Xinyu302 · Xinyu302 · Jan 15, 2024 · Jan 2, 2024 · Jan 2, 2024 · Jan 2, 2024
diff --git a/.gitignore b/.gitignore
@@ -108,7 +108,7 @@ paddle/fluid/pir/dialect/operator/ir/pd_api.*
 paddle/fluid/pir/dialect/operator/ir/op_decomp.cc
 paddle/fluid/pir/dialect/operator/ir/pd_op_vjp.cc
 paddle/fluid/pir/dialect/operator/ir/pd_op.*
-paddle/fluid/pir/dialect/operator/ir/pd_onednn_op.*
+paddle/fluid/pir/dialect/operator/ir/onednn_op.*
 paddle/fluid/pir/dialect/operator/ir/pd_onednn_op_info.*
 paddle/fluid/pir/dialect/operator/ir/pd_op_bwd.*
 paddle/fluid/pir/dialect/operator/ir/pd_op_fused.*

diff --git a/README.md b/README.md
@@ -15,11 +15,11 @@ English | [简体中文](./README_cn.md) | [日本語](./README_ja.md)
 Welcome to the PaddlePaddle GitHub.
 
 PaddlePaddle, as the first independent R&D deep learning platform in China, has been officially open-sourced to professional communities since 2016. It is an industrial platform with advanced technologies and rich features that cover core deep learning frameworks, basic model libraries, end-to-end development kits, tools & components as well as service platforms.
-PaddlePaddle is originated from industrial practices with dedication and commitments to industrialization. It has been widely adopted by a wide range of sectors including manufacturing, agriculture, enterprise service, and so on while serving more than 8 million developers, 220,000 companies and generating 800,000 models. With such advantages, PaddlePaddle has helped an increasing number of partners commercialize AI.
+PaddlePaddle is originated from industrial practices with dedication and commitments to industrialization. It has been widely adopted by a wide range of sectors including manufacturing, agriculture, enterprise service, and so on while serving more than 10.7 million developers, 235,000 companies and generating 860,000 models. With such advantages, PaddlePaddle has helped an increasing number of partners commercialize AI.
 
 ## Installation
 
-### Latest PaddlePaddle Release: [v2.5](https://github.com/PaddlePaddle/Paddle/tree/release/2.5)
+### Latest PaddlePaddle Release: [v2.6](https://github.com/PaddlePaddle/Paddle/tree/release/2.6)
 
 Our vision is to enable deep learning for everyone via PaddlePaddle.
 Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest features of PaddlePaddle.

diff --git a/README_cn.md b/README_cn.md
@@ -14,13 +14,13 @@
 
 欢迎来到 PaddlePaddle GitHub
 
-飞桨(PaddlePaddle)以百度多年的深度学习技术研究和业务应用为基础，是中国首个自主研发、功能完备、 开源开放的产业级深度学习平台，集深度学习核心训练和推理框架、基础模型库、端到端开发套件和丰富的工具组件于一体。目前，飞桨累计开发者800万，服务企业22万家，基于飞桨开源深度学习平台产生了80万个模型。飞桨助力开发者快速实现AI想法，快速上线AI业务。帮助越来越多的行业完成AI赋能，实现产业智能化升级。
+飞桨(PaddlePaddle)以百度多年的深度学习技术研究和业务应用为基础，是中国首个自主研发、功能完备、 开源开放的产业级深度学习平台，集深度学习核心训练和推理框架、基础模型库、端到端开发套件和丰富的工具组件于一体。目前，飞桨累计开发者1070万，服务企业23.5万家，基于飞桨开源深度学习平台产生了86万个模型。飞桨助力开发者快速实现AI想法，快速上线AI业务。帮助越来越多的行业完成AI赋能，实现产业智能化升级。
 
 ## 安装
 
-### PaddlePaddle最新版本: [v2.5](https://github.com/PaddlePaddle/Paddle/tree/release/2.5)
+### PaddlePaddle 最新版本: [v2.6](https://github.com/PaddlePaddle/Paddle/tree/release/2.6)
 
-跟进PaddlePaddle最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases)
+跟进 PaddlePaddle 最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases)
 
 ### 安装最新稳定版本
 

diff --git a/README_ja.md b/README_ja.md
@@ -15,11 +15,11 @@
 PaddlePaddle GitHub へようこそ。
 
