Fix

Author: co63oc

.github/CODEOWNERS

@@ -1,7 +1,7 @@
 # This file is migrated from CI script, it's an effort of modenizing our dev infra.
 # Code owners are expected to take responsibility for review patches to respective file.
 
-/CMakeLists.txt @wanghuancoder @Aurelius84 @XiaoguangHu01
+/CMakeLists.txt @wanghuancoder @XiaoguangHu01
 paddle/fluid/distributed/collective @sneaxiy @ForFishes
 paddle/fluid/eager/autograd_meta.cc @JiabinYang @phlrain
 paddle/fluid/eager/autograd_meta.h @JiabinYang @phlrain
@@ -12,20 +12,20 @@ paddle/fluid/eager/grad_node_info.h @JiabinYang @phlrain
 paddle/fluid/eager/grad_tensor_holder.cc @JiabinYang @phlrain
 paddle/fluid/eager/grad_tensor_holder.h @JiabinYang @phlrain
 paddle/fluid/eager/tensor_wrapper.h @JiabinYang @phlrain
-paddle/fluid/framework/block_desc.h @XiaoguangHu01 @zhiqiu @Xreki @Aurelius84
-paddle/fluid/framework/details/op_registry.h @XiaoguangHu01 @zhiqiu @Xreki @Aurelius84
-paddle/fluid/framework/framework.proto @XiaoguangHu01 @zhiqiu @Xreki @Aurelius84
-paddle/fluid/framework/grad_op_desc_maker.h @XiaoguangHu01 @zhiqiu @Xreki @Aurelius84
-paddle/fluid/framework/ir/graph.h @XiaoguangHu01 @zhiqiu @Xreki @Aurelius84
-paddle/fluid/framework/ir/node.h @XiaoguangHu01 @zhiqiu @Xreki @Aurelius84
-paddle/fluid/framework/lod_tensor.h @XiaoguangHu01 @zhiqiu @Xreki @Aurelius84
-paddle/fluid/framework/op_desc.h @XiaoguangHu01 @zhiqiu @Xreki @Aurelius84
-paddle/fluid/framework/operator.h @XiaoguangHu01 @zhiqiu @Xreki @Aurelius84
-paddle/fluid/framework/scope.h @XiaoguangHu01 @zhiqiu @Xreki @Aurelius84
-paddle/fluid/framework/selected_rows.h @XiaoguangHu01 @zhiqiu @Xreki @Aurelius84
-paddle/fluid/framework/tensor.h @XiaoguangHu01 @zhiqiu @Xreki @Aurelius84
+paddle/fluid/framework/block_desc.h @XiaoguangHu01 @zhiqiu @Xreki
+paddle/fluid/framework/details/op_registry.h @XiaoguangHu01 @zhiqiu @Xreki
+paddle/fluid/framework/framework.proto @XiaoguangHu01 @zhiqiu @Xreki
+paddle/fluid/framework/grad_op_desc_maker.h @XiaoguangHu01 @zhiqiu @Xreki
+paddle/fluid/framework/ir/graph.h @XiaoguangHu01 @zhiqiu @Xreki
+paddle/fluid/framework/ir/node.h @XiaoguangHu01 @zhiqiu @Xreki
+paddle/fluid/framework/lod_tensor.h @XiaoguangHu01 @zhiqiu @Xreki
+paddle/fluid/framework/op_desc.h @XiaoguangHu01 @zhiqiu @Xreki
+paddle/fluid/framework/operator.h @XiaoguangHu01 @zhiqiu @Xreki
+paddle/fluid/framework/scope.h @XiaoguangHu01 @zhiqiu @Xreki
+paddle/fluid/framework/selected_rows.h @XiaoguangHu01 @zhiqiu @Xreki
+paddle/fluid/framework/tensor.h @XiaoguangHu01 @zhiqiu @Xreki
 paddle/fluid/framework/unused_var_check.cc @zhiqiu @phlrain
-paddle/fluid/framework/var_desc.h @XiaoguangHu01 @zhiqiu @Xreki @Aurelius84
+paddle/fluid/framework/var_desc.h @XiaoguangHu01 @zhiqiu @Xreki
 paddle/fluid/operators/distributed/send_recv.proto.in @gongweibao @seiriosPlus
 paddle/fluid/prim/api/api.yaml @xiaoguoguo626807 @JiabinYang @phlrain
 paddle/fluid/prim/api/composite_backward/composite_backward_api.h @xiaoguoguo626807 @JiabinYang
@@ -44,33 +44,33 @@ paddle/phi/core/meta_tensor.h @phlrain @zyfncg @YuanRisheng
 paddle/phi/core/tensor_base.h @phlrain @zyfncg @YuanRisheng
