diff options
Diffstat (limited to 'src/stream_manager.cc')
| -rw-r--r-- | src/stream_manager.cc | 40 |
1 files changed, 33 insertions, 7 deletions
diff --git a/src/stream_manager.cc b/src/stream_manager.cc index bb95bb1..d43964a 100644 --- a/src/stream_manager.cc +++ b/src/stream_manager.cc @@ -34,15 +34,20 @@ unsigned CUstream_st::sm_next_stream_uid = 0; -// SST memcpy callbacks -extern void SST_callback_memcpy_H2D_done(); -extern void SST_callback_memcpy_D2H_done(); +// SST memcpy callbacks, called after a stream operation is done via record_next_done() +extern void SST_callback_memcpy_H2D_done(uint64_t dst, uint64_t src, size_t count, cudaStream_t stream); +extern void SST_callback_memcpy_D2H_done(uint64_t dst, uint64_t src, size_t count, cudaStream_t stream); extern void SST_callback_memcpy_to_symbol_done(); extern void SST_callback_memcpy_from_symbol_done(); -__attribute__((weak)) void SST_callback_memcpy_H2D_done() {} -__attribute__((weak)) void SST_callback_memcpy_D2H_done() {} +extern void SST_callback_cudaEventSynchronize_done(cudaEvent_t event); +extern void SST_callback_kernel_done(cudaStream_t stream); +__attribute__((weak)) void SST_callback_memcpy_H2D_done(uint64_t dst, uint64_t src, size_t count, cudaStream_t stream) {} +__attribute__((weak)) void SST_callback_memcpy_D2H_done(uint64_t dst, uint64_t src, size_t count, cudaStream_t stream) {} __attribute__((weak)) void SST_callback_memcpy_to_symbol_done() {} __attribute__((weak)) void SST_callback_memcpy_from_symbol_done() {} +__attribute__((weak)) void SST_callback_cudaEventSynchronize_done(cudaEvent_t event); +__attribute__((weak)) void SST_callback_kernel_done(cudaStream_t stream); + CUstream_st::CUstream_st() { m_pending = false; @@ -74,6 +79,10 @@ void CUstream_st::synchronize() { } while (!done); } +bool CUstream_st::synchronize_check() { + return m_operations.empty(); +} + void CUstream_st::push(const stream_operation &op) { // called by host thread pthread_mutex_lock(&m_lock); @@ -132,13 +141,15 @@ bool stream_operation::do_operation(gpgpu_sim *gpu) { if (g_debug_execution >= 3) printf("memcpy host-to-device\n"); gpu->memcpy_to_gpu(m_device_address_dst, m_host_address_src, m_cnt); m_stream->record_next_done(); - if (gpu->is_SST_mode()) SST_callback_memcpy_H2D_done(); + if (gpu->is_SST_mode()) { + SST_callback_memcpy_H2D_done((uint64_t) m_device_address_dst, (uint64_t) m_host_address_src, m_cnt, m_stream->is_stream_zero_stream() ? 0 : m_stream); + } break; case stream_memcpy_device_to_host: if (g_debug_execution >= 3) printf("memcpy device-to-host\n"); gpu->memcpy_from_gpu(m_host_address_dst, m_device_address_src, m_cnt); m_stream->record_next_done(); - if (gpu->is_SST_mode()) SST_callback_memcpy_D2H_done(); + if (gpu->is_SST_mode()) SST_callback_memcpy_D2H_done((uint64_t) m_host_address_dst, (uint64_t) m_device_address_src, m_cnt, m_stream->is_stream_zero_stream() ? 0 : m_stream); break; case stream_memcpy_device_to_device: if (g_debug_execution >= 3) printf("memcpy device-to-device\n"); @@ -194,6 +205,13 @@ bool stream_operation::do_operation(gpgpu_sim *gpu) { time_t wallclock = time((time_t *)NULL); m_event->update(gpu->gpu_tot_sim_cycle, wallclock); m_stream->record_next_done(); + if ((gpu->is_SST_mode()) && m_event->done() && + m_event->requested_synchronize()) { + // Notify that the event is done + SST_callback_cudaEventSynchronize_done(m_event); + // Reset the sync flag as we have notified SST + m_event->reset_request_synchronize(); + } } break; case stream_wait_event: // only allows next op to go if event is done @@ -252,6 +270,9 @@ stream_manager::stream_manager(gpgpu_sim *gpu, bool cuda_launch_blocking) { m_cuda_launch_blocking = cuda_launch_blocking; pthread_mutex_init(&m_lock, NULL); m_last_stream = m_streams.begin(); + + // Mark stream zero as the default stream + m_stream_zero.set_stream_zero(); } bool stream_manager::operation(bool *sim) { @@ -303,6 +324,11 @@ bool stream_manager::register_finished_kernel(unsigned grid_uid) { // grid_uid, stream->get_uid()); kernel_stat.flush(); // kernel_stat.close(); stream->record_next_done(); + // Callback to notify a kernel is done for SST's stream + // manager to support with nonblocking + blocking kernel launch + if (m_gpu->is_SST_mode()) { + SST_callback_kernel_done(stream->is_stream_zero_stream() ? 0 : stream); + } m_grid_id_to_stream.erase(grid_uid); kernel->notify_parent_finished(); delete kernel; |