AI/rocm-systems
2
0
Fork 0
Koodi Ongelmat Pull-pyynnöt Actions Paketit Projektit Julkaisut Wiki Toiminta

several fixes; adding direct loading of alprofile library

Selaa lähdekoodia 
[ROCm/rocprofiler commit: 8270530fec]
This commit is contained in:
Evgeny
2017-12-19 15:32:34 -06:00
vanhempi e1f9046f5c
commit 24695e37be
16 muutettua tiedostoa jossa 195 lisäystä ja 136 poistoa
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocprofiler/src/core/intercept_queue.h
+4 -2
Näytä tiedosto
@@ -55,12 +55,14 @@ class InterceptQueue {
ProxyQueue* proxy = ProxyQueue::Create(agent, size, type, callback, data, private_segment_size,
group_segment_size, queue, &status);
if (status != HSA_STATUS_SUCCESS) {
if (status == HSA_STATUS_SUCCESS) {
InterceptQueue* obj = new InterceptQueue(agent, proxy);
(*obj_map_)[(uint64_t)(*queue)] = obj;
status = proxy->SetInterceptCB(OnSubmitCB, obj);
}
if (status != HSA_STATUS_SUCCESS) abort();
return status;
}
@@ -97,7 +99,7 @@ class InterceptQueue {
dispatch_packet->kernel_object,
GetKernelName(dispatch_packet)};
hsa_status_t status = on_dispatch_cb_(&data, on_dispatch_cb_data_, &group);
if (status == HSA_STATUS_SUCCESS) {
if ((status == HSA_STATUS_SUCCESS) && (group.context != NULL)) {
Context* context = reinterpret_cast<Context*>(group.context);
const pkt_vector_t& start_vector = context->StartPackets(group.index);
const pkt_vector_t& stop_vector = context->StopPackets(group.index);
projects/rocprofiler/src/core/metrics.h
+6 -1
Näytä tiedosto
@@ -110,13 +110,18 @@ class MetricsDict {
MetricsDict(const util::AgentInfo* agent_info) : xml_(NULL) {
const char* xml_name = getenv("ROCP_METRICS");
if (xml_name != NULL) {
xml_ = new xml::Xml(xml_name);
xml_ = xml::Xml::Create(xml_name);
if (xml_ == NULL) EXC_RAISING(HSA_STATUS_ERROR, "metrics .xml open error '" << xml_name << "'");
std::cout << "ROCProfiler: importing metrics from '" << xml_name << "':" << std::endl;
ImportMetrics(agent_info, agent_info->gfxip);
ImportMetrics(agent_info, "global");
}
}
~MetricsDict() {
xml::Xml::Destroy(xml_);
}
void ImportMetrics(const util::AgentInfo* agent_info, const char* scope) {
auto scope_list = xml_->GetNodes("top." + std::string(scope) + ".metric");
if (!scope_list.empty()) {
projects/rocprofiler/src/core/proxy_queue.cpp
+3 -1
Näytä tiedosto
@@ -24,7 +24,7 @@ ProxyQueue* ProxyQueue::Create(hsa_agent_t agent, uint32_t size, hsa_queue_type3
ProxyQueue* instance = new SimpleProxyQueue();
#endif
if (instance != NULL) {
const auto suc = instance->Init(agent, size, type, callback, data, private_segment_size,
suc = instance->Init(agent, size, type, callback, data, private_segment_size,
group_segment_size, queue);
if (suc != HSA_STATUS_SUCCESS) {
delete instance;
@@ -32,10 +32,12 @@ ProxyQueue* ProxyQueue::Create(hsa_agent_t agent, uint32_t size, hsa_queue_type3
}
}
*status = suc;
assert(*status == HSA_STATUS_SUCCESS);
return instance;
}
hsa_status_t ProxyQueue::Destroy(const ProxyQueue* obj) {
assert(obj != NULL);
auto suc = obj->Cleanup();
delete obj;
return suc;
projects/rocprofiler/src/core/simple_proxy_queue.cpp
+1 -1
Näytä tiedosto
@@ -7,5 +7,5 @@ void SimpleProxyQueue::HsaIntercept(HsaApiTable* table) {
table->core_->hsa_queue_store_write_index_relaxed_fn = rocprofiler::SimpleProxyQueue::StoreIndex;
}