 PaddlePaddle は中国初の独立系 R&D ディープラーニングプラットフォームとして、2016年からプロのコミュニティに正式にオープンソース化されました。コアとなる深層学習フレームワーク、基本モデルライブラリ、エンドツーエンドの開発キット、ツール＆コンポーネント、さらにサービスプラットフォームを網羅する、高度な技術と豊富な機能を備えた産業プラットフォームです。
-PaddlePaddle は、工業化に対するコミットメントを持つ工業的実践から生まれたものです。製造業、農業、企業サービスなど幅広い分野で採用され、800万人以上の開発者、22万以上の企業、80万以上のモデルを生み出しています。それにより PaddlePaddle は、ますます多くのパートナーの AI 商用化を支援しています。
+PaddlePaddle は、工業化に対するコミットメントを持つ工業的実践から生まれたものです。製造業、農業、企業サービスなど幅広い分野で採用され、1070万人以上の開発者、23.5万以上の企業、86万以上のモデルを生み出しています。それにより PaddlePaddle は、ますます多くのパートナーの AI 商用化を支援しています。
 
 ## インストール
 
-### PaddlePaddle の最新リリース: [v2.5](https://github.com/PaddlePaddle/Paddle/tree/release/2.5)
+### PaddlePaddle の最新リリース: [v2.6](https://github.com/PaddlePaddle/Paddle/tree/release/2.6)
 
 私たちのビジョンは、PaddlePaddle を通じて、誰もが深層学習を行えるようにすることです。
 PaddlePaddle の最新機能を追跡するために、私たちの[リリースのお知らせ](https://github.com/PaddlePaddle/Paddle/releases)を参照してください。

diff --git a/cmake/external/xpu.cmake b/cmake/external/xpu.cmake
@@ -26,10 +26,10 @@ set(XPU_XBLAS_LIB_NAME "libxpu_blas.so")
 set(XPU_XFA_LIB_NAME "libxpu_flash_attention.so")
 
 if(NOT DEFINED XPU_BASE_DATE)
-  set(XPU_BASE_DATE "20231203")
+  set(XPU_BASE_DATE "20231218")
 endif()
 if(NOT DEFINED XPU_XHPC_BASE_DATE)
-  set(XPU_XHPC_BASE_DATE "20231229")
+  set(XPU_XHPC_BASE_DATE "20240105")
 endif()
 set(XPU_XCCL_BASE_VERSION "1.1.8.1")
 if(NOT DEFINED XPU_XFT_BASE_VERSION)

diff --git a/cmake/flags.cmake b/cmake/flags.cmake
@@ -37,6 +37,7 @@ checkcompilercxx14flag()
 if(NOT WIN32)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++17")
 else()
+  set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -std=c++17")
   set(CMAKE_CXX_STANDARD 17)
 endif()
 

diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake
@@ -328,10 +328,19 @@ copy(
   inference_lib_dist
   SRCS ${PADDLE_SOURCE_DIR}/paddle/phi/core/visit_type.h
   DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/phi/core/)
+
 copy(
   inference_lib_dist
-  SRCS ${PADDLE_SOURCE_DIR}/paddle/phi/core/hostdevice.h
-  DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/phi/core/)
+  SRCS ${PADDLE_SOURCE_DIR}/paddle/phi/core/distributed/type_defs.h
+  DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/phi/core/distributed/
+)
+
+copy(
+  inference_lib_dist
+  SRCS ${PADDLE_SOURCE_DIR}/paddle/phi/core/distributed/auto_parallel/*.h
+  DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/phi/core/distributed/auto_parallel/
+)
+
 copy(
   inference_lib_dist
   SRCS ${PADDLE_SOURCE_DIR}/paddle/fluid/platform/init_phi.h

diff --git a/paddle/cinn/backends/codegen_cuda_host.cc b/paddle/cinn/backends/codegen_cuda_host.cc
@@ -198,13 +198,23 @@ llvm::Value* CodeGenCUDA_Host::LowerHostFunc(const ir::_LoweredFunc_* func) {
                  [](auto& arg) { return std::addressof(arg); });
   // @}
 
+  // Set local scope table
+  CHECK_EQ(ll_function_args.size(), func->args.size());
+  for (int i = 0; i < ll_function_args.size(); ++i) {
+    SetVar(func->args[i].name(), ll_function_args[i]);
+  }
   llvm::BasicBlock* entry = llvm::BasicBlock::Create(
       /*Context=*/b_->getContext(),
       /*Name=*/"entry",
       /*Parent=*/f_,
       /*InsertBefore=*/nullptr);
   b_->SetInsertPoint(entry);
   CodeGenLLVM::Visit(&func->body);
+
+  // Reset local scope table
+  for (const ir::Argument& func_arg : func->args) {
+    symbol_table_->Erase(func_arg.name());
+  }
   RetVoid();
 
   return f_;

diff --git a/paddle/cinn/backends/codegen_cuda_host.h b/paddle/cinn/backends/codegen_cuda_host.h
@@ -53,7 +53,7 @@ class CodeGenCUDA_Host : public CodeGenLLVM {
     } else if (op->name == runtime::intrinsic::call_cuda_kernel) {
       return LowerCUDAKernelCall(op);
     } else {
-      CINN_NOT_IMPLEMENTED;
+      return CodeGenLLVM::Visit(op);
     }
   }
 

diff --git a/paddle/cinn/backends/codegen_cuda_util.h b/paddle/cinn/backends/codegen_cuda_util.h
@@ -31,6 +31,7 @@ namespace backends {
 #define KERNEL_ARGS "kernel_args"
 #define KERNEL_ARGS_NUM "kernel_args_num"
 #define KERNEL_STREAM "kernel_stream"
+#define TENSOR_SHAPE_ARGS "tensor_shape_args"
 
