[PROTON][XPU] Cleanup debug stuff

Signed-off-by: Anatoly Myachev <[email protected]>

Author: anmyachev

third_party/proton/csrc/lib/Profiler/Xpupti/XpuptiProfiler.cpp

@@ -75,33 +75,16 @@ processActivityKernel(XpuptiProfiler::CorrIdToExternIdMap &corrIdToExternId,
                       std::set<Data *> &dataSet,
                       xpupti::Pti_Activity *activity) {
   auto *kernel = reinterpret_cast<pti_view_record_kernel *>(activity);
-  // std::cout << "activity->_name: " << kernel->_name << "\n" << std::flush;
-  // std::cout << "activity->_sycl_queue_id: " << kernel->_sycl_queue_id << "\n"
-  // << std::flush;
   auto correlationId = kernel->_correlation_id;
-  std::cout << "kernel->_correlation_id " << kernel->_correlation_id << "\n"
-            << std::flush;
-  std::cout << "kernel->_kernel_id " << kernel->_kernel_id << "\n";
-  // here doesn't work
-  // uint64_t corr_id = 0;
-  // auto res =
-  // ptiViewPopExternalCorrelationId(pti_view_external_kind::PTI_VIEW_EXTERNAL_KIND_CUSTOM_1,
-  // &corr_id); std::cout << "ptiViewPopExternalCorrelationId res: " << res <<
-  // "\n" << std::flush; std::cout << "corr_id: " << corr_id << "\n" <<
-  // std::flush;
   if (/*Not a valid context*/ !corrIdToExternId.contain(correlationId)) {
-    // if (false) {
-    std::cout << "MARK#3\n" << std::flush;
     return correlationId;
   }
   auto [parentId, numInstances] = corrIdToExternId.at(correlationId);
-  std::cout << "parentId: " << parentId << std::endl;
   if (true) {
     // Non-graph kernels
     for (auto *data : dataSet) {
       auto scopeId = parentId;
       if (apiExternIds.contain(scopeId)) {
-        std::cout << "first branch" << std::endl;
         // It's triggered by a CUDA op but not triton op
         scopeId = data->addOp(parentId, kernel->_name);
       }
@@ -116,10 +99,8 @@ processActivityKernel(XpuptiProfiler::CorrIdToExternIdMap &corrIdToExternId,
     // 2. graphExecId -> graphId
     // --- CUPTI thread ---
     // 3. corrId -> numKernels
-    std::cout << "MARK#1\n" << std::flush;
     for (auto *data : dataSet) {
       auto externId = data->addOp(parentId, kernel->_name);
-      std::cout << "MARK#2\n" << std::flush;
       data->addMetric(externId, convertActivityToMetric(activity));
     }
   }
@@ -133,21 +114,6 @@ processActivityKernel(XpuptiProfiler::CorrIdToExternIdMap &corrIdToExternId,
   return correlationId;
 }
 
-uint32_t processActivityExternalCorrelation(
-    XpuptiProfiler::CorrIdToExternIdMap &corrIdToExternId,
-    xpupti::Pti_Activity *activity) {
-  auto *externalActivity =
-      reinterpret_cast<pti_view_record_external_correlation *>(activity);
-  std::cout << "processActivityExternalCorrelation: _correlation_id: "
-            << externalActivity->_correlation_id << "\n";
-  std::cout << "processActivityExternalCorrelation: _external_id: "
-            << externalActivity->_external_id << "\n";
-
-  // corrIdToExternId[externalActivity->_correlation_id] =
-  // {externalActivity->_external_id, 1};
-  return externalActivity->_correlation_id;
-}
-
 uint32_t processActivity(XpuptiProfiler::CorrIdToExternIdMap &corrIdToExternId,
                          XpuptiProfiler::ApiExternIdSet &apiExternIds,
                          std::set<Data *> &dataSet,
@@ -159,11 +125,6 @@ uint32_t processActivity(XpuptiProfiler::CorrIdToExternIdMap &corrIdToExternId,
                                           dataSet, activity);
     break;
   }
-  case PTI_VIEW_EXTERNAL_CORRELATION: {
-    // correlationId = processActivityExternalCorrelation(corrIdToExternId,
-    // activity);
-    break;
-  }
   default:
     break;
   }
@@ -253,8 +214,6 @@ void XpuptiProfiler::XpuptiProfilerPimpl::completeBuffer(uint8_t *buffer,
   do {
     status = xpupti::viewGetNextRecord<true>(buffer, validSize, &activity);
     if (status == pti_result::PTI_SUCCESS) {
-      std::cout << "activity->_view_kind: " << activity->_view_kind << "\n"
-                << std::flush;
       auto correlationId =
           processActivity(profiler.correlation.corrIdToExternId,
                           profiler.correlation.apiExternIds, dataSet, activity);
@@ -278,7 +237,6 @@ void XpuptiProfiler::XpuptiProfilerPimpl::callbackFn(
     pti_callback_domain domain, pti_api_group_id driver_api_group_id,
     uint32_t driver_api_id, pti_backend_ctx_t backend_context, void *cb_data,
     void *global_user_data, void **instance_user_data) {
-  std::cout << "callback\n" << std::flush;
   pti_callback_gpu_op_data *callback_data =
       static_cast<pti_callback_gpu_op_data *>(cb_data);
   if (callback_data == nullptr) {
@@ -377,25 +335,13 @@ void XpuptiProfiler::XpuptiProfilerPimpl::doStart() {
   if (profiler.utils_cache_path != "") {
     callEnumDeviceUUIDs(profiler.utils_cache_path);
   }
-  // auto res = ptiViewPushExternalCorrelationId(
-  //     pti_view_external_kind::PTI_VIEW_EXTERNAL_KIND_CUSTOM_1, 42);
-  //  std::cout << "res: " << res << "\n" << std::flush;
-  /*
-  ze_result_t status = ZE_RESULT_SUCCESS;
-  // status = zeInit(ZE_INIT_FLAG_GPU_ONLY);
-  // assert(status == ZE_RESULT_SUCCESS);
-  */
 