std::map<signal_handle_t, SimpleProxyQueue*> SimpleProxyQueue::queue_map_;
SimpleProxyQueue::queue_map_t* SimpleProxyQueue::queue_map_ = NULL;
} // namespace rocprofiler
projects/rocprofiler/src/core/simple_proxy_queue.h
+16 -8
Näytä tiedosto
@@ -24,8 +24,8 @@ class SimpleProxyQueue : public ProxyQueue {
static void HsaIntercept(HsaApiTable* table);
static void SignalStore(hsa_signal_t signal, hsa_signal_value_t que_idx) {
auto it = queue_map_.find(signal.handle);
if (it != queue_map_.end()) {
auto it = queue_map_->find(signal.handle);
if (it != queue_map_->end()) {
SimpleProxyQueue* instance = it->second;
const uint64_t begin = instance->submit_index_;
const uint64_t end = que_idx + 1;
@@ -46,8 +46,8 @@ class SimpleProxyQueue : public ProxyQueue {
static uint64_t LoadIndex(const hsa_queue_t* queue) {
uint64_t index = 0;
auto it = queue_map_.find(queue->doorbell_signal.handle);
if (it != queue_map_.end()) {
auto it = queue_map_->find(queue->doorbell_signal.handle);
if (it != queue_map_->end()) {
SimpleProxyQueue* instance = it->second;
instance->mutex_.lock();
index = instance->queue_index_;
@@ -58,8 +58,8 @@ class SimpleProxyQueue : public ProxyQueue {
}
static void StoreIndex(const hsa_queue_t* queue, uint64_t value) {
auto it = queue_map_.find(queue->doorbell_signal.handle);
if (it != queue_map_.end()) {
auto it = queue_map_->find(queue->doorbell_signal.handle);
if (it != queue_map_->end()) {
SimpleProxyQueue* instance = it->second;
instance->queue_index_ = value;
instance->mutex_.unlock();
@@ -115,6 +115,8 @@ class SimpleProxyQueue : public ProxyQueue {
~SimpleProxyQueue() {}
private:
typedef std::map<signal_handle_t, SimpleProxyQueue*> queue_map_t;
hsa_status_t Init(hsa_agent_t agent, uint32_t size, hsa_queue_type32_t type,
void (*callback)(hsa_status_t status, hsa_queue_t* source, void* data),
void* data, uint32_t private_segment_size, uint32_t group_segment_size,
@@ -129,6 +131,7 @@ class SimpleProxyQueue : public ProxyQueue {
agent_info_ = util::HsaRsrcFactory::Instance().GetAgentInfo(agent);
if (agent_info_ != NULL) {
if (agent_info_->dev_type == HSA_DEVICE_TYPE_GPU) {
printf("queue_create size 0x%x(%d)\n", size, (int)size);
status = hsa_queue_create_fn(agent, size, HSA_QUEUE_TYPE_MULTI, NULL, NULL, UINT32_MAX,
UINT32_MAX, &queue_);
if (status == HSA_STATUS_SUCCESS) {
@@ -138,11 +141,16 @@ class SimpleProxyQueue : public ProxyQueue {
uintptr_t addr = (uintptr_t)data_array_;
queue_->base_address = (void*)((addr + align_mask_) & ~align_mask_);
status = hsa_signal_create(1, 0, NULL, &(queue_->doorbell_signal));
if (status != HSA_STATUS_SUCCESS) abort();
queue_mask_ = size - 1;
queue_map_[queue_->doorbell_signal.handle] = this;
if (queue_map_ == NULL) queue_map_ = new queue_map_t;
(*queue_map_)[queue_->doorbell_signal.handle] = this;
}
else abort();
}
}
if (status != HSA_STATUS_SUCCESS) abort();
return status;
}
@@ -155,7 +163,7 @@ class SimpleProxyQueue : public ProxyQueue {
return status;
}
static std::map<signal_handle_t, SimpleProxyQueue*> queue_map_;
static queue_map_t* queue_map_;
const util::AgentInfo* agent_info_;
hsa_queue_t* queue_;
static const uintptr_t align_mask_ = sizeof(packet_t) - 1;
projects/rocprofiler/src/util/hsa_rsrc_factory.cpp
+38 -1
Näytä tiedosto
@@ -24,6 +24,7 @@ POSSIBILITY OF SUCH DAMAGE.
#include "util/hsa_rsrc_factory.h"
#include <dlfcn.h>
#include <hsa.h>
#include <hsa_ext_finalize.h>
#include <stdint.h>
@@ -85,7 +86,10 @@ HsaRsrcFactory::HsaRsrcFactory() {
// Get AqlProfile API table
aqlprofile_api_ = {0};
status = hsa_system_get_extension_table(HSA_EXTENSION_AMD_AQLPROFILE, 1, 0, &aqlprofile_api_);
CHECK_STATUS("aqlprofile API table query failed", status);
#ifdef ROCP_LOAD_AQLPROF
if (status != HSA_STATUS_SUCCESS) status = LoadAqlProfileLib(&aqlprofile_api_);
#endif
CHECK_STATUS("aqlprofile API table load failed", status);
// Get Loader API table
loader_api_ = {0};
@@ -99,6 +103,39 @@ HsaRsrcFactory::~HsaRsrcFactory() {
CHECK_STATUS("Error in hsa_shut_down", status);
}
hsa_status_t HsaRsrcFactory::LoadAqlProfileLib(aqlprofile_pfn_t* api) {
void* handle = dlopen(kAqlProfileLib, RTLD_NOW);
if (handle == NULL) {
fprintf(stderr, "Loading '%s' failed, %s\n", kAqlProfileLib, dlerror());
return HSA_STATUS_ERROR;
}
dlerror(); /* Clear any existing error */
api->hsa_ven_amd_aqlprofile_error_string =
(decltype(::hsa_ven_amd_aqlprofile_error_string)*)
dlsym(handle, "hsa_ven_amd_aqlprofile_error_string");
api->hsa_ven_amd_aqlprofile_validate_event =
(decltype(::hsa_ven_amd_aqlprofile_validate_event)*)
dlsym(handle, "hsa_ven_amd_aqlprofile_validate_event");
api->hsa_ven_amd_aqlprofile_start =
(decltype(::hsa_ven_amd_aqlprofile_start)*)
dlsym(handle, "hsa_ven_amd_aqlprofile_start");
api->hsa_ven_amd_aqlprofile_stop =
(decltype(::hsa_ven_amd_aqlprofile_stop)*)
dlsym(handle, "hsa_ven_amd_aqlprofile_stop");
api->hsa_ven_amd_aqlprofile_legacy_get_pm4 =
(decltype(::hsa_ven_amd_aqlprofile_legacy_get_pm4)*)
dlsym(handle, "hsa_ven_amd_aqlprofile_legacy_get_pm4");
api->hsa_ven_amd_aqlprofile_get_info =
(decltype(::hsa_ven_amd_aqlprofile_get_info)*)
dlsym(handle, "hsa_ven_amd_aqlprofile_get_info");
api->hsa_ven_amd_aqlprofile_iterate_data =
(decltype(::hsa_ven_amd_aqlprofile_iterate_data)*)
dlsym(handle, "hsa_ven_amd_aqlprofile_iterate_data");
return HSA_STATUS_SUCCESS;
}
// Add system agent info
const AgentInfo* HsaRsrcFactory::AddAgentInfo(const hsa_agent_t agent) {
// Determine if device is a Gpu agent
projects/rocprofiler/src/util/hsa_rsrc_factory.h
+6 -2
Näytä tiedosto
@@ -222,7 +222,8 @@ class HsaRsrcFactory {
bool PrintGpuAgents(const std::string& header);
// Return AqlProfile API table
const hsa_ven_amd_aqlprofile_1_00_pfn_t* AqlProfileApi() const { return &aqlprofile_api_; }
typedef hsa_ven_amd_aqlprofile_1_00_pfn_t aqlprofile_pfn_t;
const aqlprofile_pfn_t* AqlProfileApi() const { return &aqlprofile_api_; }
// Return Loader API table
const hsa_ven_amd_loader_1_00_pfn_t* LoaderApi() const { return &loader_api_; }
@@ -234,6 +235,9 @@ class HsaRsrcFactory {
// Callback function to find and bind kernarg region of an agent
static hsa_status_t FindMemRegionsCallback(hsa_region_t region, void* data);
// Load AQL profile HSA extension library directly
static hsa_status_t LoadAqlProfileLib(aqlprofile_pfn_t* api);
// Constructor of the class. Will initialize the Hsa Runtime and
// query the system topology to get the list of Cpu and Gpu devices
HsaRsrcFactory();
@@ -257,7 +261,7 @@ class HsaRsrcFactory {
std::map<hsa_agent_handle_t, const AgentInfo*> agent_map_;
// AqlProfile API table
hsa_ven_amd_aqlprofile_1_00_pfn_t aqlprofile_api_;
aqlprofile_pfn_t aqlprofile_api_;
// Loader API table
hsa_ven_amd_loader_1_00_pfn_t loader_api_;
projects/rocprofiler/src/xml/xml.h
+29 -16
Näytä tiedosto
@@ -27,6 +27,33 @@ class Xml {
enum { DECL_STATE, BODY_STATE };
static Xml* Create(const char* file_name) {
Xml* xml = new Xml(file_name);
if (xml->fd_ == -1) {
delete xml;
xml = NULL;
}
return xml;
}
static void Destroy(Xml *xml) { delete xml; }
std::vector<level_t*> GetNodes(std::string global_tag) { return map_[global_tag]; }
void Print() const {
for (auto& elem : map_) {
for (auto node : elem.second) {
if (node->opts.size()) {
std::cout << elem.first << ":" << std::endl;
for (auto& opt : node->opts) {
std::cout << " " << opt.first << " = " << opt.second << std::endl;
}
}
}
}
}
private:
Xml(const char* file_name)
: file_name_(file_name),
file_line_(0),
@@ -39,7 +66,7 @@ class Xml {
fd_ = open(file_name, O_RDONLY);
if (fd_ == -1) {
std::cout << "XML file not found: " << file_name << std::endl;
perror("open XML file");
return;
}
@@ -117,22 +144,8 @@ class Xml {
}
}
std::vector<level_t*> GetNodes(std::string global_tag) { return map_[global_tag]; }
~Xml() {}
void Print() const {
for (auto& elem : map_) {
for (auto node : elem.second) {
if (node->opts.size()) {
std::cout << elem.first << ":" << std::endl;
for (auto& opt : node->opts) {
std::cout << " " << opt.first << " = " << opt.second << std::endl;
}
}
}
}
}
private:
bool LineEndCheck() {
bool found = false;
if (buffer_[index_] == '\n') {