 paddle/phi/core/tensor_meta.h @phlrain @zyfncg @YuanRisheng
 paddle/phi/infermeta/spmd_rules @LiYuRio @ForFishes @zhiqiu
-paddle/scripts/paddle_build.bat @zhwesky2010 @wanghuancoder @Aurelius84
+paddle/scripts/paddle_build.bat @zhwesky2010 @wanghuancoder
 paddle/scripts/paddle_build.sh @risemeup1 @zhangbo9674 @XieYunshen
 pyproject.toml @SigureMo @gouzil
-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/autograd/backward_utils.py @xiaoguoguo626807 @changeyoung98 @phlrain
+python/paddle/autograd/ir_backward.py @xiaoguoguo626807 @changeyoung98
+python/paddle/base/backward.py @XiaoguangHu01 @zhiqiu @Xreki @qili93
+python/paddle/base/compiler.py @XiaoguangHu01 @zhiqiu @Xreki @qili93
 python/paddle/base/dygraph/layers.py @JiabinYang @phlrain
-python/paddle/base/framework.py @XiaoguangHu01 @zhiqiu @Xreki @qili93 @Aurelius84
-python/paddle/base/__init__.py @phlrain @Aurelius84 @qili93
-python/paddle/base/tests/unittests/white_list/check_op_sequence_batch_1_input_white_list.py @Aurelius84 @phlrain
-python/paddle/base/tests/unittests/white_list/check_op_sequence_instance_0_input_white_list.py @Aurelius84 @phlrain
-python/paddle/base/tests/unittests/white_list/check_shape_white_list.py @hong19860320 @Aurelius84 @phlrain
-python/paddle/base/tests/unittests/white_list/compile_vs_runtime_white_list.py @Aurelius84 @phlrain
-python/paddle/base/tests/unittests/white_list/no_check_set_white_list.py @Aurelius84 @phlrain
-python/paddle/base/tests/unittests/white_list/no_grad_set_white_list.py @Aurelius84 @phlrain
-python/paddle/base/tests/unittests/white_list/op_accuracy_white_list.py @juncaipeng @zhangting2020 @Aurelius84
-python/paddle/base/tests/unittests/white_list/op_threshold_white_list.py @juncaipeng @zhangting2020 @Aurelius84
+python/paddle/base/framework.py @XiaoguangHu01 @zhiqiu @Xreki @qili93
+python/paddle/base/__init__.py @phlrain @qili93
+test/white_list/check_op_sequence_batch_1_input_white_list.py @phlrain
+test/white_list/check_op_sequence_instance_0_input_white_list.py @phlrain
+test/white_list/check_shape_white_list.py @hong19860320 @phlrain
+test/white_list/compile_vs_runtime_white_list.py @phlrain
+test/white_list/no_check_set_white_list.py @phlrain
+test/white_list/no_grad_set_white_list.py @phlrain
+test/white_list/op_accuracy_white_list.py @juncaipeng @zhangting2020
+test/white_list/op_threshold_white_list.py @juncaipeng @zhangting2020
 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 @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
-test/sot @SigureMo @Aurelius84 @gouzil
-tools/parallel_UT_rule.py @zhwesky2010 @wanghuancoder @Aurelius84
-tools/windows/run_unittests.sh @zhwesky2010 @wanghuancoder @Aurelius84
+test/dygraph_to_static @SigureMo @zrr1999 @gouzil
+test/sot @SigureMo @zrr1999 @gouzil
+tools/parallel_UT_rule.py @zhwesky2010 @wanghuancoder
+tools/windows/run_unittests.sh @zhwesky2010 @wanghuancoder
 .pre-commit-config.yaml @SigureMo @gouzil

cmake/external/xpu.cmake

@@ -25,12 +25,13 @@ set(XPU_XFT_LIB_NAME "libxft.so")
 set(XPU_XPTI_LIB_NAME "libxpti.so")
 set(XPU_XBLAS_LIB_NAME "libxpu_blas.so")
 set(XPU_XFA_LIB_NAME "libxpu_flash_attention.so")
+set(XPU_XPUDNN_LIB_NAME "libxpu_dnn.so")
 
