[PIR] support matrix_norm and fix backward redundant cast

[gouzil]

PR Category

Others

PR Types

Others

Description

• 拆分TestSundryAPIStatic.test_linalg_norm 单测
• 适配matrix_norm 的 p_matrix_norm 和 nuclear_norm 两个方法
• 删除 backward 中不必要的pd_op.cast fix auto backward bug #63113

原:

{
    (%0) = "pd_op.full" () {dtype:(pd_op.DataType)int64,place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[2,2],stop_gradient:[true],value:(Float)0} : () -> builtin.tensor<2x2xi64>
    (%1) = "pd_op.assign_value_" (%0) {dtype:(pd_op.DataType)int64,place:(pd_op.Place)Place(undefined:0),shape:[(Int32)2,(Int32)2],stop_gradient:[true],values:[(Double)3,(Double)3,(Double)3,(Double)3]} : (builtin.tensor<2x2xi64>) -> builtin.tensor<2x2xi64>
    (%2) = "pd_op.cast" (%1) {dtype:(pd_op.DataType)float32,stop_gradient:[false]} : (builtin.tensor<2x2xi64>) -> builtin.tensor<2x2xf32>
    (%3) = "pd_op.full" () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[1],stop_gradient:[true],value:(Float)1} : () -> builtin.tensor<1xf32>
    (%4) = "pd_op.full_like" (%2, %3) {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(undefined:0),stop_gradient:[false]} : (builtin.tensor<2x2xf32>, builtin.tensor<1xf32>) -> builtin.tensor<2x2xf32>
    (%5) = "pd_op.cast" (%4) {dtype:(pd_op.DataType)int64,stop_gradient:[false]} : (builtin.tensor<2x2xf32>) -> <<NULL TYPE>>
}

现:

{
    (%0) = "pd_op.full" () {dtype:(pd_op.DataType)int64,place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[2,2],stop_gradient:[true],value:(Float)0} : () -> builtin.tensor<2x2xi64>
    (%1) = "pd_op.assign_value_" (%0) {dtype:(pd_op.DataType)int64,place:(pd_op.Place)Place(undefined:0),shape:[(Int32)2,(Int32)2],stop_gradient:[true],values:[(Double)3,(Double)3,(Double)3,(Double)3]} : (builtin.tensor<2x2xi64>) -> builtin.tensor<2x2xi64>
    (%2) = "pd_op.cast" (%1) {dtype:(pd_op.DataType)float32,stop_gradient:[false]} : (builtin.tensor<2x2xi64>) -> builtin.tensor<2x2xf32>
    (%3) = "pd_op.full" () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[1],stop_gradient:[true],value:(Float)1} : () -> builtin.tensor<1xf32>
    (%4) = "pd_op.full_like" (%2, %3) {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(undefined:0),stop_gradient:[false]} : (builtin.tensor<2x2xf32>, builtin.tensor<1xf32>) -> builtin.tensor<2x2xf32>
}

TODO: 在mac下TestSundryAPIStatic.test_linalg_cond会断错误

相关链接:

• [PIR] Adaptation of test_zero_dim_*.py - step: 2 #62652

[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.

[SigureMo]

那几个不是随机挂，得看看

[SigureMo]

另外，breakward 是啥

[gouzil]

另外，breakward 是啥

打错了！！！(超小声)

[SigureMo]

LGTMeow