dispatch vs profiling packets signals race fixed

Change-Id: I7f250d01f3d45f3f8ec2539b2187ab99503850aa


[ROCm/rocprofiler commit: caa15d7186]
Этот коммит содержится в:
Evgeny
2018-03-16 10:34:44 -05:00
родитель a900f3eda4
Коммит 6e4240aef7
9 изменённых файлов: 61 добавлений и 24 удалений
+1 -1
Просмотреть файл
@@ -169,7 +169,7 @@ typedef enum {
} rocprofiler_mode_t;
// Profiling handler, calling on profiling completion
typedef void (*rocprofiler_handler_t)(rocprofiler_group_t group, void* arg);
typedef bool (*rocprofiler_handler_t)(rocprofiler_group_t group, void* arg);
// Profiling preperties
typedef struct {
+3 -3
Просмотреть файл
@@ -6,6 +6,7 @@
#include <hsa.h>
#include <hsa_ext_amd.h>
#include <map>
#include <mutex>
#include <vector>
#include "core/metrics.h"
@@ -125,8 +126,7 @@ class Group {
}
void ResetRefs() { refs_ = n_profiles_; }
uint32_t DecrRefs() {
--refs_;
return refs_;
return (refs_ > 0) ? --refs_ : 0;
}
private:
@@ -395,7 +395,7 @@ class Context {
uint32_t r = group->DecrRefs();
group->GetContext()->mutex_.unlock();
if (r == 0) {
group->GetContext()->handler_(group->GetGroup(), group->GetContext()->handler_arg_);
return group->GetContext()->handler_(group->GetGroup(), group->GetContext()->handler_arg_);
}
return false;
}
+1
Просмотреть файл
@@ -14,4 +14,5 @@ InterceptQueue::obj_map_t* InterceptQueue::obj_map_ = NULL;
const char* InterceptQueue::kernel_none_ = "";
uint64_t InterceptQueue::timeout_ = UINT64_MAX;
Tracker* InterceptQueue::tracker_ = NULL;
bool InterceptQueue::tracker_on_ = true;
} // namespace rocprofiler
+10 -5
Просмотреть файл
@@ -41,7 +41,7 @@ class InterceptQueue {
group_segment_size, queue, &status);
if (status != HSA_STATUS_SUCCESS) abort();
if (!tracker_) tracker_ = new Tracker(timeout_);
if (tracker_on_ && (tracker_ == NULL)) tracker_ = new Tracker(timeout_);
status = hsa_amd_profiling_set_profiler_enabled(*queue, true);
if (status != HSA_STATUS_SUCCESS) abort();
@@ -89,15 +89,19 @@ class InterceptQueue {
const hsa_kernel_dispatch_packet_t* dispatch_packet =
reinterpret_cast<const hsa_kernel_dispatch_packet_t*>(packet);
const char* kernel_name = GetKernelName(dispatch_packet);
const auto* entry = tracker_->Add(obj->agent_info_->dev_id, dispatch_packet->completion_signal);
const_cast<hsa_kernel_dispatch_packet_t*>(dispatch_packet)->completion_signal = entry->signal;
const rocprofiler_dispatch_record_t* record = NULL;
if (tracker_ != NULL) {
const auto* entry = tracker_->Add(obj->agent_info_->dev_id, dispatch_packet->completion_signal);
const_cast<hsa_kernel_dispatch_packet_t*>(dispatch_packet)->completion_signal = entry->signal;
record = entry->record;
}
rocprofiler_callback_data_t data = {obj->agent_info_->dev_id,
obj->agent_info_->dev_index,
obj->queue_,
user_que_idx,
dispatch_packet->kernel_object,
kernel_name,
entry->record};
record};
hsa_status_t status = dispatch_callback_(&data, callback_data_, &group);
free(const_cast<char*>(kernel_name));
if ((status == HSA_STATUS_SUCCESS) && (group.context != NULL)) {
@@ -137,6 +141,7 @@ class InterceptQueue {
}
static void SetTimeout(uint64_t timeout) { timeout_ = timeout; }
static void TrackerOn(bool on) { tracker_on_ = on; }
private:
InterceptQueue(const hsa_agent_t& agent, hsa_queue_t* const queue, ProxyQueue* proxy) :
@@ -188,7 +193,7 @@ class InterceptQueue {
static const char* kernel_none_;
