[DeviceSanitizer] Only allows to report one error per each kernel launch (#15326)

When error is detected, the corresponding work-item will exit without
unpoisoning local/private shadow memory. Then other work-items are easy
to cause false alarms. So, we only allows to report one error per each
kernel launch by default.

UR Part: oneapi-src/unified-runtime#2069

---------

Co-authored-by: Piotr Balcer <[email protected]>

Author: zhaomaosu

libdevice/CMakeLists.txt

@@ -8,6 +8,17 @@ set(CMAKE_MODULE_PATH
   "${CMAKE_CURRENT_SOURCE_DIR}/cmake/modules"
   )
 
+if(EXISTS ${FETCHCONTENT_BASE_DIR}/unified-runtime-src)
+  set(UR_SANITIZER_INCLUDE_DIR
+      ${FETCHCONTENT_BASE_DIR}/unified-runtime-src/source/loader/layers/sanitizer)
+elseif(EXISTS ${SYCL_UR_SOURCE_DIR})
+  set(UR_SANITIZER_INCLUDE_DIR
+      ${SYCL_UR_SOURCE_DIR}/source/loader/layers/sanitizer)
+else()
+  message(WARNING "Unified Runtime source directory is not provided, \
+                   will not build libdevice sanitizer")
+endif()
+
 # Build libdevice for SYCL.
 include(SYCLLibdevice)
 

libdevice/cmake/modules/SYCLLibdevice.cmake

@@ -200,10 +200,10 @@ set(cmath_obj_deps device_math.h device.h sycl-compiler)
 set(imf_obj_deps device_imf.hpp imf_half.hpp imf_bf16.hpp imf_rounding_op.hpp imf_impl_utils.hpp device.h sycl-compiler)
 set(itt_obj_deps device_itt.h spirv_vars.h device.h sycl-compiler)
 set(bfloat16_obj_deps sycl-headers sycl-compiler)
-if (NOT MSVC)
+if (NOT MSVC AND UR_SANITIZER_INCLUDE_DIR)
   set(sanitizer_obj_deps
     device.h atomic.hpp spirv_vars.h
-    include/asan_libdevice.hpp
+    ${UR_SANITIZER_INCLUDE_DIR}/asan_libdevice.hpp
     include/sanitizer_utils.hpp
     include/spir_global_var.hpp
     sycl-compiler)
@@ -268,10 +268,12 @@ if(MSVC)
     SRC msvc_math.cpp
     DEPENDENCIES ${cmath_obj_deps})
 else()
-  add_devicelibs(libsycl-sanitizer
-    SRC sanitizer_utils.cpp
-    DEPENDENCIES ${sanitizer_obj_deps}
-    EXTRA_OPTS -fno-sycl-instrument-device-code)
+  if(UR_SANITIZER_INCLUDE_DIR)
+    add_devicelibs(libsycl-sanitizer
+      SRC sanitizer_utils.cpp
+      DEPENDENCIES ${sanitizer_obj_deps}
+      EXTRA_OPTS -fno-sycl-instrument-device-code -I${UR_SANITIZER_INCLUDE_DIR})
+  endif()
 endif()
 
 add_devicelibs(libsycl-fallback-cassert

libdevice/include/asan_libdevice.hpp

@@ -6,11 +6,11 @@
 //
 //===----------------------------------------------------------------------===//
 
+#include "asan_libdevice.hpp"
 #include "atomic.hpp"
 #include "device.h"
 #include "spirv_vars.h"
 
-#include "include/asan_libdevice.hpp"
 #include "include/sanitizer_utils.hpp"
 
 using uptr = uintptr_t;
@@ -381,8 +381,11 @@ void __asan_internal_report_save(DeviceSanitizerErrorType error_type) {
   auto &SanitizerReport = ((__SYCL_GLOBAL__ LaunchInfo *)__AsanLaunchInfo)
                               ->SanitizerReport[WG_LID % ASAN_MAX_NUM_REPORTS];
 
-  if (atomicCompareAndSet(&SanitizerReport.Flag, Desired, Expected) ==
-      Expected) {
+  if (atomicCompareAndSet(
+          &(((__SYCL_GLOBAL__ LaunchInfo *)__AsanLaunchInfo)->ReportFlag), 1,
+          0) == 0 &&
+      atomicCompareAndSet(&SanitizerReport.Flag, Desired, Expected) ==
+          Expected) {
     SanitizerReport.ErrorType = error_type;
     SanitizerReport.IsRecover = false;
 
@@ -415,8 +418,12 @@ void __asan_internal_report_save(
   auto &SanitizerReport = ((__SYCL_GLOBAL__ LaunchInfo *)__AsanLaunchInfo)
                               ->SanitizerReport[WG_LID % ASAN_MAX_NUM_REPORTS];
 
-  if (atomicCompareAndSet(&SanitizerReport.Flag, Desired, Expected) ==
-      Expected) {
+  if ((is_recover ||
+       atomicCompareAndSet(
+           &(((__SYCL_GLOBAL__ LaunchInfo *)__AsanLaunchInfo)->ReportFlag), 1,
+           0) == 0) &&
+      atomicCompareAndSet(&SanitizerReport.Flag, Desired, Expected) ==
+          Expected) {
 
     int FileLength = 0;
     int FuncLength = 0;

libdevice/sanitizer_utils.cpp

@@ -1,7 +1,5 @@
-# commit d3a518a2a621da6d78f8d25e8460fd3300fff705
-# Merge: ab0a706b 77975b87
-# Author: aarongreig <[email protected]>
-# Date:   Fri Oct 25 16:45:53 2024 +0100
-#     Merge pull request #2219 from kbenzie/benie/cl-adapter-opt-core-func
-#     [CL] Support using old ICD loaders
-set(UNIFIED_RUNTIME_TAG d3a518a2a621da6d78f8d25e8460fd3300fff705)
+# commit 66ba7970a6badf781226b75a98c9585ef30ea93a
+# Author: Maosu Zhao <[email protected]>
+# Date:   Mon Oct 28 18:39:57 2024 +0800
+#     [DeviceSanitizer] Add a report flag to LaunchInfo (#2069)
+set(UNIFIED_RUNTIME_TAG 66ba7970a6badf781226b75a98c9585ef30ea93a)