diff options
| -rw-r--r-- | libcuda/cuda_runtime_api.cc | 10 | ||||
| -rw-r--r-- | src/stream_manager.cc | 58 | ||||
| -rw-r--r-- | src/stream_manager.h | 17 |
3 files changed, 40 insertions, 45 deletions
diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 7005eef..6a7826e 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1709,13 +1709,13 @@ __host__ cudaError_t CUDARTAPI cudaStreamWaitEvent(cudaStream_t stream, cudaEven //reference: https://www.cs.cmu.edu/afs/cs/academic/class/15668-s11/www/cuda-doc/html/group__CUDART__STREAM_gfe68d207dc965685d92d3f03d77b0876.html CUevent_st *e = get_event(event); if( !e ){ - printf("GPGPU-Sim API: Error at cudaStreamWaitEvent. Event is not created .\n"); - return g_last_cudaError = cudaErrorInvalidResourceHandle; + printf("GPGPU-Sim API: Error at cudaStreamWaitEvent. Event is not created .\n"); + return g_last_cudaError = cudaErrorInvalidResourceHandle; } - else if(e->num_issued() == 0){ - printf("GPGPU-Sim API: Warning: cudaEventRecord has not been called on event before calling cudaStreamWaitEvent.\nNothing to be done.\n"); + else if(e->num_issued() == 0){ + printf("GPGPU-Sim API: Warning: cudaEventRecord has not been called on event before calling cudaStreamWaitEvent.\nNothing to be done.\n"); return g_last_cudaError = cudaSuccess; - } + } if (!stream){ g_stream_manager->pushCudaStreamWaitEventToAllStreams(e, flags); } else { diff --git a/src/stream_manager.cc b/src/stream_manager.cc index 6e02dcd..d0eecdd 100644 --- a/src/stream_manager.cc +++ b/src/stream_manager.cc @@ -198,13 +198,13 @@ bool stream_operation::do_operation( gpgpu_sim *gpu ) //only allows next op to go if event is done //otherwise stays in the stream queue printf("stream wait event processing...\n"); - if(m_event->done()){ + if(m_event->done()){ printf("stream wait event done\n"); m_stream->record_next_done(); } - else{ - return false; - } + else{ + return false; + } break; default: abort(); @@ -235,7 +235,7 @@ stream_manager::stream_manager( gpgpu_sim *gpu, bool cuda_launch_blocking ) m_service_stream_zero = false; m_cuda_launch_blocking = cuda_launch_blocking; pthread_mutex_init(&m_lock,NULL); - m_last_stream = m_streams.begin(); + m_last_stream = m_streams.begin(); } bool stream_manager::operation( bool * sim) @@ -336,29 +336,25 @@ stream_operation stream_manager::front() } if(!m_service_stream_zero) { - std::list<struct CUstream_st*>::iterator s = m_last_stream; - if(m_last_stream == m_streams.end()){ - s = m_streams.begin(); - } - else{ - s++; - } - for(size_t ii = 0 ; ii < m_streams.size(); ii++, s++) { - if(s == m_streams.end()){ - s = m_streams.begin(); - } - m_last_stream = s; - CUstream_st *stream = *s; - if( !stream->busy() && !stream->empty() ) { - result = stream->next(); - if( result.is_kernel() ) { - unsigned grid_id = result.get_kernel()->get_uid(); - m_grid_id_to_stream[grid_id] = stream; - } - break; - } - } - } + std::list<struct CUstream_st*>::iterator s = m_last_stream; + if(m_last_stream == m_streams.end()){ s = m_streams.begin(); } + else{ s++; } + for(size_t ii = 0 ; ii < m_streams.size(); ii++, s++) { + if(s == m_streams.end()){ + s = m_streams.begin(); + } + m_last_stream = s; + CUstream_st *stream = *s; + if( !stream->busy() && !stream->empty() ) { + result = stream->next(); + if( result.is_kernel() ) { + unsigned grid_id = result.get_kernel()->get_uid(); + m_grid_id_to_stream[grid_id] = stream; + } + break; + } + } + } return result; } @@ -383,8 +379,8 @@ void stream_manager::destroy_stream( CUstream_st *stream ) break; } } - delete stream; - m_last_stream = m_streams.begin(); + delete stream; + m_last_stream = m_streams.begin(); pthread_mutex_unlock(&m_lock); } @@ -398,6 +394,7 @@ bool stream_manager::concurrent_streams_empty() for( s=m_streams.begin(); s!=m_streams.end();++s ) { struct CUstream_st *stream = *s; if( !stream->empty() ) { + //stream->print(stdout); result = false; break; } @@ -450,6 +447,7 @@ void stream_manager::print_impl( FILE *fp) void stream_manager::push( stream_operation op ) { struct CUstream_st *stream = op.get_stream(); + // block if stream 0 (or concurrency disabled) and pending concurrent operations exist bool block= !stream || m_cuda_launch_blocking; while(block) { diff --git a/src/stream_manager.h b/src/stream_manager.h index ae1af9f..eb4c6ff 100644 --- a/src/stream_manager.h +++ b/src/stream_manager.h @@ -163,6 +163,7 @@ public: void print( FILE *fp ) const; struct CUstream_st *get_stream() { return m_stream; } void set_stream( CUstream_st *stream ) { m_stream = stream; } + private: struct CUstream_st *m_stream; @@ -178,9 +179,9 @@ private: const char *m_symbol; size_t m_offset; - struct CUevent_st *m_event; bool m_sim_mode; kernel_info_t *m_kernel; + struct CUevent_st *m_event; }; struct CUevent_st { @@ -192,7 +193,7 @@ public: m_updates = 0; m_wallclock = 0; m_gpu_tot_sim_cycle = 0; - m_issued = 0; + m_issued = 0; m_done = false; } void update( double cycle, time_t clk ) @@ -207,18 +208,14 @@ public: unsigned num_updates() const { return m_updates; } bool done() const { return m_done; } time_t clock() const { return m_wallclock; } - void issue(){ - m_issued++; - } - unsigned int num_issued() const{ - return m_issued; - } + void issue(){ m_issued++; } + unsigned int num_issued() const{ return m_issued; } private: int m_uid; bool m_blocking; bool m_done; int m_updates; - unsigned int m_issued; + unsigned int m_issued; time_t m_wallclock; double m_gpu_tot_sim_cycle; @@ -275,7 +272,7 @@ private: CUstream_st m_stream_zero; bool m_service_stream_zero; pthread_mutex_t m_lock; - std::list<struct CUstream_st*>::iterator m_last_stream; + std::list<struct CUstream_st*>::iterator m_last_stream; }; #endif |