static uint64_t timeout_;
static Tracker* tracker_;
static const bool tracker_on_ = true;
static bool tracker_on_;
hsa_queue_t* const queue_;
ProxyQueue* const proxy_;
+6
Просмотреть файл
@@ -148,6 +148,11 @@ CONSTRUCTOR_API void constructor() {
Context::SetTimeout(timeout_val);
InterceptQueue::SetTimeout(timeout_val);
}
const char* tracker_on_str = getenv("ROCP_TRACKER_ON");
if (tracker_on_str != NULL) {
if (strncmp(tracker_on_str, "true", 4) == 0) InterceptQueue::TrackerOn(true);
if (strncmp(tracker_on_str, "false", 4) == 0) InterceptQueue::TrackerOn(false);
}
}
DESTRUCTOR_API void destructor() {
@@ -173,6 +178,7 @@ const MetricsDict* GetMetrics(const hsa_agent_t& agent) {
return metrics;
}
rocprofiler::Tracker::mutex_t rocprofiler::Tracker::mutex_;
util::Logger::mutex_t util::Logger::mutex_;
util::Logger* util::Logger::instance_ = NULL;
uint64_t Context::timeout_ = UINT64_MAX;
+11
Просмотреть файл
@@ -6,6 +6,7 @@
#include <hsa_ext_amd.h>
#include <list>
#include <mutex>
#include "inc/rocprofiler.h"
#include "util/exception.h"
@@ -15,6 +16,7 @@ namespace rocprofiler {
class Tracker {
public:
typedef std::mutex mutex_t;
typedef rocprofiler_dispatch_record_t record_t;
struct entry_t;
typedef std::list<entry_t*> sig_list_t;
@@ -29,6 +31,7 @@ class Tracker {
Tracker(uint64_t timeout = UINT64_MAX) : timeout_(timeout) {}
~Tracker() {
mutex_.lock();
for (entry_t* entry : sig_list_) {
assert(entry != NULL);
while (1) {
@@ -43,6 +46,7 @@ class Tracker {
}
Del(entry);
}
mutex_.unlock();
}
// Add tracker entry
@@ -50,7 +54,9 @@ class Tracker {
entry_t* entry = new entry_t{};
assert(entry);
entry->tracker = this;
mutex_.lock();
entry->it = sig_list_.insert(sig_list_.begin(), entry);
mutex_.unlock();
entry->agent = agent;
entry->orig = orig;
@@ -73,7 +79,9 @@ class Tracker {
// Delete tracker entry
void Del(entry_t* entry) {
hsa_signal_destroy(entry->signal);
mutex_.lock();
sig_list_.erase(entry->it);
mutex_.unlock();
delete entry;
}
@@ -84,6 +92,7 @@ class Tracker {
hsa_status_t status = hsa_system_get_info(HSA_SYSTEM_INFO_TIMESTAMP, &record->complete);
if (status != HSA_STATUS_SUCCESS) EXC_RAISING(status, "hsa_system_get_info(HSA_SYSTEM_INFO_TIMESTAMP)");
if (record->complete == 0) EXC_RAISING(status, "hsa_system_get_info(HSA_SYSTEM_INFO_TIMESTAMP), time is zero");
hsa_amd_profiling_dispatch_time_t dispatch_time{};
status = hsa_amd_profiling_get_dispatch_time(entry->agent, entry->signal, &dispatch_time);
@@ -106,6 +115,8 @@ class Tracker {
uint64_t timeout_;
// Tracked signals list
sig_list_t sig_list_;
// Inter-thread synchronization
static mutex_t mutex_;
};
} // namespace rocprofiler
+27 -12
Просмотреть файл
@@ -8,6 +8,7 @@
#include <dirent.h>
#include <hsa.h>
#include <pthread.h>
#include <sched.h>
#include <stdio.h>
#include <stdlib.h>
#include <sys/types.h>
@@ -243,10 +244,15 @@ void output_group(FILE* file, const rocprofiler_group_t* group, const char* str)
}
// Dump stored context profiling output data
void dump_context(context_entry_t* entry) {
bool dump_context(context_entry_t* entry) {
hsa_status_t status = HSA_STATUS_ERROR;
if (entry->valid) {
const rocprofiler_dispatch_record_t* record = entry->data.record;
if (record) {
if (record->complete == 0) return false;
}
++context_collected;
entry->valid = 0;
const uint32_t index = entry->index;
