Merge branch 'PaddlePaddle:develop' into divide

PaddlePaddle · Oct 8, 2024 · 7f8d0f2 · 7f8d0f2
2 parents cdde50f + 8db2e3d
commit 7f8d0f2
Show file tree

Hide file tree

Showing 1,041 changed files with 18,004 additions and 13,528 deletions.
diff --git a/.github/CODEOWNERS b/.github/CODEOWNERS
@@ -27,10 +27,10 @@ paddle/fluid/framework/tensor.h @XiaoguangHu01 @zhiqiu @Xreki @Aurelius84
 paddle/fluid/framework/unused_var_check.cc @zhiqiu @phlrain
 paddle/fluid/framework/var_desc.h @XiaoguangHu01 @zhiqiu @Xreki @Aurelius84
 paddle/fluid/operators/distributed/send_recv.proto.in @gongweibao @seiriosPlus
-paddle/fluid/prim/api/api.yaml @cxxly @xiaoguoguo626807 @Charles-hit @cyber-pioneer @JiabinYang
-paddle/fluid/prim/api/composite_backward/composite_backward_api.h @cxxly @xiaoguoguo626807 @Charles-hit @cyber-pioneer @JiabinYang
-paddle/fluid/prim/api/composite_backward/composite_double_backward_api.h @cxxly @xiaoguoguo626807 @Charles-hit @cyber-pioneer @JiabinYang
-paddle/fluid/prim/api/manual_prim/prim_manual_api.h @cxxly @xiaoguoguo626807 @Charles-hit @cyber-pioneer @JiabinYang
+paddle/fluid/prim/api/api.yaml @xiaoguoguo626807 @JiabinYang @phlrain
+paddle/fluid/prim/api/composite_backward/composite_backward_api.h @xiaoguoguo626807 @JiabinYang
+paddle/fluid/prim/api/composite_backward/composite_double_backward_api.h @xiaoguoguo626807 @JiabinYang
+paddle/fluid/prim/api/manual_prim/prim_manual_api.h @xiaoguoguo626807 @JiabinYang
 paddle/phi/api/include/tensor.h @phlrain @zyfncg @YuanRisheng
 paddle/phi/core/attribute.h @phlrain @zyfncg @YuanRisheng
 paddle/phi/core/dense_tensor.h @phlrain @zyfncg @YuanRisheng
@@ -47,8 +47,8 @@ paddle/phi/infermeta/spmd_rules @LiYuRio @ForFishes @zhiqiu
 paddle/scripts/paddle_build.bat @zhwesky2010 @wanghuancoder @Aurelius84
 paddle/scripts/paddle_build.sh @risemeup1 @zhangbo9674 @XieYunshen
 pyproject.toml @SigureMo @gouzil
-python/paddle/autograd/backward_utils.py @Aurelius84 @cxxly @xiaoguoguo626807 @changeyoung98 @cyber-pioneer
-python/paddle/autograd/ir_backward.py @Aurelius84 @cxxly @xiaoguoguo626807 @changeyoung98
+python/paddle/autograd/backward_utils.py @Aurelius84 @xiaoguoguo626807 @changeyoung98 @phlrain
+python/paddle/autograd/ir_backward.py @Aurelius84 @xiaoguoguo626807 @changeyoung98
 python/paddle/base/backward.py @XiaoguangHu01 @zhiqiu @Xreki @qili93 @Aurelius84
 python/paddle/base/compiler.py @XiaoguangHu01 @zhiqiu @Xreki @qili93 @Aurelius84
 python/paddle/base/dygraph/layers.py @JiabinYang @phlrain
@@ -65,8 +65,8 @@ python/paddle/base/tests/unittests/white_list/op_threshold_white_list.py @juncai
 python/paddle/distributed/fleet/__init__.py @sneaxiy @raindrops2sea
 python/paddle/distributed/fleet/launch.py @sneaxiy @raindrops2sea
 python/paddle/distributed/__init__.py @sneaxiy @raindrops2sea
-python/paddle/incubate/autograd/composite_rules.py @cyber-pioneer @xiaoguoguo626807 @Charles-hit @JiabinYang
-python/paddle/incubate/autograd/primitives.py @cyber-pioneer @xiaoguoguo626807 @Charles-hit @JiabinYang
+python/paddle/incubate/autograd/composite_rules.py @xiaoguoguo626807 @JiabinYang
+python/paddle/incubate/autograd/primitives.py @xiaoguoguo626807 @JiabinYang @phlrain
 python/paddle/_typing @SigureMo @zrr1999 @gouzil
 python/requirements.txt @phlrain @jzhang533 @kolinwei
 test/dygraph_to_static @SigureMo @Aurelius84 @gouzil

diff --git a/cmake/external/dlpack.cmake b/cmake/external/dlpack.cmake
@@ -15,7 +15,7 @@
 include(ExternalProject)
 
