Merge address sanitizer build to 6.3

Author: malcolmroberts

CHANGELOG.md

@@ -28,6 +28,9 @@ Documentation for rocFFT is available at
 
 * Compilation uses amdclang++ instead of hipcc.
 * CLI11 replaces Boost Program Options as the command line parser for clients and samples.
+* Building with the address sanitizer option sets xnack+ on relevant GPU
+  architectures and address-sanitizer support is added to runtime-compiled
+  kernels.
 
 ## rocFFT 1.0.30 for ROCm 6.2.4
 

CMakeLists.txt

@@ -116,12 +116,6 @@ option(ROCFFT_RUNTIME_COMPILE_DEFAULT "Compile kernels at runtime by default" OF
 # Set default to OFF since users are not likely to tune
 option(ROCFFT_BUILD_OFFLINE_TUNER "Build with offline tuner executable rocfft_offline_tuner" OFF)
 
-if(BUILD_ADDRESS_SANITIZER)
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=address -shared-libasan")
-  set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fsanitize=address -shared-libasan")
-  add_link_options(-fuse-ld=lld)
-endif()
-
 # FOR HANDLING ENABLE/DISABLE OPTIONAL BACKWARD COMPATIBILITY for FILE/FOLDER REORG
 option(BUILD_FILE_REORG_BACKWARD_COMPATIBILITY "Build with file/folder reorg with backward compatibility enabled" OFF)
 if(BUILD_FILE_REORG_BACKWARD_COMPATIBILITY AND NOT WIN32)
@@ -138,6 +132,52 @@ if( WERROR )
   set( WARNING_FLAGS ${WARNING_FLAGS} -Werror )
 endif( )
 
+set(DEFAULT_GPUS
+  gfx803
+  gfx900
+  gfx906
+  gfx908
+  gfx90a
+  gfx940
+  gfx941
+  gfx942
+  gfx1030
+  gfx1100
+  gfx1101
+  gfx1102
+  gfx1151
+  gfx1200
+  gfx1201)
+
+if(BUILD_ADDRESS_SANITIZER)
+  add_compile_options(-fsanitize=address)
+  add_link_options(-fsanitize=address)
+  add_link_options(-shared-libasan)
+  SET(DEFAULT_GPUS
+    gfx908:xnack+
+    gfx90a:xnack+
+    gfx940:xnack+
+    gfx941:xnack+
+    gfx942:xnack+)
+  add_link_options(-fuse-ld=lld)
+  set(ROCFFT_KERNEL_CACHE_ENABLE off)
+  add_compile_definitions(ADDRESS_SANITIZER)
+endif()
+
+# Build only for local GPU architecture
+if (BUILD_LOCAL_GPU_TARGET_ONLY)
+  message(STATUS "Building only for local GPU target")
+  if (COMMAND rocm_local_targets)
+    rocm_local_targets(DEFAULT_GPUS)
+  else()
+    message(WARNING "Unable to determine local GPU targets. Falling back to default GPUs.")
+  endif()
+endif()
+
+
+set(AMDGPU_TARGETS "${DEFAULT_GPUS}" CACHE STRING "Target default GPUs if AMDGPU_TARGETS is not defined.")
+rocm_check_target_ids(AMDGPU_TARGETS TARGETS "${AMDGPU_TARGETS}")
+
 # Use target ID syntax if supported for AMDGPU_TARGETS
 rocm_check_target_ids(DEFAULT_AMDGPU_TARGETS
   TARGETS "gfx803;gfx900;gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201")

library/src/rtc_cache.cpp

@@ -565,6 +565,11 @@ std::vector<char> RTCCache::cached_compile(const std::string&          kernel_na
                                            kernel_src_gen_t            generate_src,
                                            const std::array<char, 32>& generator_sum)
 {
+#ifdef ADDRESS_SANITIZER
+    // The address sanitizer is reported to work better when we include xnack+, so don't strip this
+    // from the architecture string when building:
+    const std::string gpu_arch = gpu_arch_with_flags;
+#else
     // Supplied gpu arch may have extra flags on it
     // (e.g. gfx90a:sramecc+:xnack-), Strip those from the arch name
     // since omitting them will generate code that handles either
@@ -573,7 +578,8 @@ std::vector<char> RTCCache::cached_compile(const std::string&          kernel_na
     // As of this writing, there are no known performance benefits to
     // including the flags.  If that changes, we may need to be more
     // selective about which flags to strip.
-    std::string gpu_arch = gpu_arch_strip_flags(gpu_arch_with_flags);
+    const std::string gpu_arch = gpu_arch_strip_flags(gpu_arch_with_flags);
+#endif
 
     std::shared_future<std::vector<char>> result;
 

library/src/rtc_compile.cpp

@@ -36,11 +36,13 @@ std::vector<char> compile_inprocess(const std::string& kernel_src, const std::st
     std::string gpu_arch_arg = "--gpu-architecture=" + gpu_arch;
 
     std::vector<const char*> options;
-    options.reserve(4);
     options.push_back("-O3");
     options.push_back("-std=c++14");
     options.push_back(gpu_arch_arg.c_str());
     options.push_back("-mcumode");
+#ifdef ADDRESS_SANITIZER
+    options.push_back("-fsanitize=address");
+#endif
 
     auto compileResult = hiprtcCompileProgram(prog, options.size(), options.data());
     if(compileResult != HIPRTC_SUCCESS)