fix(mgb): fix the compatibility issue of cuda stub with older version drivers

GitOrigin-RevId: 628afbf3cf

dnn/cuda-stub/src/dlopen_helper.h

@@ -111,7 +111,7 @@ static void* resolve_library_func(void* handle, const char* func) {
     }
     auto ret = dlsym(handle, func);
     if (!ret) {
-        LOGD("failed to load %s func: %s", g_default_api_name, func);
+        LOGD("failed to load %s func: %s.(It seems that the currently used version of the library containing the function is too old.)", g_default_api_name, func);
     }
     return ret;
 }

dnn/cuda-stub/src/libcuda-wrap_10.1.h

@@ -3829,6 +3829,7 @@ static void load_library() {
         return;
 
     void* handle = get_library_handle();
+    int num = 0;
     for (size_t i = 0; i < NR_FUNC; ++i) {
         void* func;
         if (!handle) {
@@ -3837,9 +3838,16 @@ static void load_library() {
             func = resolve_library_func(handle, g_func_name[i]);
         }
         if (!func) {
-            func = g_func_table_error[i];
+            break;
         }
+        num++;
         __atomic_store_n(g_func_table + i, func, __ATOMIC_RELAXED);
+    }
+	if (num != NR_FUNC) {
+        for (size_t i = 0; i < NR_FUNC; ++i) {
+            void* func = g_func_table_error[i];
+            __atomic_store_n(g_func_table + i, func, __ATOMIC_RELAXED);
+        }
     }
     done = true;
 }

dnn/cuda-stub/src/libcuda-wrap_10.2.h

@@ -3999,6 +3999,7 @@ static void load_library() {
         return;
 
     void* handle = get_library_handle();
+    int num = 0;
     for (size_t i = 0; i < NR_FUNC; ++i) {
         void* func;
         if (!handle) {
@@ -4007,9 +4008,16 @@ static void load_library() {
             func = resolve_library_func(handle, g_func_name[i]);
         }
         if (!func) {
-            func = g_func_table_error[i];
+            break;
         }
+        num++;
         __atomic_store_n(g_func_table + i, func, __ATOMIC_RELAXED);
+    }
+	if (num != NR_FUNC) {
+        for (size_t i = 0; i < NR_FUNC; ++i) {
+            void* func = g_func_table_error[i];
+            __atomic_store_n(g_func_table + i, func, __ATOMIC_RELAXED);
+        }
     }
     done = true;
 }

dnn/cuda-stub/src/libcuda-wrap_11.1.h

@@ -4369,6 +4369,7 @@ static void load_library() {
         return;
 
     void* handle = get_library_handle();
+    int num = 0;
     for (size_t i = 0; i < NR_FUNC; ++i) {
         void* func;
         if (!handle) {
@@ -4377,9 +4378,16 @@ static void load_library() {
             func = resolve_library_func(handle, g_func_name[i]);
         }
         if (!func) {
-            func = g_func_table_error[i];
+            break;
         }
+        num++;
         __atomic_store_n(g_func_table + i, func, __ATOMIC_RELAXED);
+    }
+	if (num != NR_FUNC) {
+        for (size_t i = 0; i < NR_FUNC; ++i) {
+            void* func = g_func_table_error[i];
+            __atomic_store_n(g_func_table + i, func, __ATOMIC_RELAXED);
+        }
     }
     done = true;
 }

dnn/cuda-stub/src/libcuda-wrap_11.2.h

@@ -4659,6 +4659,7 @@ static void load_library() {
         return;
 
     void* handle = get_library_handle();
+    int num = 0;
     for (size_t i = 0; i < NR_FUNC; ++i) {
         void* func;
         if (!handle) {
@@ -4667,9 +4668,16 @@ static void load_library() {
             func = resolve_library_func(handle, g_func_name[i]);
         }
         if (!func) {
-            func = g_func_table_error[i];
+            break;
         }
+        num++;
         __atomic_store_n(g_func_table + i, func, __ATOMIC_RELAXED);
+    }
+	if (num != NR_FUNC) {
+        for (size_t i = 0; i < NR_FUNC; ++i) {
+            void* func = g_func_table_error[i];
+            __atomic_store_n(g_func_table + i, func, __ATOMIC_RELAXED);
+        }
     }
     done = true;
 }

dnn/cuda-stub/src/libcuda-wrap_11.4.h

@@ -4899,6 +4899,7 @@ static void load_library() {
         return;
 
     void* handle = get_library_handle();
+    int num = 0;
     for (size_t i = 0; i < NR_FUNC; ++i) {
         void* func;
         if (!handle) {
@@ -4907,9 +4908,16 @@ static void load_library() {
             func = resolve_library_func(handle, g_func_name[i]);
         }
         if (!func) {
-            func = g_func_table_error[i];
+            break;
         }
+        num++;
         __atomic_store_n(g_func_table + i, func, __ATOMIC_RELAXED);
+    }
+	if (num != NR_FUNC) {
+        for (size_t i = 0; i < NR_FUNC; ++i) {
+            void* func = g_func_table_error[i];
+            __atomic_store_n(g_func_table + i, func, __ATOMIC_RELAXED);
+        }
     }
     done = true;
 }

dnn/cuda-stub/src/libcuda.cpp

@@ -3,8 +3,8 @@
 #pragma GCC visibility push(default)
 
 #include <cstdio>
-#define LOGI(fmt, v...) fprintf(stdout, "info: " fmt "\n", ##v)
-#define LOGD(fmt, v...) fprintf(stdout, "debug: " fmt "\n", ##v)
+#define LOGI(fmt, v...) fprintf(stderr, "info: " fmt "\n", ##v)
+#define LOGD(fmt, v...) fprintf(stderr, "debug: " fmt "\n", ##v)
 
 extern "C" {
 #include "cuda.h"