New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[CINN] Strong constraint branch support dynamic shape #59309

Merged

jiahy0825 merged 42 commits into PaddlePaddle:develop from jiahy0825:adt-support-dynamic-shape

Dec 6, 2023

Contributor

jiahy0825 commented Nov 23, 2023 •

edited

Loading

PR types

New features

PR changes

Others

Description

pcard-76996

强约束分支支持动态形状，使用 ShapeDialect 中的约束关系构建方程组，然后使用方程组进行符号化推导，得到每个迭代量对应的符号表达式。

exp + sub 最小子图，求解得到的 MapExpr 如下：

MapExprTest(t_var_2, t_var_1) {
  AnchoredMapStmt(t_var_0) {
    MapStmt([i_59, i_60]) {
      exp(
          &t_var[IndexDot([BI(i_59, sym_17), 0], [sym_17, 1])],
          t_var_1[IndexDot([BI(i_59, sym_17), 0], [sym_17, 1])]);
      subtract(
          &t_var_0[IndexDot([i_59, i_60], [sym_17, 1])],
          t_var_2[IndexDot([BI(i_59, sym_17), 0], [sym_17, 1])],
          t_var[IndexDot([BI(i_59, sym_17), 0], [sym_17, 1])]);
    }
  }
}

jiahy0825 added 24 commits

November 14, 2023 09:19


          Strong Constraint Branch

c393fdd


          NoInlineTranslator (#84)

5ec684c


          Adapt adt to pir

7d31ca8


          Move FLAGS_cinn_enable_map_expr_schedule location

880e8fa


          Apply new group schedule

b867a18


          Remove useless log

948eae2


          Remove adt unittest

2ace0f1


          Solve merge conflicts

e06cfd3


          Fix typo


          Fix merge conflicts

41f5a80


          Add unit test

f84f518


          Fix cmake

673277a


          Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

c4be097

… rebase-adapt-to-pir


          Add test_cinn_sub_graph_map_expr

a4d44ea


          Merge commit 'refs/pull/58993/head' of github.com:PaddlePaddle/Paddle…

… into map-expr-pir-shape-dialect


          transfer origin code to develop

1821acd


          Fully support ShapeDialect

f58de6c


          TranslateDimExpr

fcad232


          Fix dim_expr_simplifier

122dd15


          Generate ir with symbolic

8a9a113


          Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

c3c540f

… map-expr-pir-shape-dialect


          Solve compile error

28d9b95


          SymbolicDim to SymbolicDimOp when Translate ir

7403d9e


          Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

60d8f31

… adt-support-dynamic-shape

paddle-bot bot commented Nov 23, 2023

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.

jiahy0825 added 5 commits

November 23, 2023 12:49


          Solve some conflict

3c7f2fa


          Solve conflict

254ac49


          ShapeAnalysis

a9d8026


          fix compile error

fd5e3f7


          Disable dynamic shape

a11a164

jiahy0825 added 12 commits

November 24, 2023 11:58


          Add scale generate_equation

483d69d


          Add cpp unittest

c46b794


          Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

b011f6a

… adt-support-dynamic-shape


          Add cpp unittest

ac68071


          Change VLOG priority

04ef03a


          Clean IndexDot

8e110ae


          Unittest

9a3031e


          Cancel SimplifyDotBI

1177f46


          UniqueId ResetSeqNumber

6ef3753


          input_spec and kDynamic

154915e


          map_expr_test

39394b3


          Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

602df3e

… adt-support-dynamic-shape

Aurelius84 previously approved these changes

View reviewed changes

Contributor

Aurelius84 left a comment

Great work!

paddle/cinn/adt/adapter_dynamic_tensor.h

+                }
+              };
+              inline std::size_t GetHashValueImpl(const DynamicTensor& tensor) {

Contributor

Aurelius84 Dec 1, 2023

size_t 作为hash_key是有隐藏的hash冲突问题。对于Hash Map，当出现key的冲突时，会再次调用operator==。但直接以size_t作为key，二阶段的判断operator==是无效的，请注意这个风险会带来难排查的问题。

Contributor

Aurelius84 Dec 1, 2023

其实 Value 可以直接作为HashKey的

Contributor Author

jiahy0825 Dec 1, 2023 •

edited

Loading

感谢提醒 😃 目前是直接把 pir::Value 当作 HashKey 来用的～
此处在重载 Tensor 的 std::hash 函数，由于 DynamicTensor 是 Tensor 的子类型，因此强约束分支内须实现 DynamicTensor 的 GetHashValueImpl 方法

paddle/cinn/adt/adt.h

@@ @@ -74,6 +84,7 @@ class Tuple { @@
                           std::make_shared<std::tuple<Ts...>>(std::forward<Args>(args)...)) {}
                 const std::tuple<Ts...>& tuple() const { return *tuple_; }
+                std::tuple<Ts...>* mut_tuple() { return &*tuple_; }

Contributor

Aurelius84 Dec 1, 2023

mut 是表示 mutable？，有点Rust 的风格~~，框架内部用的最多的是mutable全称，保持一致是否好一些？

Contributor Author

jiahy0825 Dec 1, 2023 •

edited

Loading

是的，表示 mutable～这里后续提个 PR 修改下

paddle/cinn/adt/dim_expr_simplifier.cc

+                  } else {
+                    return Op<DimExpr>{ret_operand};
+                  }
+                  LOG(FATAL) << "Dead code.";