 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 "eb35/20240927")
+  set(XPU_XHPC_BASE_DATE "eb35/20241015")
 endif()
 set(XPU_XCCL_BASE_VERSION "1.2.11e")
 if(NOT DEFINED XPU_XFT_BASE_VERSION)
@@ -139,6 +140,7 @@ set(XPU_XBLAS_LIB "${XPU_LIB_DIR}/${XPU_XBLAS_LIB_NAME}")
 set(XPU_RT_LIB "${XPU_LIB_DIR}/${XPU_RT_LIB_NAME}")
 set(XPU_CUDA_LIB "${XPU_LIB_DIR}/${XPU_CUDA_LIB_NAME}")
 set(XPU_XFA_LIB "${XPU_LIB_DIR}/${XPU_XFA_LIB_NAME}")
+set(XPU_XPUDNN_LIB "${XPU_LIB_DIR}/${XPU_XPUDNN_LIB_NAME}")
 
 set(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${XPU_INSTALL_DIR}/lib")
 
@@ -175,7 +177,9 @@ ExternalProject_Add(
   UPDATE_COMMAND ""
   CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${XPU_INSTALL_ROOT}
   CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${XPU_INSTALL_ROOT}
-  BUILD_BYPRODUCTS ${XPU_API_LIB} BUILD_BYPORDUCTS ${XPU_XBLAS_LIB}
+  BUILD_BYPRODUCTS ${XPU_API_LIB}
+  BUILD_BYPRODUCTS ${XPU_XBLAS_LIB}
+  BUILD_BYPRODUCTS ${XPU_XPUDNN_LIB}
   BUILD_BYPRODUCTS ${XPU_XFA_LIB}
   BUILD_BYPRODUCTS ${XPU_RT_LIB}
   BUILD_BYPRODUCTS ${XPU_BKCL_LIB})
@@ -203,6 +207,8 @@ set(XPU_XHPC_INC_DIR "${XPU_INC_DIR}/xhpc")
 include_directories(${XPU_XHPC_INC_DIR})
 set(XPU_XBLAS_INC_DIR "${XPU_INC_DIR}/xhpc/xblas")
 include_directories(${XPU_XBLAS_INC_DIR})
+set(XPU_XPUDNN_INC_DIR "${XPU_INC_DIR}/xhpc/xpudnn")
+include_directories(${XPU_XPUDNN_INC_DIR})
 
 if(WITH_XPU_XRE5)
   add_definitions(-DPADDLE_WITH_XPU_XRE5)
@@ -227,8 +233,14 @@ if(WITH_XPTI)
 endif()
 
 if(WITH_XPU_XRE5)
-  target_link_libraries(xpulib ${XPU_RT_LIB} ${XPU_BKCL_LIB} ${XPU_XBLAS_LIB}
-                        ${XPU_API_LIB} ${XPU_XFA_LIB})
+  target_link_libraries(
+    xpulib
+    ${XPU_RT_LIB}
+    ${XPU_BKCL_LIB}
+    ${XPU_XBLAS_LIB}
+    ${XPU_API_LIB}
+    ${XPU_XFA_LIB}
+    ${XPU_XPUDNN_LIB})
 else()
   target_link_libraries(xpulib ${XPU_RT_LIB} ${XPU_BKCL_LIB} ${XPU_XBLAS_LIB}
                         ${XPU_API_LIB})

cmake/inference_lib.cmake

@@ -304,7 +304,7 @@ endif()
 
 copy(
   inference_lib_dist
-  SRCS ${CMAKE_BINARY_DIR}/paddle/fluid/framework/framework.pb.h
+  SRCS ${CMAKE_BINARY_DIR}/paddle/phi/core/framework/framework.pb.h
   DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/internal)
 copy(
   inference_lib_dist

paddle/cinn/adt/equation_solver.cc

@@ -37,8 +37,8 @@ std::unordered_map<Variable, Value> InferValuesImpl(
   PADDLE_ENFORCE_EQ(
       ctx->HasValue(in_variable),
       true,
-      phi::errors::NotFound("The param id's out_iter must contain "
-                            "its in_iter's value"));
+      ::common::errors::NotFound("The param id's out_iter must contain "
+                                 "its in_iter's value"));
   return {{out_iter.value(), ctx->GetValue(in_variable)}};
 }
 
