SWDEV-569101 - increase signal list size to at least DEBUG_HIP_GRAPH_BATCH_SIZE

Author: iassiour

projects/clr/rocclr/device/rocm/rocvirtual.cpp

@@ -396,7 +396,7 @@ bool VirtualGPU::HwQueueTracker::CreateSignal(ProfilingSignal* signal, bool inte
 
 // ================================================================================================
 bool VirtualGPU::HwQueueTracker::Create() {
-  const uint kSignalListSize = ROC_SIGNAL_POOL_SIZE;
+  const uint kSignalListSize = std::max(ROC_SIGNAL_POOL_SIZE, DEBUG_HIP_GRAPH_BATCH_SIZE);
   signal_list_.resize(kSignalListSize);
   for (uint i = 0; i < kSignalListSize; ++i) {
     std::unique_ptr<ProfilingSignal> signal(new ProfilingSignal());
@@ -436,8 +436,8 @@ hsa_signal_t VirtualGPU::HwQueueTracker::ActiveSignal(hsa_signal_value_t init_va
 
   bool new_signal = false;
 
-  // Peep signal +2 ahead to see if its done
-  auto temp_id = (current_id_ + 2) % signal_list_.size();
+  // Peep signal +DEBUG_HIP_GRAPH_BATCH_SIZE ahead to see if its done
+  auto temp_id = (current_id_ + DEBUG_HIP_GRAPH_BATCH_SIZE) % signal_list_.size();
 
   // If GPU is still busy with processing, then add more signals to avoid more frequent stalls
   if (Hsa::signal_load_relaxed(signal_list_[temp_id]->signal_) > 0) {

projects/clr/rocclr/utils/flags.hpp

@@ -221,7 +221,7 @@ release(uint, ROC_P2P_SDMA_SIZE, 1024,                                        \
         "The minimum size in KB for P2P transfer with SDMA")                  \
 release(uint, ROC_AQL_QUEUE_SIZE, 16384,                                      \
         "AQL queue size in AQL packets")                                      \
-release(uint, ROC_SIGNAL_POOL_SIZE, 64,                                       \
+release(uint, ROC_SIGNAL_POOL_SIZE, 256,                                      \
         "Initial size of HSA signal pool")                                    \
 release(uint, DEBUG_CLR_LIMIT_BLIT_WG, 16,                                    \
         "Limit the number of workgroups in blit operations")                  \