 /**
  * Split a CINN Module into two separate modules, one cantains the host
@@ -150,7 +151,8 @@ struct CollectBucketStrategyHostFunctionVisitor
       : CollectHostFunctionVisitor(module_name),
         kernel_args_(KERNEL_ARGS, type_of<void*>()),
         kernel_args_num_(KERNEL_ARGS_NUM, type_of<int>()),
-        kernel_stream_(KERNEL_STREAM, type_of<void*>()) {}
+        kernel_stream_(KERNEL_STREAM, type_of<void*>()),
+        tensor_shape_args_(TENSOR_SHAPE_ARGS, type_of<int32_t**>()) {}
 
   std::tuple<ir::Module, ir::Module> operator()(Expr* expr) {
     ir::IRMutator<>::Visit(expr, expr);
@@ -181,6 +183,25 @@ struct CollectBucketStrategyHostFunctionVisitor
                                 {});
     host_module_builder.AddFunctionWithoutOptim(
         host_func.as_lowered_func_ref());
+
+    // Parse LoweredFunc to infer output tensor's shape
+    std::vector<ir::Expr> infer_shape_func_body_stmts(arg_defs_);
+    infer_shape_func_body_stmts.insert(
+        infer_shape_func_body_stmts.end(),
+        op->infer_shape_func.as_lowered_func()->body);
+
+    std::vector<ir::Argument> infer_shape_arguments = {
+        ir::Argument(kernel_args_, ir::Argument::IO::kOutput),
+        ir::Argument(kernel_args_num_, ir::Argument::IO::kInput),
+        ir::Argument(tensor_shape_args_, ir::Argument::IO::kOutput)};
+
+    ir::Expr host_infer_shape_func =
+        ir::_LoweredFunc_::Make(op->infer_shape_func.as_lowered_func()->name,
+                                infer_shape_arguments,
+                                ir::Block::Make(infer_shape_func_body_stmts),
+                                {});
+    host_module_builder.AddFunctionWithoutOptim(
+        host_infer_shape_func.as_lowered_func_ref());
   }
 
   void ProcessLoweredFunc(ir::Expr func, ir::Expr predicate);
@@ -199,6 +220,7 @@ struct CollectBucketStrategyHostFunctionVisitor
   ir::Var kernel_args_;
   ir::Var kernel_args_num_;
   ir::Var kernel_stream_;
+  ir::Var tensor_shape_args_;
 };
 
 }  // namespace detail

diff --git a/paddle/cinn/backends/llvm/codegen_llvm.cc b/paddle/cinn/backends/llvm/codegen_llvm.cc
@@ -818,7 +818,8 @@ llvm::Value *CodeGenLLVM::Visit(const ir::_Var_ *op) {
   // TODO(fc500110) hard coding
   if (LLVM_WillVarLowerAsPointer(op->name)) {
     result = value;
-  } else if (value->getType()->isPointerTy()) {
+  } else if (value->getType()->isPointerTy() &&
+             !value->getType()->getPointerElementType()->isPointerTy()) {
     result = Load(value, op->name + "_load");
   } else {
     result = value;

diff --git a/paddle/cinn/common/CMakeLists.txt b/paddle/cinn/common/CMakeLists.txt
@@ -23,7 +23,9 @@ gather_srcs(
   nvgpu_dev_info.cc
   integer_set.cc
   dim_expr_simplify.cc
-  dim_expr_converter.cc)
+  dim_expr_converter.cc
+  broadcast_tree.cc
+  dim_expr_util.cc)
 
 cinn_cc_test(test_equation_graph_topo_walker SRCS
              equation_graph_topo_walker_test.cc DEPS gtest glog)
@@ -48,8 +50,10 @@ if(WITH_CUDA)
                gtest glog)
 endif()
 if(NOT CINN_ONLY)
+  cinn_cc_test(dim_expr_util_test SRCS dim_expr_util_test.cc DEPS cinncore)
   cinn_cc_test(dim_expr_simplify_test SRCS dim_expr_simplify_test.cc DEPS
                cinncore)
   cinn_cc_test(dim_expr_converter_test SRCS dim_expr_converter_test.cc DEPS
                cinncore)
+  cinn_cc_test(broadcast_tree_test SRCS broadcast_tree_test.cc DEPS cinncore)
 endif()