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

Author: zhwesky2010

.github/workflows/_Linux-XPU.yml

@@ -206,7 +206,7 @@ jobs:
           CCACHE_DIR: /root/.ccache
           CCACHE_MAXSIZE: 150G
           CCACHE_LIMIT_MULTIPLE: 0.8
-          IF_KUNLUN3: "OFF"
+          IF_KUNLUN3: "ON"
           GITHUB_API_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           home_dir: ${{ github.workspace }}/../../../..
           FLAGS_use_stride_kernel: "0"

.github/workflows/_Slice.yml

@@ -53,6 +53,7 @@ jobs:
 
   slice:
     name: Slice test
+    needs: check-bypass
     if: ${{ inputs.can-skip != 'true' && needs.check-bypass.outputs.can-skip != 'true' }}
     runs-on:
       group: slice

ci/kunlun_test.sh

@@ -159,9 +159,10 @@ set +x
             git clone --depth 1000 https://gitee.com/paddlepaddle/PaddleX.git
             cd PaddleX
             pip install -e .
+            pip install numpy==1.24.4 pypdfium2
 
             #install paddle x dependency
-            paddlex --install PaddleClas
+            paddlex --install PaddleClas -y
 
             #download paddle dataset
             wget -q https://paddle-model-ecology.bj.bcebos.com/paddlex/data/cls_flowers_examples.tar -P ./dataset

cmake/cinn.cmake

@@ -26,12 +26,21 @@ else()
 endif()
 
 if(NOT DEFINED ENV{runtime_include_dir})
-  message(
-    STATUS
-      "set runtime_include_dir: ${CMAKE_SOURCE_DIR}/paddle/cinn/runtime/cuda")
-  set(ENV{runtime_include_dir} "${CMAKE_SOURCE_DIR}/paddle/cinn/runtime/cuda")
-  add_definitions(
-    -DRUNTIME_INCLUDE_DIR="${CMAKE_SOURCE_DIR}/paddle/cinn/runtime/cuda")
+  if(WITH_GPU)
+    message(
+      STATUS
+        "set runtime_include_dir: ${CMAKE_SOURCE_DIR}/paddle/cinn/runtime/cuda")
+    set(ENV{runtime_include_dir} "${CMAKE_SOURCE_DIR}/paddle/cinn/runtime/cuda")
+    add_definitions(
+      -DRUNTIME_INCLUDE_DIR="${CMAKE_SOURCE_DIR}/paddle/cinn/runtime/cuda")
+  elseif(WITH_ROCM)
+    message(
+      STATUS
+        "set runtime_include_dir: ${CMAKE_SOURCE_DIR}/paddle/cinn/runtime/hip")
+    set(ENV{runtime_include_dir} "${CMAKE_SOURCE_DIR}/paddle/cinn/runtime/hip")
+    add_definitions(
+      -DRUNTIME_INCLUDE_DIR="${CMAKE_SOURCE_DIR}/paddle/cinn/runtime/hip")
+  endif()
 endif()
 
 if(WITH_TESTING)
@@ -118,6 +127,10 @@ if(WITH_ROCM)
     add_definitions(-DCINN_WITH_HIP)
   endif()
   link_libraries(${ROCM_HIPRTC_LIB})
+
+  message(
+    STATUS "copy paddle/cinn/common/float16.h to $ENV{runtime_include_dir}")
+  file(COPY paddle/cinn/common/float16.h DESTINATION $ENV{runtime_include_dir})
 endif()
 
 set(cinnapi_src CACHE INTERNAL "" FORCE)

doc/README_cn.md

@@ -0,0 +1,6 @@
+# 致读者和开发者
+感谢您阅读 PaddlePaddle 文档。
+
+自 **2018年9月17日** 起，**0.15.0 及 develop** 分支的文档源码已迁移至 [FluidDoc Repo](https://github.com/PaddlePaddle/FluidDoc) 仓库 ，并将在该仓库中持续更新。
+
+请前往 FluidDoc 仓库获取最新文档。

paddle/cinn/backends/codegen_device_util.cc

@@ -44,13 +44,12 @@ ir::Module CreateSwitchWithBroadcastConditionModule(
       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)};
-
   const auto &symbolic_arg_define = [&]() -> std::vector<ir::Expr> {
     std::vector<ir::Expr> arg_defs;
     for (const auto &item : symbolic_shape_var_index) {
       ir::Expr call_get_value_in_kernel_args =
           ir::Call::Make(Int(64),
-                         runtime::intrinsic::get_value_in_cuda_kernel_args,
+                         runtime::intrinsic::get_value_in_kernel_args,
                          {kernel_args, ir::Expr(item.first)},
                          {},
                          ir::CallType::Extern,
@@ -384,7 +383,7 @@ void detail::CollectBucketStrategyHostFunctionVisitor::ProcessArgs(
     if (args[i].is_var()) {
       ir::Expr call_get_value_in_kernel_args =
           ir::Call::Make(Int(64),
-                         runtime::intrinsic::get_value_in_cuda_kernel_args,
+                         runtime::intrinsic::get_value_in_kernel_args,
                          {kernel_args_, ir::Expr(i)},
                          {},
                          ir::CallType::Extern,

paddle/cinn/backends/codegen_invoke_module.h

@@ -68,7 +68,7 @@ class CodeGenSwitchHost : public CodeGenInvokeModule {
       : CodeGenInvokeModule(m, b, vars) {}
   // 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) {
+    if (op->name == runtime::intrinsic::get_value_in_kernel_args) {
       return CodeGenLLVM::Visit(op);
     } else {
       return LowerInnerCaseCall(op);

paddle/cinn/hlir/dialect/operator/transforms/pd_to_cinn_pass.cc

@@ -742,9 +742,7 @@ class SplitOpPattern : public pir::OpRewritePattern<paddle::dialect::SplitOp> {
   using pir::OpRewritePattern<paddle::dialect::SplitOp>::OpRewritePattern;
 
   bool Match(paddle::dialect::SplitOp op) const override {
-    const bool is_denied = CompatibleInfo::IsDeniedForCinn(*op.operation());
-
-    return !is_denied && PatternConstraint(op);
+    return PatternConstraint(op);
   }
 
   void Rewrite(paddle::dialect::SplitOp op,

paddle/cinn/runtime/cpu/host_intrinsics.cc

@@ -280,6 +280,12 @@ inline int64_t FN_INT64(logical_right_shift)(int64_t x, int64_t y) {
 }
 
 #undef FN_INT64
+
+int64_t cinn_get_value_in_kernel_args(void* v_args, int idx) {
+  cinn_pod_value_t* args = static_cast<cinn_pod_value_t*>(v_args);
+  return args[idx].operator int64_t();
+}
+
 }  // extern "C"
 
 CINN_REGISTER_HELPER(host_intrinsics) {
@@ -469,5 +475,11 @@ CINN_REGISTER_HELPER(host_intrinsics) {
       .AddInputType<int>()
       .End();
 
+  REGISTER_EXTERN_FUNC_HELPER(cinn_get_value_in_kernel_args, host_target)
+      .SetRetType<int64_t>()
+      .AddInputType<void*>()  // args
+      .AddInputType<int>()    // index
+      .End();
+
   return true;
 }

paddle/cinn/runtime/cpu/host_intrinsics.h

@@ -121,4 +121,6 @@ inline float FN_FP32(cbrt)(float x);
 inline double FN_FP64(cbrt)(double x);
 
 #undef FN_FP64
+
+int64_t cinn_get_value_in_kernel_args(void* v_args, int idx);
 }