From 8e9f0375a6a8bed217b0634a08f7f1f0cd149744 Mon Sep 17 00:00:00 2001 From: boyealkim Date: Thu, 25 Jul 2019 20:31:21 +0900 Subject: fix stream fix bug and fix cudaStreamWaitEvent impl --- libcuda/cuda_runtime_api.cc | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 44f0f4e..7005eef 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1696,6 +1696,7 @@ __host__ cudaError_t CUDARTAPI cudaEventRecord(cudaEvent_t event, cudaStream_t s if( !e ) return g_last_cudaError = cudaErrorUnknown; struct CUstream_st *s = (struct CUstream_st *)stream; stream_operation op(e,s); + e->issue(); g_stream_manager->push(op); return g_last_cudaError = cudaSuccess; } @@ -1708,9 +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: Warning: cudaEventRecord has not been called on event before calling cudaStreamWaitEvent.\nNothing to be done.\n"); - return g_last_cudaError = cudaSuccess; + 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"); + return g_last_cudaError = cudaSuccess; + } if (!stream){ g_stream_manager->pushCudaStreamWaitEventToAllStreams(e, flags); } else { -- cgit v1.3 From b13557baa4ea6d1084bdcc9b28b9eb660c394427 Mon Sep 17 00:00:00 2001 From: boyealkim Date: Fri, 26 Jul 2019 11:55:45 +0900 Subject: fix indentation --- libcuda/cuda_runtime_api.cc | 10 ++++---- src/stream_manager.cc | 58 ++++++++++++++++++++++----------------------- src/stream_manager.h | 17 ++++++------- 3 files changed, 40 insertions(+), 45 deletions(-) (limited to 'libcuda') 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::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::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::iterator m_last_stream; + std::list::iterator m_last_stream; }; #endif -- cgit v1.3 From 555cce395bb8457a209b26f6f7ff62cc09839fdf Mon Sep 17 00:00:00 2001 From: bftf Date: Mon, 11 Nov 2019 12:29:15 -0800 Subject: Enabled CUDA 10. Added implementations for __cudaPushCallConfiguration and __cudaPopCallConfiguration, neither of which are documented in the CUDA API but they preceed any cudaLaunchKernel() call and push the call configuration on a stack which is then popped when cudaLaunchKernel() is called. Currently the max stack size is 1, an assert will be triggered when this is violated. --- CHANGES | 1 + Makefile | 1 + libcuda/cuda_api.h | 3 ++- libcuda/cuda_runtime_api.cc | 54 +++++++++++++++++++++++++++++++++++++++++++-- linux-so-version.txt | 2 ++ setup_environment | 5 +++-- 6 files changed, 61 insertions(+), 5 deletions(-) (limited to 'libcuda') diff --git a/CHANGES b/CHANGES index 2a4222b..31ae672 100644 --- a/CHANGES +++ b/CHANGES @@ -37,6 +37,7 @@ Version 4.0.0 (development branch) versus 3.2.3 3- Addig new system stats: gpu occupancy, L2BW, etc -Library: 1 Enabled CUTLASS Library on GPGPU-Sim +2 Enabled CUDA 10 -Regression: 1- Added TensorCore Regression Kernel -Configs: diff --git a/Makefile b/Makefile index 3db8ce8..8df3b57 100644 --- a/Makefile +++ b/Makefile @@ -164,6 +164,7 @@ $(SIM_LIB_DIR)/libcudart.so: makedirs $(LIBS) cudalib if [ ! -f $(SIM_LIB_DIR)/libcudart.so.9.0 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.9.0; fi if [ ! -f $(SIM_LIB_DIR)/libcudart.so.9.1 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.9.1; fi if [ ! -f $(SIM_LIB_DIR)/libcudart.so.9.2 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.9.2; fi + if [ ! -f $(SIM_LIB_DIR)/libcudart.so.10.0 ]; then ln -s libcudart.so $(SIM_LIB_DIR)/libcudart.so.10.0; fi $(SIM_LIB_DIR)/libcudart.dylib: makedirs $(LIBS) cudalib g++ -dynamiclib -Wl,-headerpad_max_install_names,-undefined,dynamic_lookup,-compatibility_version,1.1,-current_version,1.1\ diff --git a/libcuda/cuda_api.h b/libcuda/cuda_api.h index 3808e8a..b304a59 100644 --- a/libcuda/cuda_api.h +++ b/libcuda/cuda_api.h @@ -234,10 +234,11 @@ typedef struct CUgraphicsResource_st *CUgraphicsResource; /**< CUDA graphics int typedef unsigned long long CUtexObject; /**< An opaque value that represents a CUDA texture object */ typedef unsigned long long CUsurfObject; /**< An opaque value that represents a CUDA surface object */ +#if CUDA_VERSION < 1000 typedef struct CUuuid_st { /**< CUDA definition of UUID */ char bytes[16]; } CUuuid; - +#endif #if __CUDA_API_VERSION >= 4010 diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 44f0f4e..fd092e5 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1448,6 +1448,50 @@ __host__ cudaError_t CUDARTAPI cudaConfigureCall(dim3 gridDim, dim3 blockDim, si return g_last_cudaError = cudaSuccess; } + +#if CUDA_VERSION >= 1000 +/* +* CUDA 10 requires a new CUDA kernel launch sequence +* A call to __cudaPushCallConfiguration() preceeds any call to cudaLaunchKernel() +* __cudaPushCallConfiguration is undocumented in the API but it simply sets up a buffer with the arguments which is accessed in cudaLaunchKernel() +* __cudaPopCallConfiguration is undocumented in the API but it simply pops the configuration set in cudaLaunchKernel() +* +* pushing more than 1 configuration without popping is currently not implemented in GPGPU-Sim and will result in an assert error +*/ +namespace g_cudaPushArgsBuffer +{ + bool g_is_initialized = false; + dim3 g_gridDim; + dim3 g_blockDim; + size_t g_sharedMem; + cudaStream_t g_stream; +} + +__host__ cudaError_t CUDARTAPI __cudaPushCallConfiguration(dim3 gridDim, dim3 blockDim, size_t sharedMem, cudaStream_t stream) +{ + assert(g_cudaPushArgsBuffer::g_is_initialized == false); + printf("Pushing cuda call configuration \n"); + g_cudaPushArgsBuffer::g_is_initialized = true; + g_cudaPushArgsBuffer::g_gridDim = gridDim; + g_cudaPushArgsBuffer::g_blockDim = blockDim; + g_cudaPushArgsBuffer::g_sharedMem = sharedMem; + g_cudaPushArgsBuffer::g_stream = stream; + + return cudaSuccess; +} + +__host__ cudaError_t CUDARTAPI __cudaPopCallConfiguration() +{ + printf("Inside __cudaPopCallConfiguration\n"); + assert(g_cudaPushArgsBuffer::g_is_initialized == true); + printf("Poping cuda call configuration \n"); + g_cudaPushArgsBuffer::g_is_initialized = false; + return cudaSuccess; +} + +#endif // #if CUDA_VERSION >= 1000 + + __host__ cudaError_t CUDARTAPI cudaSetupArgument(const void *arg, size_t size, size_t offset) { if(g_debug_execution >= 3){ @@ -1551,8 +1595,14 @@ __host__ cudaError_t CUDARTAPI cudaLaunchKernel ( const char* hostFun, dim3 grid } CUctx_st *context = GPGPUSim_Context(); function_info *entry = context->get_kernel(hostFun); - - cudaConfigureCall(gridDim, blockDim, sharedMem, stream); + +#if CUDA_VERSION >= 1000 + assert(g_cudaPushArgsBuffer::g_is_initialized == false); + cudaConfigureCall(g_cudaPushArgsBuffer::g_gridDim, g_cudaPushArgsBuffer::g_blockDim, g_cudaPushArgsBuffer::g_sharedMem, g_cudaPushArgsBuffer::g_stream); +#else + cudaConfigureCall(gridDim, blockDim, sharedMem, stream); +#endif // #if CUDA_VERSION >= 1000 + for(unsigned i = 0; i < entry->num_args(); i++){ std::pair p = entry->get_param_config(i); cudaSetupArgument(args[i], p.first, p.second); diff --git a/linux-so-version.txt b/linux-so-version.txt index a7c2d3c..357c6cd 100644 --- a/linux-so-version.txt +++ b/linux-so-version.txt @@ -6,3 +6,5 @@ libcudart.so.9.2{ }; libcuda.so.1{ }; +libcudart.so.10.0{ +}; \ No newline at end of file diff --git a/setup_environment b/setup_environment index b420584..154dea0 100644 --- a/setup_environment +++ b/setup_environment @@ -51,9 +51,10 @@ CC_VERSION=`gcc --version | head -1 | awk '{for(i=1;i<=NF;i++){ if(match($i,/^[0 CUDA_VERSION_STRING=`$CUDA_INSTALL_PATH/bin/nvcc --version | awk '/release/ {print $5;}' | sed 's/,//'`; export CUDA_VERSION_NUMBER=`echo $CUDA_VERSION_STRING | sed 's/\./ /' | awk '{printf("%02u%02u", 10*int($1), 10*$2);}'` -if [ $CUDA_VERSION_NUMBER -gt 9100 -o $CUDA_VERSION_NUMBER -lt 2030 ]; then +if [ $CUDA_VERSION_NUMBER -gt 10000 -o $CUDA_VERSION_NUMBER -lt 2030 ]; then echo "ERROR ** GPGPU-Sim version $GPGPUSIM_VERSION_STRING not tested with CUDA version $CUDA_VERSION_STRING (please see README)"; - return + echo $CUDA_VERSION_NUMBER + return fi if [ $CUDA_VERSION_NUMBER -ge 6000 ]; then -- cgit v1.3 From 57d2df04da439b1055590f10cff5f5f50791b9d9 Mon Sep 17 00:00:00 2001 From: bftf Date: Tue, 12 Nov 2019 19:14:16 -0800 Subject: Updated the ifdef flag which enables/disables the relevant code blocks for cuda 10; while assuring backwards compatibility for older cuda versions --- libcuda/cuda_api.h | 4 ++-- libcuda/cuda_runtime_api.cc | 10 +++++----- 2 files changed, 7 insertions(+), 7 deletions(-) (limited to 'libcuda') diff --git a/libcuda/cuda_api.h b/libcuda/cuda_api.h index b304a59..0ded242 100644 --- a/libcuda/cuda_api.h +++ b/libcuda/cuda_api.h @@ -234,11 +234,11 @@ typedef struct CUgraphicsResource_st *CUgraphicsResource; /**< CUDA graphics int typedef unsigned long long CUtexObject; /**< An opaque value that represents a CUDA texture object */ typedef unsigned long long CUsurfObject; /**< An opaque value that represents a CUDA surface object */ -#if CUDA_VERSION < 1000 +#if CUDART_VERSION < 10000 typedef struct CUuuid_st { /**< CUDA definition of UUID */ char bytes[16]; } CUuuid; -#endif +#endif // #if CUDART_VERSION < 10000 #if __CUDA_API_VERSION >= 4010 diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index fd092e5..71b8722 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1449,7 +1449,7 @@ __host__ cudaError_t CUDARTAPI cudaConfigureCall(dim3 gridDim, dim3 blockDim, si } -#if CUDA_VERSION >= 1000 +#if CUDART_VERSION >= 10000 /* * CUDA 10 requires a new CUDA kernel launch sequence * A call to __cudaPushCallConfiguration() preceeds any call to cudaLaunchKernel() @@ -1489,7 +1489,7 @@ __host__ cudaError_t CUDARTAPI __cudaPopCallConfiguration() return cudaSuccess; } -#endif // #if CUDA_VERSION >= 1000 +#endif // #if CUDART_VERSION >= 10000 __host__ cudaError_t CUDARTAPI cudaSetupArgument(const void *arg, size_t size, size_t offset) @@ -1596,12 +1596,12 @@ __host__ cudaError_t CUDARTAPI cudaLaunchKernel ( const char* hostFun, dim3 grid CUctx_st *context = GPGPUSim_Context(); function_info *entry = context->get_kernel(hostFun); -#if CUDA_VERSION >= 1000 +#if CUDART_VERSION >= 10000 assert(g_cudaPushArgsBuffer::g_is_initialized == false); cudaConfigureCall(g_cudaPushArgsBuffer::g_gridDim, g_cudaPushArgsBuffer::g_blockDim, g_cudaPushArgsBuffer::g_sharedMem, g_cudaPushArgsBuffer::g_stream); #else - cudaConfigureCall(gridDim, blockDim, sharedMem, stream); -#endif // #if CUDA_VERSION >= 1000 + cudaConfigureCall(gridDim, blockDim, sharedMem, stream); +#endif // #if CUDART_VERSION >= 10000 for(unsigned i = 0; i < entry->num_args(); i++){ std::pair p = entry->get_param_config(i); -- cgit v1.3