multiple processes tracing support
Этот коммит содержится в:
+70
-102
@@ -5,46 +5,59 @@
|
||||
#include <mutex>
|
||||
#include <dlfcn.h>
|
||||
|
||||
#define LOADER_INSTANTIATE() \
|
||||
std::atomic<roctracer::HipLoader*> roctracer::HipLoader::instance_{}; \
|
||||
std::atomic<roctracer::HccLoader*> roctracer::HccLoader::instance_{}; \
|
||||
std::atomic<roctracer::KfdLoader*> roctracer::KfdLoader::instance_{}; \
|
||||
std::atomic<roctracer::RocTxLoader*> roctracer::RocTxLoader::instance_{}; \
|
||||
roctracer::Loader::mutex_t roctracer::Loader::mutex_;
|
||||
|
||||
namespace roctracer {
|
||||
|
||||
// Base runtime loader class
|
||||
class Loader {
|
||||
template <class T>
|
||||
class BaseLoader : public T {
|
||||
public:
|
||||
typedef std::mutex mutex_t;
|
||||
typedef BaseLoader<T> loader_t;
|
||||
|
||||
template <class fun_t>
|
||||
fun_t* GetFun(const char* fun_name) { return (fun_t*) dlsym(handle_, fun_name); }
|
||||
|
||||
protected:
|
||||
Loader(const char* lib_name) {
|
||||
handle_ = dlopen(lib_name, RTLD_LAZY|RTLD_NOLOAD);
|
||||
if (handle_ == NULL) {
|
||||
fprintf(stderr, "roctracer: Loading '%s' failed, %s\n", lib_name, dlerror());
|
||||
abort();
|
||||
static inline loader_t& Instance(const bool& preload = false) {
|
||||
loader_t* obj = instance_.load(std::memory_order_acquire);
|
||||
if (obj == NULL) {
|
||||
std::lock_guard<mutex_t> lck(mutex_);
|
||||
if (instance_.load(std::memory_order_relaxed) == NULL) {
|
||||
obj = new loader_t(preload);
|
||||
instance_.store(obj, std::memory_order_release);
|
||||
}
|
||||
}
|
||||
return *instance_;
|
||||
}
|
||||
|
||||
~Loader() {
|
||||
static loader_t* GetRef() { return instance_; }
|
||||
|
||||
private:
|
||||
BaseLoader(bool preload) {
|
||||
const int flags = (preload) ? RTLD_LAZY : RTLD_LAZY|RTLD_NOLOAD;
|
||||
handle_ = dlopen(lib_name_, flags);
|
||||
if (handle_ == NULL) {
|
||||
fprintf(stderr, "roctracer: Loading '%s' failed, preload(%d), %s\n", lib_name_, (int)preload, dlerror());
|
||||
abort();
|
||||
}
|
||||
|
||||
T::init(this);
|
||||
}
|
||||
|
||||
~BaseLoader() {
|
||||
if (handle_ != NULL) dlclose(handle_);
|
||||
}
|
||||
|
||||
protected:
|
||||
static mutex_t mutex_;
|
||||
|
||||
private:
|
||||
static const char* lib_name_;
|
||||
static std::atomic<loader_t*> instance_;
|
||||
void* handle_;
|
||||
};
|
||||
|
||||
// HIP runtime library loader class
|
||||
class HipLoader : protected Loader {
|
||||
class HipApi {
|
||||
public:
|
||||
typedef BaseLoader<HipApi> Loader;
|
||||
|
||||
typedef decltype(hipRegisterApiCallback) RegisterApiCallback_t;
|
||||
typedef decltype(hipRemoveApiCallback) RemoveApiCallback_t;
|
||||
typedef decltype(hipRegisterActivityCallback) RegisterActivityCallback_t;
|
||||
@@ -52,18 +65,6 @@ class HipLoader : protected Loader {
|
||||
typedef decltype(hipKernelNameRef) KernelNameRef_t;
|
||||
typedef decltype(hipApiName) ApiName_t;
|
||||
|
||||
static HipLoader& Instance() {
|
||||
HipLoader* obj = instance_.load(std::memory_order_acquire);
|
||||
if (obj == NULL) {
|
||||
std::lock_guard<mutex_t> lck(mutex_);
|
||||
if (instance_.load(std::memory_order_relaxed) == NULL) {
|
||||
obj = new HipLoader();
|
||||
instance_.store(obj, std::memory_order_release);
|
||||
}
|
||||
}
|
||||
return *instance_;
|
||||
}
|
||||
|
||||
RegisterApiCallback_t* RegisterApiCallback;
|
||||
RemoveApiCallback_t* RemoveApiCallback;
|
||||
