Add NVTX ranges for Operators run (#73)

* Add NVTX ranges for Operators run Signed-off-by: Serge Panev <spanev@nvidia.com> * Add colors for NVTX and TimeRange for Support Operators Signed-off-by: Serge Panev <spanev@nvidia.com>
NVIDIA · Jul 28, 2018 · 22f6527 · 22f6527
1 parent 400d637
commit 22f6527
Show file tree

Hide file tree

Showing 2 changed files with 20 additions and 6 deletions.
diff --git a/dali/common.h b/dali/common.h
@@ -142,6 +142,7 @@ struct TimeRange {
   static const uint32_t kBlue1   = 0x268BD2;
   static const uint32_t kCyan    = 0x2AA198;
   static const uint32_t kGreen1  = 0x859900;
+  static const uint32_t knvGreen = 0x76B900;
 
   TimeRange(std::string name, const uint32_t rgb = kBlue) { // NOLINT
 #ifdef DALI_USE_NVTX

diff --git a/dali/pipeline/executor/executor.cc b/dali/pipeline/executor/executor.cc
@@ -74,8 +74,11 @@ void Executor::RunCPU() {
   try {
     WorkspaceBlob &wsb = wss_[queue_idx];
     for (int i = 0; i < graph_->NumSupportOp(); ++i) {
-      OperatorBase &op = graph_->support_op(i);
+      OpNode &op_node = graph_->support_node(i);
+      OperatorBase &op = *op_node.op;
       SupportWorkspace &ws = wsb.support_op_data[i];
+      TimeRange tr("[Executor] Run Support op " + op_node.instance_name,
+          TimeRange::kCyan);
       op.Run(&ws);
     }
   } catch (std::runtime_error &e) {
@@ -94,9 +97,13 @@ void Executor::RunCPU() {
             TimeRange tr("[Executor] RunCPU on " + to_string(data_idx));
             SampleWorkspace ws;
             for (int j = 0; j < graph_->NumCPUOp(); ++j) {
-            OperatorBase &op = graph_->cpu_op(j);
-            wsb.cpu_op_data[j].GetSample(&ws, data_idx, tid);
-            op.Run(&ws);
+              OpNode &op_node = graph_->cpu_node(j);
+              OperatorBase &op = *op_node.op;
+              wsb.cpu_op_data[j].GetSample(&ws, data_idx, tid);
+              TimeRange tr("[Executor] Run CPU op " + op_node.instance_name
+                  + " on " + to_string(data_idx),
+                  TimeRange::kBlue1);
+              op.Run(&ws);
             }
             }, i, std::placeholders::_1));
     }
@@ -129,8 +136,11 @@ void Executor::RunMixed() {
 
   try {
     for (int i = 0; i < graph_->NumMixedOp(); ++i) {
-      OperatorBase &op = graph_->mixed_op(i);
+      OpNode &op_node = graph_->mixed_node(i);
+      OperatorBase &op = *op_node.op;
       MixedWorkspace &ws = wsb.mixed_op_data[i];
+      TimeRange tr("[Executor] Run Mixed op " + op_node.instance_name,
+          TimeRange::kOrange);
       op.Run(&ws);
       if (ws.has_stream() && ws.has_event()) {
         CUDA_CALL(cudaEventRecord(ws.event(), ws.stream()));
@@ -172,14 +182,17 @@ void Executor::RunGPU() {
   try {
     WorkspaceBlob &wsb = wss_[queue_idx];
     for (int i = 0; i < graph_->NumGPUOp(); ++i) {
-      OperatorBase &op = graph_->gpu_op(i);
+      OpNode &op_node = graph_->gpu_node(i);
+      OperatorBase &op = *op_node.op;
       DeviceWorkspace &ws = wsb.gpu_op_data[i];
       auto parent_events = ws.ParentEvents();
 
       for (auto &event : parent_events) {
         CUDA_CALL(cudaStreamWaitEvent(ws.stream(), event, 0));
       }
 
+      TimeRange tr("[Executor] Run GPU op " + op_node.instance_name,
+          TimeRange::knvGreen);
       op.Run(&ws);
       if (ws.has_event()) {
         CUDA_CALL(cudaEventRecord(ws.event(), ws.stream()));