 set(DLPACK_PREFIX_DIR ${THIRD_PARTY_PATH}/dlpack)
-set(DLPACK_TAG v0.4)
+set(DLPACK_TAG v0.8)
 set(DLPACK_INCLUDE_DIR ${THIRD_PARTY_PATH}/dlpack/src/extern_dlpack/include)
 set(SOURCE_DIR ${PADDLE_SOURCE_DIR}/third_party/dlpack)
 include_directories(${SOURCE_DIR}/include)

diff --git a/cmake/external/xpu.cmake b/cmake/external/xpu.cmake
@@ -30,9 +30,9 @@ if(NOT DEFINED XPU_XRE_BASE_VERSION)
   set(XPU_XRE_BASE_VERSION "4.32.0.1")
 endif()
 if(NOT DEFINED XPU_XHPC_BASE_DATE)
-  set(XPU_XHPC_BASE_DATE "eb35/20240906")
+  set(XPU_XHPC_BASE_DATE "eb35/20240927")
 endif()
-set(XPU_XCCL_BASE_VERSION "1.2.11c")
+set(XPU_XCCL_BASE_VERSION "1.2.11e")
 if(NOT DEFINED XPU_XFT_BASE_VERSION)
   set(XPU_XFT_BASE_VERSION "20230602")
 endif()

diff --git a/paddle/cinn/adt/equation_solver.cc b/paddle/cinn/adt/equation_solver.cc
@@ -12,6 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <typeinfo>
 #include <unordered_map>
 #include <variant>
 
@@ -33,15 +34,23 @@ std::unordered_map<Variable, Value> InferValuesImpl(
     IndexExprInferContext* ctx) {
   const auto& [out_iter, in_iter] = id.tuple();
   Variable in_variable{in_iter.value()};
-  CHECK(ctx->HasValue(in_variable));
+  PADDLE_ENFORCE_EQ(
+      ctx->HasValue(in_variable),
+      true,
+      ::common::errors::NotFound("The param id's out_iter must contain "
+                                 "its in_iter's value"));
   return {{out_iter.value(), ctx->GetValue(in_variable)}};
 }
 
 std::unordered_map<Variable, Value> InferValuesImpl(
     const Identity<tOut<Index>, tIn<Index>>& id, IndexExprInferContext* ctx) {
   const auto& [out_index, in_index] = id.tuple();
   Variable in_variable{in_index.value()};
-  CHECK(ctx->HasValue(in_variable));
+  PADDLE_ENFORCE_EQ(
+      ctx->HasValue(in_variable),
+      true,
+      ::common::errors::NotFound("The param id's out_iter must contain "
+                                 "its in_iter's value"));
   return {{out_index.value(), ctx->GetValue(in_variable)}};
 }
 
@@ -203,13 +212,30 @@ std::unordered_map<Variable, Value> InferValuesImpl(
           in_msg_out_indexes.value()->size()));
   for (std::size_t i = 0; i < out_msg_in_indexes.value()->size(); ++i) {
     const auto& value = ctx->GetValue(in_msg_in_indexes.value()->at(i));
-    CHECK(ret.emplace(out_msg_in_indexes.value()->at(i), value).second);
+    PADDLE_ENFORCE_EQ(
+        ret.emplace(out_msg_in_indexes.value()->at(i), value).second,
+        true,
+        ::common::errors::AlreadyExists([&]() {
+          std::ostringstream oss;
+          oss << "Failed to insert the variable '"
+              << "out_msg_in_indexes.value()->at(" << i
+              << ")' into the map: key already exists.";
+          return oss.str();
+        }()));
   }
   for (std::size_t i = 0; i < out_msg_out_indexes.value()->size(); ++i) {
     const auto& value = ctx->GetValue(in_msg_out_indexes.value()->at(i));
     const auto& out_index = out_msg_out_indexes.value()->at(i);
     if (out_index.has_value()) {
-      CHECK(ret.emplace(out_index.value(), value).second);
+      PADDLE_ENFORCE_EQ(ret.emplace(out_index.value(), value).second,
+                        true,
+                        ::common::errors::AlreadyExists([&]() {
+                          std::ostringstream oss;
+                          oss << "Failed to insert the variable '"
+                              << "out_index.value()"
+                              << "' into the map: key already exists.";
+                          return oss.str();
+                        }()));
     }
   }
   return ret;
@@ -265,6 +291,12 @@ tValueInferSuccess<bool> MergeInferedValuesIntoCtx(const Function* function,
       });
 }
 
