Summarizing device timing regardless of kernel shapes by default (#37)

* Initial version of unitrace * Initial commits of unitrace * Initial commits of unitrace * Initial commits of unitrace * Initial commits of unitrace * Initial commits of unitrace * Initial commits of unitrace * Initial commits of unitrace * Initial commits of unitrace * Initial commits of unitrace * Initial commits of unitrace * Initial commits of unitrace * Unhide Symbols Required By XPTI * Initial commits of unitrace * Initial commits of unitrace * Summarizing device timing regardless of kernel shapes by default * Summarizing device timing with out kernel shapes by default * Summarizing device timing with out kernel shapes by default * Summarizing device timing with out kernel shapes by default --------- Co-authored-by: Schilling, Matthew <matthew.schilling@intel.com>
intel · Dec 11, 2023 · 6a85c1d · 6a85c1d
1 parent 93f66e7
commit 6a85c1d
Show file tree

Hide file tree

Showing 8 changed files with 46 additions and 67 deletions.
diff --git a/tools/unitrace/README.md b/tools/unitrace/README.md
@@ -131,18 +131,26 @@ To trace/profile device and kernel activities, one can use one or more of the fo
 
 The **--device-timing [-d]** option outputs a timing summary of kernels and commands executed on the device:
 
-![Device Timing!](/tools/unitrace/doc/images/device-timing.png)
+
+![Device Timing With No Shape!](/tools/unitrace/doc/images/device-timing-with-no-shape.png)
 
 In addition, it also outputs kernel information that helps to identify kernel performance issues that relate to occupancy caused by shared local memory usage and register spilling.
 
-![Kernel Info!](/tools/unitrace/doc/images/kernel-info.png)
+![Kernel Info With No Shape!](/tools/unitrace/doc/images/kernel-info-with-no-shape.png)
 
 Here, the **"SLM Per Work Group"** shows the amount of shared local memory needed for each work group in bytes. This size can potentially affect occupancy.
 
 The **"Private Memory Per Thread"** is the private memory allocated for each thread in bytes. A non-zero value indicates that one or more thread private variables are not in registers.
 
 The **"Spill Memory Per Thread"** is the memory used for register spilled for each thread in bytes. A non-zero value indicates that one or more thread private variables are allocated in registers but are later spilled to memory.
 
+By default, the kernel timing is summarized regardless of shapes. In case the kernel has different shapes, using **-v** along with **-d** is strongly recommended:
+
+![Device Timing!](/tools/unitrace/doc/images/device-timing.png)
+
+![Kernel Info!](/tools/unitrace/doc/images/kernel-info.png)
+
+
 The **--kernel-submission [-s]** option outputs a time summary of kernels spent in queuing, submission and execution:
 ![Kernel Submissions!](/tools/unitrace/doc/images/kernel-submissions.png)
 

diff --git a/tools/unitrace/doc/images/device-timing-with-no-shape.png b/tools/unitrace/doc/images/device-timing-with-no-shape.png
diff --git a/tools/unitrace/doc/images/kernel-info-with-no-shape.png b/tools/unitrace/doc/images/kernel-info-with-no-shape.png
diff --git a/tools/unitrace/src/chromelogger.h b/tools/unitrace/src/chromelogger.h
@@ -39,7 +39,7 @@ static uint32_t mpi_rank = std::atoi(rank.c_str());
 static std::string process_start_time = std::to_string(UniTimer::GetEpochTimeInUs(UniTimer::GetHostTimestamp()));
 static std::string pmi_hostname = GetHostName();
 
-std::string GetZeKernelCommandName(uint64_t id, ze_group_count_t& group_count, size_t size);
+std::string GetZeKernelCommandName(uint64_t id, ze_group_count_t& group_count, size_t size, bool detailed);
 ze_pci_ext_properties_t *GetZeDevicePciPropertiesAndId(ze_device_handle_t device, int32_t *parent_device_id, int32_t *device_id, int32_t *subdevice_id);
 
 static Logger* logger_ = nullptr;

diff --git a/tools/unitrace/src/levelzero/ze_collector.h b/tools/unitrace/src/levelzero/ze_collector.h
@@ -942,41 +942,43 @@ typedef void (*OnZeKernelFinishCallback)(uint64_t kid, uint64_t tid, uint64_t st
 
 ze_result_t (*zexKernelGetBaseAddress)(ze_kernel_handle_t hKernel, uint64_t *baseAddress) = nullptr;
 