Contributor

Aurelius84 Dec 1, 2023

这一行是不是永远都不会触发？

Contributor Author

jiahy0825 Dec 1, 2023

随着未来可能的不断迭代，if 分支的逻辑可能变复杂，这里是写的防御性代码，保证触发 if-else 分支，不会到达这一行。

paddle/cinn/adt/dim_expr_simplifier.cc

+                  } else {
+                    return expr;
+                  }
+                  LOG(FATAL) << "Dead code";

Contributor

Aurelius84 Dec 1, 2023

同上，下面似乎也有

Contributor Author

jiahy0825 Dec 1, 2023

同上

paddle/cinn/adt/dim_expr_simplifier.cc

+              template <>
+              struct GetOrderValue<BroadcastedDim<DimExpr>> {
+                static constexpr int value = 10;

Contributor

Aurelius84 Dec 1, 2023

这里的value 似乎与IsLhsBeforeRhs 逻辑有关系，想多了解下这里定义order的依据是什么，比如为什么Product 放在 Sum 之前？ std::int64_t> 对应的order既没有放着最前面，也没有放到最后面，而是重新中间，这个又是为什么？

Contributor

Aurelius84 Dec 1, 2023

是运算的优先级？类似乘法高于加法？

Contributor Author

jiahy0825 Dec 1, 2023

此处 order 的值没有特别的依据，这里的 order 主要还是为了排序时相同类型的 DimExpr 可以放在一起。
这样做有利于简化 DimExpr 的判等

paddle/cinn/adt/dim_expr_simplifier.cc

+              };
+              std::int64_t GetInteger(const DimExpr& expr) {
+                if (expr.Has<Negative<DimExpr>>()) {

Contributor

Aurelius84 Dec 1, 2023

这里的Has感觉像是 Is的语义？或者对最外层判断Is语义？

Contributor Author

jiahy0825 Dec 1, 2023 •

edited

Loading

Has 在判断 DimExpr 最外层的子类型是什么，确实很贴近 Is 的语义

paddle/cinn/adt/dim_expr_simplifier.cc

+              template <>
+              struct GetInversed<Sum> {
+                static DimExpr Call(const DimExpr& expr) { return Negative<DimExpr>(expr); }

Contributor

Aurelius84 Dec 1, 2023

S1 + S2 的反义是 -(S1+S2) ? 还是S1 - S2 ?

Contributor Author

jiahy0825 Dec 1, 2023

如果只看这里的 Inversed 实现，S1 + S2 对应的是 -(S1 + S2)
实际上调用的地方在 FlattenOperands pass 内，会先把 S1 + S2 展开，然后再取负，经过 pass 优化后表达式会变为 -S1 - S2

paddle/cinn/adt/dim_expr_simplifier.cc

+                DimExpr ret = expr;
+                for (bool keep_rewrite = true; keep_rewrite;) {
+                  keep_rewrite = false;
+                  DoPass<SimplifyOneOperand<Negative>>(&keep_rewrite, &ret);

Contributor

Aurelius84 Dec 1, 2023

这个是一个逐步收敛的优化过程是吧？这里有评估执行效率如何么，会有可感知的耗时么？

Contributor Author

jiahy0825 Dec 1, 2023

这个是一个逐步收敛的优化过程是吧？

是的。没有实际测过，但 DimExpr 相对来说都比较简单，预计不会成为性能瓶颈

paddle/cinn/adt/igroup.h

-                        constants_provider)
+                explicit IGroup(const List<OpStmt>& op_stmts,
+                                const AnchorIndex& anchor_index,
+                                const EquationCtx4OpStmtT& EquationCtx4OpStmt)

Contributor

Aurelius84 Dec 1, 2023

这里4 表示 For ?

Contributor Author

jiahy0825 Dec 1, 2023

是的～

ZzSean reviewed

View reviewed changes

paddle/cinn/runtime/flags.cc

+              PD_DEFINE_bool(cinn_enable_map_expr_index_detail,
+                             BoolFromEnv("FLAGS_cinn_enable_map_expr_index_detail", false),
+                             "It controls whether to display datail tensor index");

Contributor

ZzSean Dec 4, 2023 •

edited

Loading

datail->detailed

Contributor Author

jiahy0825 Dec 5, 2023

Thanks～下个 PR 改

ZzSean reviewed

View reviewed changes

paddle/cinn/runtime/flags.cc

@@ @@ -91,6 +91,14 @@ PD_DEFINE_bool(cinn_enable_map_expr_inline, @@
                              BoolFromEnv("FLAGS_cinn_enable_map_expr_inline", false),
                              "It controls whether to inline by map_expr");
+              PD_DEFINE_bool(cinn_enable_map_expr_dynamic_shape,
+                             BoolFromEnv("FLAGS_cinn_enable_map_expr_dynamic_shape", false),

Contributor

ZzSean Dec 4, 2023

这个flag好像没用到

Contributor Author

jiahy0825 Dec 5, 2023

在 test/cpp/pir/cinn/adt/map_expr_test.cc 单测以及 paddle/cinn/adt/graph_symbolic_dim_infer_ctx.cc 文件中使用，表示会根据 shape dialect 进行符号推导


          Merge branch 'develop' of github.com:PaddlePaddle/Paddle into adt-sup…

7c68623

…port-dynamic-shape

jiahy0825 dismissed Aurelius84’s stale review via

7c68623

December 5, 2023 02:39

ZzSean approved these changes

View reviewed changes

Contributor

ZzSean left a comment

LGTM

tianshuo78520a approved these changes

View reviewed changes

XiaoguangHu01 approved these changes

View reviewed changes

Contributor

XiaoguangHu01 left a comment

LGTM

jiahy0825 merged commit 74dd107 into PaddlePaddle:develop

29 checks passed

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment