[PIR][Dy2St] Fix ReshapeOp cannot inplace in PIR mode

[SigureMo]

PR Category

Execute Infrastructure

PR Types

Bug fixes

Description

based on #69142

import paddle


def foo(x):
    y = x + 1
    z = y.reshape([2, 6])
    return z


static_foo = paddle.jit.to_static(foo, full_graph=True)
x = paddle.rand([3, 4], dtype='float32')
x.stop_gradient = False
out = static_foo(x)

ForwardProgram is :
{
    (%0) = "pd_op.data" [id:33] () {dtype:(pd_op.DataType)float32,name:"_jst.0.x.0",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[3,4],stop_gradient:[false]} : () -> builtin.tensor<3x4xf32>
    (%1) = "pd_op.full" [id:34] () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[1],stop_gradient:[true],value:(Double)1} : () -> builtin.tensor<1xf32>
    (%2) = "pd_op.scale" [id:35] (%0, %1) {bias:(Float)1,bias_after_scale:true,stop_gradient:[false]} : (builtin.tensor<3x4xf32>, builtin.tensor<1xf32>) -> builtin.tensor<3x4xf32>
    (%3) = "pd_op.full_int_array" [id:36] () {dtype:(pd_op.DataType)int64,place:(pd_op.Place)Place(cpu),stop_gradient:[true],value:[(Int64)2,(Int64)6]} : () -> builtin.tensor<2xi64>
    (%4) = "pd_op.reshape" [id:37] (%2, %3) {stop_gradient:[false]} : (builtin.tensor<3x4xf32>, builtin.tensor<2xi64>) -> builtin.tensor<2x6xf32>
    () = "builtin.shadow_output" [id:38] (%1) {output_name:"middle_0"} : (builtin.tensor<1xf32>) -> 
    () = "builtin.shadow_output" [id:39] (%2) {output_name:"middle_1"} : (builtin.tensor<3x4xf32>) -> 
    () = "builtin.shadow_output" [id:40] (%4) {output_name:"output_0"} : (builtin.tensor<2x6xf32>) -> 
}
BackwardProgram is:
{
    ^kw:%kwarg_output_grad_0, %kwarg_middle_0, %kwarg_middle_1
    (%0) = "pd_op.reshape_grad" [id:41] (%kwarg_middle_1, %kwarg_output_grad_0) {stop_gradient:[false]} : (builtin.tensor<3x4xf32>, builtin.tensor<2x6xf32>) -> builtin.tensor<3x4xf32>
    (%1) = "pd_op.scale" [id:42] (%0, %kwarg_middle_0) {bias:(Float)0,bias_after_scale:true,stop_gradient:[false]} : (builtin.tensor<3x4xf32>, builtin.tensor<1xf32>) -> builtin.tensor<3x4xf32>
    () = "builtin.shadow_output" [id:43] (%1) {output_name:"input_grad_0"} : (builtin.tensor<3x4xf32>) -> 
}

PIR 下，动转静 ReshapeOP 在存在反向的时候，输入会被反向用到（ShadowOutput），导致 ReshapeOp 不能 inplace，但这个输入实际是 no_need_buffer 的，因此在 Inplace Pass 中传入 no_need_buffer 信息，以确保其能成功 inplace

PCard-66972

[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.

[Copilot]

Copilot wasn't able to review any files in this pull request.

Files not reviewed (5)

• paddle/fluid/eager/to_static/run_program_op_node.h: Language not supported
• paddle/fluid/framework/executor_cache.cc: Language not supported
• paddle/fluid/framework/executor_cache.h: Language not supported
• paddle/fluid/pir/transforms/general/inplace_pass.cc: Language not supported
• paddle/fluid/pir/transforms/general/inplace_pass.h: Language not supported

[zhangbo9674]

LGTM