) -> 
}

Step 2: After CINN related Pass

{
 (%0) = "pd_op.data" () {dtype:(pd_op.DataType)float32,name:"_jst.0.x.0",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[64,128],stop_gradient:[false]} : () -> pd_op.tensor<64x128xf32>
 (%1) = "cinn_runtime.jit_kernel" (%0) {jit_info:(0x46aa1f90)} : (pd_op.tensor<64x128xf32>) -> pd_op.tensor<64x128xf32>
 () = "builtin.set_parameter" (%1) {parameter_name:"output_0"} : (pd_op.tensor<64x128xf32>) -> 
}

Step 3: After Lowering and ToKernelPass

{
 (%0) = "data(phi_kernel)" () {dtype:(pd_op.DataType)float32,kernel_key:<backend:Undefined|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"data",name:"_jst.0.x.0",op_name:"pd_op.data",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[64,128],stop_gradient:[false]} : () -> undefined_tensor<64x128xf32>
 (%1) = "shadow_feed(phi_kernel)" (%0) {kernel_key:<backend:GPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"shadow_feed",op_name:"pd_op.shadow_feed"} : (undefined_tensor<64x128xf32>) -> gpu_tensor<64x128xf32>
 (%2) = "cinn_runtime.jit_kernel" (%1) {jit_info:(0x46aa1f90)} : (gpu_tensor<64x128xf32>) -> gpu_tensor<64x128xf32>
 () = "(phi_kernel)" (%2) {kernel_key:<backend:GPUDNN|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"",op_name:"builtin.set_parameter",parameter_name:"output_0"} : (gpu_tensor<64x128xf32>) -> 
}

[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.

[Aurelius84]

这里统一对global namespace 的全局static 变量命名加上 pir 前缀，避免libpaddle.so 与 libpir.so 中同名pass导致符号变量仅初始化一次，引起运行时另一个同名Pass未注册的错误。

[6clc]

LGTM of cmake