ADD: add kernel launching latency from stream to distributor

5 files changed, 37 insertions, 14 deletions

diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc
index be5c5b9..819ad35 100644
--- a/src/abstract_hardware_model.cc
+++ b/src/abstract_hardware_model.cc
@@ -550,6 +550,8 @@ void warp_inst_t::completed( unsigned long long cycle ) const
    ptx_file_line_stats_add_latency(pc, latency * active_count());  
 }
 
+//Jin: kernel launch latency
+unsigned g_kernel_launch_latency;
 
 unsigned kernel_info_t::m_next_uid = 1;
 
@@ -569,6 +571,8 @@ kernel_info_t::kernel_info_t( dim3 gridDim, dim3 blockDim, class function_info *
     //Jin: parent and child kernel management for CDP
     m_parent_kernel = NULL;
    
+    //Jin: launch latency management
+    m_launch_latency = g_kernel_launch_latency;
 }
 
 kernel_info_t::~kernel_info_t()
diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h
index e6ef521..ab77d63 100644
--- a/src/abstract_hardware_model.h
+++ b/src/abstract_hardware_model.h
@@ -298,6 +298,7 @@ public:
    unsigned long long launch_cycle;
    unsigned long long start_cycle;
    unsigned long long end_cycle;
+   unsigned m_launch_latency;
 };
 
 struct core_config {
diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc
index a9da1c9..7fb9ab3 100644
--- a/src/gpgpu-sim/gpu-sim.cc
+++ b/src/gpgpu-sim/gpu-sim.cc
@@ -443,6 +443,12 @@ void gpgpu_sim_config::reg_options(option_parser_t opp)
                           &Trace::sampling_memory_partition, "The memory partition which is printed using MEMPART_DPRINTF. Default -1 (i.e. all)",
                           "-1");
    ptx_file_line_stats_options(opp);
+
+    //Jin: kernel launch latency
+    extern unsigned g_kernel_launch_latency;
+    option_parser_register(opp, "-gpgpu_kernel_launch_latency", OPT_INT32, 
+                          &g_kernel_launch_latency, "Kernel launch latency in cycles. Default: 0",
+                          "0");
 }
 
 /////////////////////////////////////////////////////////////////////////////
diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc
index fb17eed..04845e7 100644
--- a/src/gpgpusim_entrypoint.cc
+++ b/src/gpgpusim_entrypoint.cc
@@ -153,6 +153,7 @@ void *gpgpu_sim_thread_concurrent(void*)
            printf("GPGPU-Sim: ** STOP simulation thread (no work) **\n");
            fflush(stdout);
         }
+        g_the_gpu->print_stats();
         if(sim_cycles) {
             g_the_gpu->update_stats();
             print_simulation_time();
diff --git a/src/stream_manager.cc b/src/stream_manager.cc
index 3459c6d..3b6cbd5 100644
--- a/src/stream_manager.cc
+++ b/src/stream_manager.cc
@@ -160,20 +160,31 @@ bool stream_operation::do_operation( gpgpu_sim *gpu )
         m_stream->record_next_done();
         break;
     case stream_kernel_launch:
-        if( gpu->can_start_kernel() ) {
-        	gpu->set_cache_config(m_kernel->name());
-            if(g_debug_execution >= 3)
-        	    printf("kernel %d: \'%s\' transfer to GPU hardware scheduler\n", m_kernel->get_uid(), m_kernel->name().c_str() );
-            m_kernel->print_parent_info();
-            if( m_sim_mode )
-                gpu->functional_launch( m_kernel );
-            else
-                gpu->launch( m_kernel );
+        if( m_sim_mode ) { //Functional Sim
+            if(g_debug_execution >= 3) {
+                printf("kernel %d: \'%s\' transfer to GPU hardware scheduler\n", m_kernel->get_uid(), m_kernel->name().c_str() );
+                m_kernel->print_parent_info();
+            }
+            gpu->set_cache_config(m_kernel->name());
+            gpu->functional_launch( m_kernel );
         }
-        else {
-            if(g_debug_execution >= 3)
-        	    printf("kernel %d: \'%s\' not ready to transfer to GPU hardware scheduler\n", m_kernel->get_uid(), m_kernel->name().c_str() );
-            return false;    
+        else { //Performance Sim
+            if( gpu->can_start_kernel() && m_kernel->m_launch_latency == 0) {
+                if(g_debug_execution >= 3) {
+                    printf("kernel %d: \'%s\' transfer to GPU hardware scheduler\n", m_kernel->get_uid(), m_kernel->name().c_str() );
+                    m_kernel->print_parent_info();
+                }
+                gpu->set_cache_config(m_kernel->name());
+                gpu->launch( m_kernel );
+            }
+            else {
+                if(m_kernel->m_launch_latency)
+                    m_kernel->m_launch_latency--;
+                if(g_debug_execution >= 3)
+                    printf("kernel %d: \'%s\', latency %u not ready to transfer to GPU hardware scheduler\n", 
+                        m_kernel->get_uid(), m_kernel->name().c_str(), m_kernel->m_launch_latency);
+                return false;    
+            }
         }
         break;
     case stream_event: {
@@ -255,7 +266,7 @@ bool stream_manager::register_finished_kernel(unsigned grid_uid)
         //Jin: should check children kernels for CDP
         if(kernel->is_finished()) {
 //            std::ofstream kernel_stat("kernel_stat.txt", std::ofstream::out | std::ofstream::app);
-//            kernel_stat<< " kernel " << grid_uid;
+//            kernel_stat<< " kernel " << grid_uid << ": " << kernel->name();
 //            if(kernel->get_parent())
 //                kernel_stat << ", parent " << kernel->get_parent()->get_uid() <<
 //                ", launch " << kernel->launch_cycle;