@@ -49,8 +49,8 @@ std::unordered_map<Variable, Value> InferValuesImpl(
   PADDLE_ENFORCE_EQ(
       ctx->HasValue(in_variable),
       true,
-      phi::errors::NotFound("The param id's out_iter must contain "
-                            "its in_iter's value"));
+      ::common::errors::NotFound("The param id's out_iter must contain "
+                                 "its in_iter's value"));
   return {{out_index.value(), ctx->GetValue(in_variable)}};
 }
 
@@ -215,7 +215,7 @@ std::unordered_map<Variable, Value> InferValuesImpl(
     PADDLE_ENFORCE_EQ(
         ret.emplace(out_msg_in_indexes.value()->at(i), value).second,
         true,
-        phi::errors::AlreadyExists([&]() {
+        ::common::errors::AlreadyExists([&]() {
           std::ostringstream oss;
           oss << "Failed to insert the variable '"
               << "out_msg_in_indexes.value()->at(" << i
@@ -229,7 +229,7 @@ std::unordered_map<Variable, Value> InferValuesImpl(
     if (out_index.has_value()) {
       PADDLE_ENFORCE_EQ(ret.emplace(out_index.value(), value).second,
                         true,
-                        phi::errors::AlreadyExists([&]() {
+                        ::common::errors::AlreadyExists([&]() {
                           std::ostringstream oss;
                           oss << "Failed to insert the variable '"
                               << "out_index.value()"
@@ -306,7 +306,9 @@ void SolveEquations(
         tValueInferSuccess<bool> has_unique_value =
             MergeInferedValuesIntoCtx(function, ctx);
         PADDLE_ENFORCE_EQ(
-            has_unique_value.value(), true, phi::errors::InvalidArgument([&]() {
+            has_unique_value.value(),
+            true,
+            ::common::errors::InvalidArgument([&]() {
               std::ostringstream oss;
               oss << "Failed to merge inferred values into the context for "
                      "function '"

paddle/cinn/backends/codegen_device_util.cc

@@ -262,6 +262,24 @@ void detail::CollectBucketStrategyHostFunctionVisitor::ProcessLoweredFunc(
                      ir::CallType::Extern,
                      ir::FunctionRef(),
                      0);
+
+  // create memset calls for temp_spaces if needed
+  std::vector<ir::Expr> call_kernel_stmts;
+  for (auto &temp_space : func_node->temp_spaces) {
+    if (temp_space.need_zero_init()) {
+      ir::Expr size = common::cast(temp_space.size(), common::UInt(64));
+      ir::Expr call_get_arg =
+          lang::CallExtern(runtime::intrinsic::get_item_in_cuda_kernel_args,
+                           {kernel_args_, ir::Expr(temp_space.arg_idx())});
+      ir::Expr call_memset = lang::CallExtern(
+          runtime::intrinsic::call_cuda_memset,
+          {call_get_arg, ir::Expr(1), ir::Expr(0), size, kernel_stream_});
+      call_kernel_stmts.push_back(call_memset);
+    }
+  }
+  call_kernel_stmts.push_back(call_extern_api);
+  call_extern_api = ir::Block::Make(call_kernel_stmts);
+
   if (buckets_.empty()) {
     buckets_.emplace_back(ir::IfThenElse::Make(predicate, call_extern_api));
   } else {
@@ -270,6 +288,26 @@ void detail::CollectBucketStrategyHostFunctionVisitor::ProcessLoweredFunc(
     buckets_.emplace_back(
         ir::IfThenElse::Make(predicate, call_extern_api, false_expr));
   }
+
+  // create infer shape calls for temp_spaces
+  std::vector<ir::Expr> temp_space_infer_shape_stmts;
+  for (int i = 0; i < func_node->temp_spaces.size(); ++i) {
+    ir::Var tensor_shape_args(TENSOR_SHAPE_ARGS, type_of<int64_t **>());
+    ir::Expr size =
+        common::cast(func_node->temp_spaces[i].size(), common::Int(64));
+    ir::Expr call_set_value =
+        lang::CallExtern(runtime::intrinsic::infer_shape_set_value,
+                         {ir::Expr(func_node->num_output_tensors + i),
+                          ir::Expr(0),
+                          size,
+                          tensor_shape_args});
+    temp_space_infer_shape_stmts.push_back(call_set_value);
+  }
+  if (!temp_space_infer_shape_stmts.empty()) {
+    ir::Expr if_body = ir::Block::Make(temp_space_infer_shape_stmts);
+    temp_space_infer_shape_body_ =
+        ir::IfThenElse::Make(predicate, if_body, temp_space_infer_shape_body_);
+  }
 }
 