-inline std::string GetZeKernelCommandName(uint64_t id, const ze_group_count_t& group_count, size_t size) {
+inline std::string GetZeKernelCommandName(uint64_t id, const ze_group_count_t& group_count, size_t size, bool detailed = true) {
   std::stringstream s;
   kernel_command_properties_mutex_.lock_shared();
   auto it = kernel_command_properties_->find(id);
   if (it != kernel_command_properties_->end()) {
     s << utils::Demangle(it->second.name_.c_str());
-    if (it->second.type_ == KERNEL_COMMAND_TYPE_COMPUTE) {
-      if (it->second.simd_width_ > 0) {
-        s << "[SIMD";
-        if (it->second.simd_width_ == 1) {
-          s << "_ANY";
-        } else {
-          s << it->second.simd_width_;
+    if (detailed) {
+      if (it->second.type_ == KERNEL_COMMAND_TYPE_COMPUTE) {
+        if (it->second.simd_width_ > 0) {
+          s << "[SIMD";
+          if (it->second.simd_width_ == 1) {
+            s << "_ANY";
+          } else {
+            s << it->second.simd_width_;
+          }
         }
+        s << " {" <<
+          group_count.groupCountX << "; " <<
+          group_count.groupCountY << "; " <<
+          group_count.groupCountZ << "} {" <<
+          it->second.group_size_.x << "; " <<
+          it->second.group_size_.y << "; " <<
+          it->second.group_size_.z << "}]";
+      }
+      else if ((it->second.type_ == KERNEL_COMMAND_TYPE_MEMORY) && (size > 0)) {
+        s << "[" << size << "]";
       }
-      s << " {" <<
-        group_count.groupCountX << "; " <<
-        group_count.groupCountY << "; " <<
-        group_count.groupCountZ << "} {" <<
-        it->second.group_size_.x << "; " <<
-        it->second.group_size_.y << "; " <<
-        it->second.group_size_.z << "}]";
-    }
-    else if ((it->second.type_ == KERNEL_COMMAND_TYPE_MEMORY) && (size > 0)) {
-      s << "[" << size << "]";
     }
   }
 
   kernel_command_properties_mutex_.unlock_shared();
   return s.str();
 }
 
-inline std::string GetZeKernelCommandName(uint64_t id, ze_group_count_t& group_count, size_t size) {
+inline std::string GetZeKernelCommandName(uint64_t id, ze_group_count_t& group_count, size_t size, bool detailed = true) {
   const ze_group_count_t& gcount = group_count;
-  return GetZeKernelCommandName(id, gcount, size);
+  return GetZeKernelCommandName(id, gcount, size, detailed);
 }
 
 inline ze_pci_ext_properties_t *GetZeDevicePciPropertiesAndId(ze_device_handle_t device, int32_t *parent_device_id, int32_t *device_id, int32_t *subdevice_id){
@@ -1115,10 +1117,10 @@ class ZeCollector {
       total_time += it.second.execute_time_;
       std::string kname;
       if (it.first.tile_ >= 0) {
-        kname = "Tile #" + std::to_string(it.first.tile_) + ": " + GetZeKernelCommandName(it.first.kernel_command_id_, it.first.group_count_, it.first.mem_size_);
+        kname = "Tile #" + std::to_string(it.first.tile_) + ": " + GetZeKernelCommandName(it.first.kernel_command_id_, it.first.group_count_, it.first.mem_size_, options_.verbose);
       }
       else {
-        kname = GetZeKernelCommandName(it.first.kernel_command_id_, it.first.group_count_, it.first.mem_size_);
+        kname = GetZeKernelCommandName(it.first.kernel_command_id_, it.first.group_count_, it.first.mem_size_, options_.verbose);
       }
       if (kname.size() > max_name_size) {
         max_name_size = kname.size();
@@ -1204,10 +1206,10 @@ class ZeCollector {
       total_submit_time += it.second.submit_time_;
       std::string kname;
       if (it.first.tile_ >= 0) {
-        kname = "Tile #" + std::to_string(it.first.tile_) + ": " + GetZeKernelCommandName(it.first.kernel_command_id_, it.first.group_count_, it.first.mem_size_);
+        kname = "Tile #" + std::to_string(it.first.tile_) + ": " + GetZeKernelCommandName(it.first.kernel_command_id_, it.first.group_count_, it.first.mem_size_, options_.verbose);
       }
       else {
-        kname = GetZeKernelCommandName(it.first.kernel_command_id_, it.first.group_count_, it.first.mem_size_);
+        kname = GetZeKernelCommandName(it.first.kernel_command_id_, it.first.group_count_, it.first.mem_size_, options_.verbose);
       }
       if (kname.size() > max_name_size) {
         max_name_size = kname.size();
@@ -1615,37 +1617,6 @@ class ZeCollector {
 
