Zgc/diopi ascend fix python gil bug (DeepLink-org#762)

* delete python relate depandcy

* optimize the location of releasing python GIL lock

* add Python.h dir to include path

* rename gil_scoped_release to GilScopedRelease

---------

Co-authored-by: zhaoguochun1995 <zhaoguochun1995@sensetime.com>

Author: zhaoguochun1995

diopi_test/diopi_stub/codegen/gen.py

@@ -166,11 +166,7 @@ def gen_functions(options, functions_fm):
     with open(os.path.join(_cur_dir, options.get('source_dir'), 'functions.h'), 'r', encoding='utf8')as f:
         content = f.readlines()
     exports = []
-    device = options.get('device')
-    if device == 'ascend':
-        ft = OT.function_ascend_template
-    else:
-        ft = OT.function_template
+    ft = OT.function_template
     exports = get_export(content, ft, exports)
     with open(os.path.join(_cur_dir, options.get('source_dir'), 'functions_ext.h'), 'r', encoding='utf8')as f:
         content_ext = f.readlines()

diopi_test/diopi_stub/codegen/op_template.py

@@ -41,20 +41,6 @@ class OpTemplate(object):
         return diopiError_t::diopiNoImplement;
     }
 });
-""")
-
-    function_ascend_template = CodeTemplate("""\
-m.def("${func_name}", [](${attrs}) {
-    if (${func_name}) {
-        ${convert}
-        py::gil_scoped_release no_gil;
-        diopiError_t ret = ${call_func};
-        ${out_copy}
-        return ret;
-    } else {
-        return diopiError_t::diopiNoImplement;
-    }
-});
 """)
 
     vector_template = CodeTemplate("""\

impl/ascend/common/acloprunner.hpp

@@ -21,6 +21,7 @@
 #include "acl/acl_op.h"
 #include "acl/acl_op_compiler.h"
 #include "debug.hpp"
+#include "gil_scoped_release.hpp"
 #include "impl_functions.hpp"
 #include "utils.hpp"
 
@@ -619,6 +620,7 @@ class AclOpRunner {
     AclOpRunner& run() {
         diopiStreamHandle_t stream;
         diopiGetStream(context_, &stream);
+        diopi::GilScopedRelease gilReleaeGuard;
         if (sync_) {
             CALL_ACLRT(aclopCompileAndExecuteV2(opname_.data(),
                                                 inputIndex_,

impl/ascend/common/gil_scoped_release.hpp

@@ -0,0 +1,28 @@
+#pragma once
+
+#include "Python.h"
+
+namespace diopi {
+
+// https://docs.python.org/zh-cn/3/c-api/init.html?highlight=pygilstate_check
+
+class GilScopedRelease {
+private:
+    PyThreadState* state_ = nullptr;
+
+public:
+    GilScopedRelease() {
+        if (PyGILState_Check()) {
+            state_ = PyEval_SaveThread();
+        }
+    }
+
+    ~GilScopedRelease() {
+        if (state_ != nullptr) {
+            PyEval_RestoreThread(state_);
+            state_ = nullptr;
+        }
+    }
+};
+
+}  // namespace diopi

impl/ascend_npu/CMakeLists.txt

@@ -21,6 +21,14 @@ add_custom_target(op_plugin_gen
 add_subdirectory(third_party/acl)
 include_directories(SYSTEM ${CMAKE_CURRENT_SOURCE_DIR}/third_party/acl/inc)
 
+if(NOT DEFINED PYTHON_INCLUDE_DIR)
+execute_process(
+  COMMAND sh -c "dirname $(find  $(dirname $(which python))/../ -name Python.h)"
+  OUTPUT_VARIABLE PYTHON_INCLUDE_DIR)
+endif()
+include_directories(SYSTEM ${PYTHON_INCLUDE_DIR})
+message(STATUS "PYTHON_INCLUDE_DIR: " ${PYTHON_INCLUDE_DIR})
+
 if(NOT DEFINED PYTORCH_DIR)
   execute_process(
     COMMAND sh -c "dirname $(python -c 'import torch;print(torch.__path__[0])')"

impl/ascend_npu/torch_npu/csrc/DIOPIAdapter.cpp

@@ -6,6 +6,7 @@
 #include <diopi/diopirt.h>
 #include <torch/library.h>
 
+#include "../../../ascend/common/gil_scoped_release.hpp"
 #include "diopi_impl/helper.hpp"
 #include "op_plugin/AclOpsInterface.h"
 
@@ -2005,6 +2006,7 @@ class OpCommandImpl {
 void OpCommandImpl::Run(bool sync, c10::SmallVector<int64_t, N>& sync_index, c10::SmallVector<at::Tensor, N>& outputTensor) {
     NPU_LOGD("Op %s start run.", opName.c_str());
     // RECORD_FUNCTION(opName, std::vector<c10::IValue>({}));
+    diopi::GilScopedRelease gilReleaeGuard;
     ACL_REQUIRE_OK_OP(InnerRun(opName, execParam, sync, sync_index, outputTensor), opName.c_str());
     NPU_LOGD("Op %s run over.", opName.c_str());
 }