   xpupti::viewSetCallbacks<true>(allocBuffer, completeBuffer);
   xpupti::viewEnable<true>(PTI_VIEW_DEVICE_GPU_KERNEL);
   xpupti::viewEnable<true>(PTI_VIEW_DEVICE_GPU_MEM_FILL);
   xpupti::viewEnable<true>(PTI_VIEW_DEVICE_GPU_MEM_COPY);
   xpupti::subscribe<true>(&subscriber, callbackFn, &subscriber);
-  // xpupti::viewEnable<true>(PTI_VIEW_DEVICE_GPU_MEM_COPY);
-  // xpupti::viewEnable<true>(PTI_VIEW_DEVICE_GPU_MEM_FILL);
   // xpupti::viewEnable<true>(PTI_VIEW_SYCL_RUNTIME_CALLS);
-  // xpupti::viewEnable<true>(PTI_VIEW_COLLECTION_OVERHEAD);
-  // xpupti::viewEnable<true>(PTI_VIEW_EXTERNAL_CORRELATION);
   // xpupti::viewEnable<true>(PTI_VIEW_LEVEL_ZERO_CALLS);
   // setGraphCallbacks(subscriber, /*enable=*/true);
   // setRuntimeCallbacks(subscriber, /*enable=*/true);
@@ -405,7 +351,6 @@ void XpuptiProfiler::XpuptiProfilerPimpl::doStart() {
 }
 
 void XpuptiProfiler::XpuptiProfilerPimpl::doFlush() {
-  std::cout << "flush\n" << std::flush;
   XpuptiProfiler &profiler = threadState.profiler;
   if (profiler.syclQueue != nullptr) {
     callWaitOnSyclQueue(profiler.utils_cache_path, profiler.syclQueue);
@@ -420,11 +365,7 @@ void XpuptiProfiler::XpuptiProfilerPimpl::doStop() {
   xpupti::viewDisable<true>(PTI_VIEW_DEVICE_GPU_KERNEL);
   xpupti::viewDisable<true>(PTI_VIEW_DEVICE_GPU_MEM_FILL);
   xpupti::viewDisable<true>(PTI_VIEW_DEVICE_GPU_MEM_COPY);
-  // xpupti::viewDisable<true>(PTI_VIEW_DEVICE_GPU_MEM_COPY);
-  // xpupti::viewDisable<true>(PTI_VIEW_DEVICE_GPU_MEM_FILL);
   // xpupti::viewDisable<true>(PTI_VIEW_SYCL_RUNTIME_CALLS);
-  // xpupti::viewDisable<true>(PTI_VIEW_COLLECTION_OVERHEAD);
-  // xpupti::viewDisable<true>(PTI_VIEW_EXTERNAL_CORRELATION);
   // xpupti::viewDisable<true>(PTI_VIEW_LEVEL_ZERO_CALLS);
   // setGraphCallbacks(subscriber, /*enable=*/false);
   // setRuntimeCallbacks(subscriber, /*enable=*/false);