+std::string GetFunctionName(const Function* function) {
+  return std::visit(
+      [](auto&& arg) -> std::string { return typeid(arg).name(); },
+      function->variant());
+}
+
 void SolveEquations(
     const EquationGraphTopoWalker<Variable, const Function*>& walker,
     const std::vector<Variable>& starts,
@@ -273,7 +305,16 @@ void SolveEquations(
       starts.begin(), starts.end(), [&](const Function* function) {
         tValueInferSuccess<bool> has_unique_value =
             MergeInferedValuesIntoCtx(function, ctx);
-        CHECK(has_unique_value.value());
+        PADDLE_ENFORCE_EQ(
+            has_unique_value.value(),
+            true,
+            ::common::errors::InvalidArgument([&]() {
+              std::ostringstream oss;
+              oss << "Failed to merge inferred values into the context for "
+                     "function '"
+                  << GetFunctionName(function) << "'.";
+              return oss.str();
+            }()));
       });
 }
 

diff --git a/paddle/cinn/backends/codegen_invoke_module.cc b/paddle/cinn/backends/codegen_invoke_module.cc
@@ -62,38 +62,6 @@ llvm::Value* CodeGenInvokeModule::LowerInvokeFunc(
   return f_;
 }
 
-llvm::Value* CodeGenInvokeModule::LowerParseArgsValueCall(
-    const ir::Call* call_ir) {
-  auto ret_type = CinnTypeToLLVMType(Int(64), m_);
-  std::vector<llvm::Type*> args_type;
-  PADDLE_ENFORCE_EQ(
-      call_ir->read_args.size(),
-      2,
-      ::common::errors::InvalidArgument(
-          "The number of arguments of ParseArgsValue should be 2"));
-  PADDLE_ENFORCE_EQ(call_ir->read_args[0].is_var() &&
-                        call_ir->read_args[0].as_var()->type().is_cpp_handle(),
-                    true,
-                    ::common::errors::InvalidArgument(
-                        "The first read argument must be a variable "
-                        "with a C++ handle type."));
-
-  PADDLE_ENFORCE_EQ(call_ir->read_args[1].type().is_int(32),
-                    true,
-                    ::common::errors::InvalidArgument(
-                        "The second read argument must be of type int32."));
-  args_type.push_back(CinnTypeToLLVMType(type_of<void*>(), m_));
-  args_type.push_back(CinnTypeToLLVMType(type_of<int32_t>(), m_));
-
-  auto func_type = llvm::FunctionType::get(ret_type, args_type, false);
-  auto call_func = m_->getOrInsertFunction(call_ir->name, func_type);
-
-  std::vector<llvm::Value*> call_args;
-  call_args.push_back(std::addressof(*f_->arg_begin()));
-  call_args.push_back(b_->getInt32(call_ir->read_args[1].as_int32()));
-  return b_->CreateCall(call_func, call_args);
-}
-
 llvm::Value* CodeGenSwitchHost::LowerInnerCaseCall(const ir::Call* op) {
   std::vector<llvm::Value*> ll_function_args;
   std::transform(f_->arg_begin(),

diff --git a/paddle/cinn/backends/codegen_invoke_module.h b/paddle/cinn/backends/codegen_invoke_module.h
@@ -43,19 +43,8 @@ class CodeGenInvokeModule : public CodeGenLLVM {
     return LowerInvokeFunc(func);
   }
 
-  llvm::Value *Visit(const ir::Call *op) override {
-    // TODO(Hongqing-work): change intrinsic name to get_value_in_kernel_args
-    if (op->name == runtime::intrinsic::get_value_in_cuda_kernel_args) {
-      return LowerParseArgsValueCall(op);
-    } else {
-      return CodeGenLLVM::Visit(op);
-    }
-  }
-
  protected:
   llvm::Value *LowerInvokeFunc(const ir::_LoweredFunc_ *func);
-
-  llvm::Value *LowerParseArgsValueCall(const ir::Call *call_ir);
 };
 
 class CodeGenHost : public CodeGenInvokeModule {
@@ -80,7 +69,7 @@ class CodeGenSwitchHost : public CodeGenInvokeModule {
   // only support call of args get function and inner case host function call
   llvm::Value *Visit(const ir::Call *op) override {
     if (op->name == runtime::intrinsic::get_value_in_cuda_kernel_args) {
-      return CodeGenInvokeModule::LowerParseArgsValueCall(op);
+      return CodeGenLLVM::Visit(op);
     } else {
       return LowerInnerCaseCall(op);
     }

diff --git a/paddle/cinn/backends/llvm/codegen_llvm.cc b/paddle/cinn/backends/llvm/codegen_llvm.cc
@@ -511,53 +511,6 @@ llvm::Value *CodeGenLLVM::Visit(const ir::Cast *op) {
 llvm::Value *CodeGenLLVM::CreateSerialFor(const ir::For *op, int stride) {
   SymbolTableGuard symbol_table_guard(*symbol_table_);
 