RegisterActivityCallback_t* RegisterActivityCallback;
|
||||
@@ -72,122 +73,89 @@ class HipLoader : protected Loader {
|
||||
ApiName_t* ApiName;
|
||||
|
||||
protected:
|
||||
HipLoader() : Loader("libhip_hcc.so") {
|
||||
RegisterApiCallback = GetFun<RegisterApiCallback_t>("hipRegisterApiCallback");
|
||||
RemoveApiCallback = GetFun<RemoveApiCallback_t>("hipRemoveApiCallback");
|
||||
RegisterActivityCallback = GetFun<RegisterActivityCallback_t>("hipRegisterActivityCallback");
|
||||
RemoveActivityCallback = GetFun<RemoveActivityCallback_t>("hipRemoveActivityCallback");
|
||||
KernelNameRef = GetFun<KernelNameRef_t>("hipKernelNameRef");
|
||||
ApiName = GetFun<ApiName_t>("hipApiName");
|
||||
void init(Loader* loader) {
|
||||
RegisterApiCallback = loader->GetFun<RegisterApiCallback_t>("hipRegisterApiCallback");
|
||||
RemoveApiCallback = loader->GetFun<RemoveApiCallback_t>("hipRemoveApiCallback");
|
||||
RegisterActivityCallback = loader->GetFun<RegisterActivityCallback_t>("hipRegisterActivityCallback");
|
||||
RemoveActivityCallback = loader->GetFun<RemoveActivityCallback_t>("hipRemoveActivityCallback");
|
||||
KernelNameRef = loader->GetFun<KernelNameRef_t>("hipKernelNameRef");
|
||||
ApiName = loader->GetFun<ApiName_t>("hipApiName");
|
||||
}
|
||||
|
||||
private:
|
||||
static std::atomic<HipLoader*> instance_;
|
||||
};
|
||||
|
||||
// HCC runtime library loader class
|
||||
class HccLoader : protected Loader {
|
||||
class HccApi {
|
||||
public:
|
||||
typedef std::mutex mutex_t;
|
||||
typedef BaseLoader<HccApi> Loader;
|
||||
|
||||
typedef decltype(Kalmar::CLAMP::InitActivityCallback) InitActivityCallback_t;
|
||||
typedef decltype(Kalmar::CLAMP::EnableActivityCallback) EnableActivityCallback_t;
|
||||
typedef decltype(Kalmar::CLAMP::GetCmdName) GetCmdName_t;
|
||||
|
||||
static HccLoader* GetRef() { return instance_; }
|
||||
|
||||
static HccLoader& Instance() {
|
||||
HccLoader* obj = instance_.load(std::memory_order_acquire);
|
||||
if (obj == NULL) {
|
||||
std::lock_guard<mutex_t> lck(mutex_);
|
||||
if (instance_.load(std::memory_order_relaxed) == NULL) {
|
||||
obj = new HccLoader();
|
||||
instance_.store(obj, std::memory_order_release);
|
||||
}
|
||||
}
|
||||
return *obj;
|
||||
}
|
||||
|
||||
InitActivityCallback_t* InitActivityCallback;
|
||||
EnableActivityCallback_t* EnableActivityCallback;
|
||||
GetCmdName_t* GetCmdName;
|
||||
|
||||
protected:
|
||||
HccLoader() : Loader("libmcwamp_hsa.so") {
|
||||
void init(Loader* loader) {
|
||||
// Kalmar::CLAMP::InitActivityCallback
|
||||
InitActivityCallback = GetFun<InitActivityCallback_t>("InitActivityCallbackImpl");
|
||||
InitActivityCallback = loader->GetFun<InitActivityCallback_t>("InitActivityCallbackImpl");
|
||||
// Kalmar::CLAMP::EnableActivityIdCallback
|
||||
EnableActivityCallback = GetFun<EnableActivityCallback_t>("EnableActivityCallbackImpl");
|
||||
EnableActivityCallback = loader->GetFun<EnableActivityCallback_t>("EnableActivityCallbackImpl");
|
||||
// Kalmar::CLAMP::GetCmdName
|
||||
GetCmdName = GetFun<GetCmdName_t>("GetCmdNameImpl");
|
||||
GetCmdName = loader->GetFun<GetCmdName_t>("GetCmdNameImpl");
|
||||
}
|
||||
|
||||
private:
|
||||
static std::atomic<HccLoader*> instance_;
|
||||
};
|
||||
|
||||
// KFD runtime library loader class
|
||||
class KfdLoader : protected Loader {
|
||||
class KfdApi {
|
||||
public:
|
||||
typedef BaseLoader<KfdApi> Loader;
|
||||
|
||||