 void detail::CollectBucketStrategyHostFunctionVisitor::ProcessArgs(

paddle/cinn/backends/codegen_device_util.h

@@ -280,6 +280,9 @@ struct CollectBucketStrategyHostFunctionVisitor
     infer_shape_func_body_stmts.insert(
         infer_shape_func_body_stmts.end(),
         op->infer_shape_func.as_lowered_func()->body);
+    if (temp_space_infer_shape_body_.defined()) {
+      infer_shape_func_body_stmts.push_back(temp_space_infer_shape_body_);
+    }
 
     std::vector<ir::Argument> infer_shape_arguments = {
         ir::Argument(kernel_args_, ir::Argument::IO::kOutput),
@@ -307,6 +310,7 @@ struct CollectBucketStrategyHostFunctionVisitor
  private:
   std::vector<ir::Expr> buckets_;
   std::vector<ir::Expr> arg_defs_;
+  ir::Expr temp_space_infer_shape_body_;
 
   ir::Var kernel_args_;
   ir::Var kernel_args_num_;

paddle/cinn/common/const_fold.h

@@ -71,5 +71,41 @@ inline std::optional<ir::Expr> TryConstFold<ir::Mul>(ir::Expr a, ir::Expr b) {
   return std::nullopt;
 }
 
+template <>
+inline std::optional<ir::Expr> TryConstFold<ir::Div>(ir::Expr a, ir::Expr b) {
+  const ir::IntImm* pa = a.As<ir::IntImm>();
+  const ir::IntImm* pb = b.As<ir::IntImm>();
+  const auto& rtype = a.type();
+  if (pa && pb) {
+    int64_t res = pa->value / pb->value;
+    return cinn::common::make_shared<ir::IntImm>(rtype, res);
+  }
+  if (pa) {
+    if (pa->value == 0) return a;
+  }
+  if (pb) {
+    if (pb->value == 1) return a;
+  }
+  return std::nullopt;
+}
+
+template <>
+inline std::optional<ir::Expr> TryConstFold<ir::Mod>(ir::Expr a, ir::Expr b) {
+  const ir::IntImm* pa = a.As<ir::IntImm>();
+  const ir::IntImm* pb = b.As<ir::IntImm>();
+  const auto& rtype = a.type();
+  if (pa && pb) {
+    int64_t res = pa->value % pb->value;
+    return cinn::common::make_shared<ir::IntImm>(rtype, res);
+  }
+  if (pa) {
+    if (pa->value == 0) return a;
+  }
+  if (pb) {
+    if (pb->value == 1) return ir::Zero(rtype);
+  }
+  return std::nullopt;
+}
+
 }  // namespace common
 }  // namespace cinn

paddle/cinn/common/dim_expr_converter.cc

@@ -153,7 +153,8 @@ struct DimExprConverterWithSymbolBindings::
       return inputs_[input_idx]->sym_shape[input_dim_idx]->GetDimExpr();
     }
     // for data binding [S0, a, b], inputs[a] is Tensor A, return A(b)
-    return inputs_[input_idx](cinn::ir::Expr(input_dim_idx));
+    return ir::Cast::Make(cinn::common::I64(),
+                          inputs_[input_idx](cinn::ir::Expr(input_dim_idx)));
   }
 
   DimExprToIrExprVisitorWithSymbolBinding(