@@ -254,18 +260,22 @@ void dump_context(context_entry_t* entry) {
const rocprofiler_feature_t* features = entry->features;
const unsigned feature_count = entry->feature_count;
fprintf(file_handle, "dispatch[%u], queue_index(%lu), kernel_name(\"%s\"), time(%lu,%lu,%lu,%lu)\n",
fprintf(file_handle, "dispatch[%u], queue_index(%lu), kernel_name(\"%s\")",
index,
entry->data.queue_index,
entry->data.kernel_name,
entry->data.record->dispatch,
entry->data.record->begin,
entry->data.record->end,
entry->data.record->complete);
entry->data.kernel_name);
if (record) fprintf(file_handle, ", time(%lu,%lu,%lu,%lu)",
record->dispatch,
record->begin,
record->end,
record->complete);
fprintf(file_handle, "\n");
fflush(file_handle);
delete entry->data.record;
entry->data.record = NULL;
if (record) {
delete record;
entry->data.record = NULL;
}
rocprofiler_group_t& group = entry->group;
if (group.context != NULL) {
@@ -285,6 +295,8 @@ void dump_context(context_entry_t* entry) {
rocprofiler_close(group.context);
}
}
return true;
}
// Dump all stored contexts profiling output data
@@ -303,8 +315,9 @@ void dump_context_array() {
}
// Profiling completion handler
void handler(rocprofiler_group_t group, void* arg) {
bool handler(rocprofiler_group_t group, void* arg) {
context_entry_t* entry = reinterpret_cast<context_entry_t*>(arg);
bool ret = false;
if (pthread_mutex_lock(&mutex) != 0) {
perror("pthread_mutex_lock");
@@ -312,14 +325,16 @@ void handler(rocprofiler_group_t group, void* arg) {
}
if (context_array->find(entry->index) != context_array->end()) {
dump_context(entry);
dealloc_context_entry(entry);
if (dump_context(entry)) dealloc_context_entry(entry);
else ret = true;
}
if (pthread_mutex_unlock(&mutex) != 0) {
perror("pthread_mutex_unlock");
exit(1);
}
return ret;
}
// Kernel disoatch callback
+1 -2
Просмотреть файл
@@ -40,7 +40,6 @@
<metric name="TCC_MC_WRREQ_STALL" block=TCC event=28 descr="Number of cycles a write request was stalled."></metric>
<metric name="TCP_TA_DATA_STALL_CYCLES" block=TCP event=3 descr="TCP stalls TA data interface. Now Windowed."></metric>
<metric name="TCP_TCP_TA_DATA_STALL_CYCLES" block=TCP event=3 descr="TCP stalls TA data interface. Not Windowed."></metric>
<metric name=CPC_ALWAYS_COUNT block=CPC event=0 ></metric>
<metric name=CPC_ME1_STALL_WAIT_ON_RCIU_READ block=CPC event=8 ></metric>
@@ -86,7 +85,7 @@
<metric name=TCC_EA_RDREQ block=TCC event=41 ></metric>
<metric name=TCC_EA_RDREQ_32B block=TCC event=42 ></metric>
<metric name=TCP_TA_DATA_STALL_CYCLES block=TCP event=6 descr="TCP stalls TA data interface. Now Windowed."></metric>
<metric name="TCP_TA_DATA_STALL_CYCLES" block=TCP event=6 descr="TCP stalls TA data interface. Now Windowed."></metric>
<metric name=CPC_ALWAYS_COUNT block=CPC event=0 ></metric>
<metric name=CPC_ME1_STALL_WAIT_ON_RCIU_READ block=CPC event=8 ></metric>
+1 -1
Просмотреть файл
@@ -178,7 +178,7 @@
<metric
name="MemUnitStalled"
descr="The percentage of GPUTime the memory unit is stalled. Try reducing the number or size of fetches and writes if possible. Value range: 0% (optimal) to 100% (bad)."
expr=100*max(TCP_TCP_TA_DATA_STALL_CYCLES,16)/GRBM_GUI_ACTIVE/SE_NUM
expr=100*max(TCP_TA_DATA_STALL_CYCLES,16)/GRBM_GUI_ACTIVE/SE_NUM
></metric>
# WriteUnitStalled The percentage of GPUTime the Write unit is stalled. Value range: 0% to 100% (bad).