Refactor acl_usm.cpp with new multi-threading routines

Author: ericxu233

include/acl_thread.h

@@ -14,6 +14,7 @@
 
 // System headers.
 #include <thread>
+#include <mutex>
 
 #if defined(__cplusplus)
 extern "C" {
@@ -35,7 +36,7 @@ extern ACL_TLS int acl_inside_sig_old_lock_count;
 
 extern aclMutex amtex;
 
-#define ALC_LOCK_GUARD std::lock_guard<aclMutex> l(amtex)
+using acl_lock_guard = std::lock_guard<aclMutex>;
 
 // -- signal handler functions --
 // When we enter a signal handler, we save "acl_global_lock_count" to

include/acl_util.h

@@ -115,6 +115,15 @@ void acl_dump_mem(cl_mem mem);
 #define acl_dump_mem(x)
 #endif
 /////////////////////
+#define RETURN(ret)                                                            \
+  do {                                                                         \
+  return (ret);                                                                \
+  } while (0)
+
+#define RETURN_VOID                                                            \
+  do {                                                                         \
+  return;                                                                      \
+} while (0)                                                                   
 
 #define UNLOCK_RETURN(ret)                                                     \
   do {                                                                         \
@@ -190,6 +199,26 @@ void acl_dump_mem(cl_mem mem);
     }                                                                          \
   } while (0)
 
+#define VALIDATE_ARRAY_OUT_ARGS(buf_size, buf, answer_size_out,         \
+                                       context)                                \
+  do {                                                                         \
+    if (buf && buf_size <= 0) {                                                \
+      acl_context_callback(context,                                            \
+                           #buf " is specified but " #buf_size " is zero");    \
+      RETURN(CL_INVALID_VALUE);                                         \
+    }                                                                          \
+    if (buf == 0 && buf_size > 0) {                                            \
+      acl_context_callback(context, #buf " is not specified but " #buf_size    \
+                                         " is positive");                      \
+      RETURN(CL_INVALID_VALUE);                                         \
+    }                                                                          \
+    if (answer_size_out == 0 && buf == 0) {                                    \
+      acl_context_callback(context,                                            \
+                           #buf " and " #answer_size_out " are both zero");    \
+      RETURN(CL_INVALID_VALUE);                                         \
+    }                                                                          \
+  } while (0)
+
 // On 32-bit ARM, pointers are 32-bit. 2GB value becomes a huge negative number
 // unless it's cast to a 64-bit integer. Need to cast to size_t first to avoid
 // warning of "casting pointer to integer of different size". Note that using