typedef bool (RegisterApiCallback_t)(uint32_t op, void* callback, void* arg);
|
||||
typedef bool (RemoveApiCallback_t)(uint32_t op);
|
||||
|
||||
static KfdLoader& Instance() {
|
||||
KfdLoader* obj = instance_.load(std::memory_order_acquire);
|
||||
if (obj == NULL) {
|
||||
std::lock_guard<mutex_t> lck(mutex_);
|
||||
if (instance_.load(std::memory_order_relaxed) == NULL) {
|
||||
obj = new KfdLoader();
|
||||
instance_.store(obj, std::memory_order_release);
|
||||
}
|
||||
}
|
||||
return *instance_;
|
||||
}
|
||||
|
||||
RegisterApiCallback_t* RegisterApiCallback;
|
||||
RemoveApiCallback_t* RemoveApiCallback;
|
||||
|
||||
protected:
|
||||
KfdLoader() : Loader("libkfdwrapper64.so") {
|
||||
RegisterApiCallback = GetFun<RegisterApiCallback_t>("RegisterApiCallback");
|
||||
RemoveApiCallback = GetFun<RemoveApiCallback_t>("RemoveApiCallback");
|
||||
void init(Loader* loader) {
|
||||
RegisterApiCallback = loader->GetFun<RegisterApiCallback_t>("RegisterApiCallback");
|
||||
RemoveApiCallback = loader->GetFun<RemoveApiCallback_t>("RemoveApiCallback");
|
||||
}
|
||||
|
||||
private:
|
||||
static std::atomic<KfdLoader*> instance_;
|
||||
};
|
||||
|
||||
// rocTX runtime library loader class
|
||||
class RocTxLoader : protected Loader {
|
||||
class RocTxApi {
|
||||
public:
|
||||
typedef BaseLoader<RocTxApi> Loader;
|
||||
|
||||
typedef bool (RegisterApiCallback_t)(uint32_t op, void* callback, void* arg);
|
||||
typedef bool (RemoveApiCallback_t)(uint32_t op);
|
||||
|
||||
static RocTxLoader& Instance() {
|
||||
RocTxLoader* obj = instance_.load(std::memory_order_acquire);
|
||||
if (obj == NULL) {
|
||||
std::lock_guard<mutex_t> lck(mutex_);
|
||||
if (instance_.load(std::memory_order_relaxed) == NULL) {
|
||||
obj = new RocTxLoader();
|
||||
instance_.store(obj, std::memory_order_release);
|
||||
}
|
||||
}
|
||||
return *instance_;
|
||||
}
|
||||
|
||||
RegisterApiCallback_t* RegisterApiCallback;
|
||||
RemoveApiCallback_t* RemoveApiCallback;
|
||||
|
||||
protected:
|
||||
RocTxLoader() : Loader("libroctx64.so") {
|
||||
RegisterApiCallback = GetFun<RegisterApiCallback_t>("RegisterApiCallback");
|
||||
RemoveApiCallback = GetFun<RemoveApiCallback_t>("RemoveApiCallback");
|
||||
void init(Loader* loader) {
|
||||
RegisterApiCallback = loader->GetFun<RegisterApiCallback_t>("RegisterApiCallback");
|
||||
RemoveApiCallback = loader->GetFun<RemoveApiCallback_t>("RemoveApiCallback");
|
||||
}
|
||||
|
||||
private:
|
||||
static std::atomic<RocTxLoader*> instance_;
|
||||
};
|
||||
|
||||
typedef BaseLoader<HipApi> HipLoader;
|
||||
typedef BaseLoader<HccApi> HccLoader;
|
||||
typedef BaseLoader<KfdApi> KfdLoader;
|
||||
typedef BaseLoader<RocTxApi> RocTxLoader;
|
||||
|
||||
} // namespace roctracer
|
||||
|
||||
#define LOADER_INSTANTIATE() \
|
||||
template<class T> typename roctracer::BaseLoader<T>::mutex_t roctracer::BaseLoader<T>::mutex_; \
|
||||
template<class T> std::atomic<roctracer::BaseLoader<T>*> roctracer::BaseLoader<T>::instance_{}; \
|
||||
template<> const char* roctracer::HipLoader::lib_name_ = "libhip_hcc.so"; \
|
||||
template<> const char* roctracer::HccLoader::lib_name_ = "libmcwamp_hsa.so"; \
|
||||
template<> const char* roctracer::KfdLoader::lib_name_ = "libkfdwrapper64.so"; \
|
||||
template<> const char* roctracer::RocTxLoader::lib_name_ = "libroctx64.so";
|
||||
|
||||
#endif // SRC_CORE_LOADER_H_
|
||||
|
||||
@@ -92,6 +92,9 @@ THE SOFTWARE.