-  do {
-    break;
-    llvm::BasicBlock *preheader_bb = b_->GetInsertBlock();
-    auto *for_begin = llvm::BasicBlock::Create(
-        b_->getContext(), "for_begin", b_->GetInsertBlock()->getParent());
-    auto *for_body = llvm::BasicBlock::Create(
-        b_->getContext(), "for_body", b_->GetInsertBlock()->getParent());
-    auto *for_end = llvm::BasicBlock::Create(
-        b_->getContext(), "for_end", b_->GetInsertBlock()->getParent());
-
-    Br(for_begin);
-    b_->SetInsertPoint(for_begin);
-
-    auto *begin = Visit(&op->min);
-    auto *loop_value = PHI(begin->getType(), 2);
-    loop_value->addIncoming(begin, preheader_bb);
-
-    llvm::Value *old_var = GetVar(op->loop_var->name);
-    SetVar(op->loop_var->name, loop_value);
-    auto *end = Visit(&op->extent);
-    CondBr(ICmpSLT(loop_value, end), for_body, for_end);
-    b_->SetInsertPoint(for_body);
-    Visit(&op->body);
-
-    if (old_var) {
-      SetVar(op->loop_var->name, old_var);
-    } else {
-      symbol_table_->Erase(op->loop_var->name);
-    }
-
-    auto loop_next = Add(loop_value,
-                         llvm::ConstantInt::get(b_->getInt32Ty(), stride),
-                         "indvar.inc",
-                         true,
-                         true);
-    loop_value->addIncoming(loop_next, b_->GetInsertBlock());
-
-    Br(for_begin);
-    b_->SetInsertPoint(for_end);
-
-    return nullptr;
-    // llvm::AllocaInst *loop_var = Alloca(b_->getInt32Ty(), nullptr,
-    // op->loop_var->name); loop_var->setAlignment(llvm::Align(4));
-    // SetVar(op->loop_var->name, loop_var);
-  } while (false);
-
-  ////////////////////////////////////
   llvm::BasicBlock *preheader_bb = b_->GetInsertBlock();
   llvm::BasicBlock *exit_bb = nullptr;
 
@@ -814,20 +767,13 @@ llvm::Value *CodeGenLLVM::Visit(const ir::_Module_ *op) {
 }
 
 llvm::Value *CodeGenLLVM::Visit(const ir::_Var_ *op) {
-  llvm::Value *value = GetVar(op->name, false);
-  llvm::Value *result{};
-  CHECK(value) << "ir::_Var_[" << op->name << "]: value is null";
-  // TODO(fc500110) hard coding
-  if (LLVM_WillVarLowerAsPointer(op->name)) {
-    result = value;
-  } else if (value->getType()->isPointerTy() &&
-             !value->getType()->getPointerElementType()->isPointerTy()) {
-    result = Load(value, op->name + "_load");
-  } else {
-    result = value;
+  llvm::Value *value = GetVar(op->name, /* lazy= */ false);
+  // When visiting a Var that is allocated on the stack, we are actually
+  // reading its value instead of its address.
+  if (llvm::AllocaInst::classof(value)) {
+    return Load(value, op->name + "_load");
   }
-
-  return result;
+  return value;
 }
 
 void CodeGenLLVM::Scalarize(
@@ -1043,12 +989,6 @@ llvm::Value *CodeGenLLVM::Visit(const ir::_Buffer_ *op) {
 
 llvm::Value *CodeGenLLVM::Visit(const ir::_Tensor_ *op) {
   return GetVar(op->name);
-  auto *buffer_op = op->buffer.As<ir::_Buffer_>();
-  if (symbol_table_->Lookup(buffer_op->name)) {
-    return Visit(buffer_op);
-  }
-
-  return SetVar(buffer_op->name, Visit(buffer_op));
 }
 
 template <typename T,
@@ -1437,10 +1377,6 @@ void CodeGenLLVM::InitTarget(const Target &target) {
   naive_vec_alignment_ = GetNaiveVecAlignment(target);
 }
 
-bool LLVM_WillVarLowerAsPointer(const std::string &var_name) {
-  return var_name == "_args" || utils::EndsWith(var_name, "__ptr");
-}
-
 void CodeGenLLVM::AddTbaaMetadata(llvm::Instruction *inst,
                                   absl::string_view buffer,
                                   Expr index) {

diff --git a/paddle/cinn/backends/llvm/codegen_llvm.h b/paddle/cinn/backends/llvm/codegen_llvm.h
@@ -49,14 +49,6 @@ class LLVMIRVisitor : public ir::IRVisitorRequireReImpl<llvm::Value *> {
 #undef __m
 };
 
-/**
- * Tell whether a variable called \p \var_name will lowered to a pointer type in
- * LLVM.
- * @param var_name name of the variable.
- * @return a boolean.
- */
-bool LLVM_WillVarLowerAsPointer(const std::string &var_name);
-
 class SymbolTable {
  public:
   SymbolTable() = default;