                 sub_desc.driver_ = driver;
                 sub_desc.context_ = context;
-#if 0
-                if (options_.metric_query) {
-                  zet_metric_group_handle_t group = nullptr;
-                  uint32_t num_groups = 0;
-                  status = zetMetricGroupGet(sub_devices[j], &num_groups, nullptr);
-                  PTI_ASSERT(status == ZE_RESULT_SUCCESS);
-                  if (num_groups > 0) {
-                    std::vector<zet_metric_group_handle_t> groups(num_groups, nullptr);
-                    status = zetMetricGroupGet(sub_devices[j], &num_groups, groups.data());
-                    PTI_ASSERT(status == ZE_RESULT_SUCCESS);
-
-                    for (uint32_t k = 0; k < num_groups; ++k) {
-                      zet_metric_group_properties_t group_props{};
-                      group_props.stype = ZET_STRUCTURE_TYPE_METRIC_GROUP_PROPERTIES;
-                      status = zetMetricGroupGetProperties(groups[k], &group_props);
-                      PTI_ASSERT(status == ZE_RESULT_SUCCESS);
-
-
-                      if ((strcmp(group_props.name, utils::GetEnv("UNITRACE_MetricGroup").c_str()) == 0) && (group_props.samplingType & ZET_METRIC_GROUP_SAMPLING_TYPE_FLAG_EVENT_BASED)) {
-                        group = groups[k];
-                        break;
-                      }
-                    }
-                  }
-                  status = zetContextActivateMetricGroups(context, sub_devices[j], 1, &group);
-                  PTI_ASSERT(status == ZE_RESULT_SUCCESS);
-                  metric_activations_.insert({context, sub_devices[j]});
-
-                  sub_desc.metric_group_ = group;
-                }
-#endif /* 0 */
 
                 sub_desc.metric_group_ = nullptr;
 

diff --git a/tools/unitrace/src/tracer.cc b/tools/unitrace/src/tracer.cc
@@ -155,7 +155,7 @@ static TraceOptions ReadArgs() {
 }
 
 std::string get_version() {
-  return std::string(VERSION) + " ("+ std::string(COMMIT_HASH) + ")";
+  return std::string(UNITRACE_VERSION) + " ("+ std::string(COMMIT_HASH) + ")";
 }
 
 void __attribute__((constructor)) Init(void) {
@@ -168,7 +168,7 @@ void __attribute__((constructor)) Init(void) {
   if (unitrace_version.size() > 0) {
     auto libunitrace_version = get_version();
     if (unitrace_version.compare(libunitrace_version) != 0) {
-      std::cerr << "[ERROR] Versions of Unitrace and libUnitrace_tool.so do not match." << std::endl;
+      std::cerr << "[ERROR] Versions of unitrace and libunitrace_tool.so do not match." << std::endl;
       exit(-1);
     }
   }

diff --git a/tools/unitrace/src/unitrace.cc b/tools/unitrace/src/unitrace.cc
@@ -119,8 +119,8 @@ void Usage(char * progname) {
     std::endl;
   std::cout <<
     "--verbose [-v]                 " <<
-    "Enable verbose mode to show more kernel information. For OpenCL backend only." << std::endl <<
-    "                               Verbose is always enabled for Level Zero backend" <<
+    "Enable verbose mode to show kernel shapes" << std::endl <<
+    "                               Kernel shapes are always enabled in timelines for Level Zero backend" <<
     std::endl;
   std::cout <<
     "--demangle                     " <<
@@ -384,7 +384,7 @@ int ParseArgs(int argc, char* argv[]) {
       show_metric_list = true;
       ++app_index;
     } else if (strcmp(argv[i], "--version") == 0) {
-      std::cout << VERSION << " (" << COMMIT_HASH << ")" << std::endl;
+      std::cout << UNITRACE_VERSION << " (" << COMMIT_HASH << ")" << std::endl;
       return 0;
     } else {
       break;
@@ -555,7 +555,7 @@ int main(int argc, char *argv[]) {
 #endif
 
   // Set unitrace version
-  auto unitrace_version =  std::string(VERSION) + " (" +  std::string(COMMIT_HASH) + ")";
+  auto unitrace_version =  std::string(UNITRACE_VERSION) + " (" +  std::string(COMMIT_HASH) + ")";
   utils::SetEnv("UNITRACE_VERSION", unitrace_version.c_str());
 
   SetProfilingEnvironment();
@@ -589,11 +589,11 @@ int main(int argc, char *argv[]) {
   if (utils::GetEnv("UNITRACE_ChromeMpiLogging") == "1") {
     preload = preload + ":" + mpi_interceptor_path;
     // For tracing MPI calls from oneCCL, we need to set CCL_MPI_LIBRARY_PATH
-    // with Unitrace's MPI intercepter path, because oneCCL directly picks up
+    // with unitrace's MPI intercepter path, because oneCCL directly picks up
     // MPI functions with dlopen/dlsym, not through the dynamic linker. Thus,
     // LD_PRELOAD would not work.
     // TODO: We have to consider a case where CCL_MPI_LIBRARY_PATH is already
-    //       set. Unitrace will need to call the MPIs in the specified libs
+    //       set. In this case, unitrace needs to call MPIs in the specified libs
     //       before/after ITT annotation.
     utils::SetEnv("CCL_MPI_LIBRARY_PATH", mpi_interceptor_path.c_str());
   }

diff --git a/tools/unitrace/src/version.h b/tools/unitrace/src/version.h
@@ -1,7 +1,7 @@
 #ifndef PTI_TOOLS_UNITRACE_VERSION_H_
 #define PTI_TOOLS_UNITRACE_VERSION_H_
 
-#define VERSION	"2.0.0"
+#define UNITRACE_VERSION	"2.0.1"
 
 std::string get_version();