|
||||
#define onload_debug false
|
||||
#endif
|
||||
|
||||
static inline uint32_t GetPid() { return syscall(__NR_getpid); }
|
||||
static inline uint32_t GetTid() { return syscall(__NR_gettid); }
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Mark callback
|
||||
//
|
||||
@@ -454,7 +457,7 @@ FILE* open_output_file(const char* prefix, const char* name) {
|
||||
FILE* file_handle = NULL;
|
||||
if (prefix != NULL) {
|
||||
std::ostringstream oss;
|
||||
oss << prefix << "/" << name;
|
||||
oss << prefix << "/" << GetPid() << "_" << name;
|
||||
file_handle = fopen(oss.str().c_str(), "w");
|
||||
if (file_handle == NULL) {
|
||||
std::ostringstream errmsg;
|
||||
|
||||
@@ -136,7 +136,7 @@ void hsa_api_callback(
|
||||
void hsa_api_flush_cb(hsa_api_trace_entry_t* entry) {
|
||||
std::ostringstream os;
|
||||
os << entry->begin << ":" << entry->end << " " << entry->pid << ":" << entry->tid << " " << hsa_api_data_pair_t(entry->cid, entry->data);
|
||||
fprintf(hsa_api_file_handle, "%s\n", os.str().c_str());
|
||||
fprintf(hsa_api_file_handle, "%s\n", os.str().c_str()); fflush(hsa_api_file_handle);
|
||||
}
|
||||
|
||||
void hsa_activity_callback(
|
||||
@@ -145,7 +145,7 @@ void hsa_activity_callback(
|
||||
void* arg)
|
||||
{
|
||||
static uint64_t index = 0;
|
||||
fprintf(hsa_async_copy_file_handle, "%lu:%lu async-copy%lu\n", record->begin_ns, record->end_ns, index);
|
||||
fprintf(hsa_async_copy_file_handle, "%lu:%lu async-copy%lu\n", record->begin_ns, record->end_ns, index); fflush(hsa_async_copy_file_handle);
|
||||
index++;
|
||||
}
|
||||
|
||||
@@ -356,7 +356,7 @@ FILE* open_output_file(const char* prefix, const char* name) {
|
||||
FILE* file_handle = NULL;
|
||||
if (prefix != NULL) {
|
||||
std::ostringstream oss;
|
||||
oss << prefix << "/" << name;
|
||||
oss << prefix << "/" << GetPid() << "_" << name;
|
||||
file_handle = fopen(oss.str().c_str(), "w");
|
||||
if (file_handle == NULL) {
|
||||
std::ostringstream errmsg;
|
||||
@@ -409,7 +409,7 @@ extern "C" PUBLIC_API bool OnLoad(HsaApiTable* table, uint64_t runtime_version,
|
||||
// API trace vector
|
||||
std::vector<std::string> hsa_api_vec;
|
||||
|
||||
printf("ROCTracer: "); fflush(stdout);
|
||||
printf("ROCTracer (pid=%d): ", (int)GetPid()); fflush(stdout);
|
||||
// XML input
|
||||
const char* xml_name = getenv("ROCP_INPUT");
|
||||
if (xml_name != NULL) {
|
||||
|
||||
Ссылка в новой задаче
Block a user