From 0c16df3c8b108d8720846bb44b9abcc60ddf42f9 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 15 May 2019 20:16:54 -0400 Subject: make gpu_tot_cycle local variable not global variable --- src/cuda-sim/cuda_device_runtime.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'src/cuda-sim/cuda_device_runtime.cc') diff --git a/src/cuda-sim/cuda_device_runtime.cc b/src/cuda-sim/cuda_device_runtime.cc index 917e7a8..86e8147 100644 --- a/src/cuda-sim/cuda_device_runtime.cc +++ b/src/cuda-sim/cuda_device_runtime.cc @@ -200,7 +200,7 @@ void gpgpusim_cuda_launchDeviceV2(const ptx_instruction * pI, ptx_thread_info * //create child kernel_info_t and index it with parameter_buffer address gpgpu_t* gpu=thread->get_gpu(); device_grid = new kernel_info_t(config.grid_dim, config.block_dim, device_kernel_entry, gpu->getNameArrayMapping(), gpu->getNameInfoMapping()); - device_grid->launch_cycle = gpu_sim_cycle + gpu_tot_sim_cycle; + device_grid->launch_cycle = gpu->gpu_sim_cycle + gpu->gpu_tot_sim_cycle; kernel_info_t & parent_grid = thread->get_kernel(); DEV_RUNTIME_REPORT("child kernel launched by " << parent_grid.name() << ", cta (" << thread->get_ctaid().x << ", " << thread->get_ctaid().y << ", " << thread->get_ctaid().z << -- cgit v1.3 From ed9f0e6b2a99840e9649551825a40a04e236dcd9 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Thu, 30 May 2019 18:28:15 -0400 Subject: adding new values to gpu context --- libcuda/cuda_runtime_api.cc | 42 +++++++++++++++++++------------------ src/abstract_hardware_model.cc | 9 ++++---- src/abstract_hardware_model.h | 2 +- src/cuda-sim/cuda-sim.cc | 36 +++++++++++++++---------------- src/cuda-sim/cuda_device_runtime.cc | 5 +++-- src/gpgpusim_entrypoint.cc | 12 +++++++++-- src/gpgpusim_entrypoint.h | 24 ++++++++++++++++----- 7 files changed, 78 insertions(+), 52 deletions(-) (limited to 'src/cuda-sim/cuda_device_runtime.cc') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index fcd5b07..4990edf 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -185,7 +185,7 @@ struct cudaArray cudaError_t g_last_cudaError = cudaSuccess; -extern stream_manager *g_stream_manager; +//extern stream_manager *g_stream_manager(); void register_ptx_function( const char *name, function_info *impl ) { @@ -375,7 +375,8 @@ private: struct _cuda_device_id *GPGPUSim_Init() { - static _cuda_device_id *the_device = NULL; + //static _cuda_device_id *the_device = NULL; + _cuda_device_id *the_device = GPGPUsim_ctx_ptr()->the_cude_device; if( !the_device ) { gpgpu_sim *the_gpu = gpgpu_ptx_sim_init_perf(); @@ -428,7 +429,8 @@ struct _cuda_device_id *GPGPUSim_Init() static CUctx_st* GPGPUSim_Context() { - static CUctx_st *the_context = NULL; + //static CUctx_st *the_context = NULL; + CUctx_st *the_context = GPGPUsim_ctx_ptr()->the_context; if( the_context == NULL ) { _cuda_device_id *the_gpu = GPGPUSim_Init(); the_context = new CUctx_st(the_gpu); @@ -699,21 +701,21 @@ __host__ cudaError_t CUDARTAPI cudaMemcpy(void *dst, const void *src, size_t cou if(g_debug_execution >= 3) printf("GPGPU-Sim PTX: cudaMemcpy(): devPtr = %p\n", dst); if( kind == cudaMemcpyHostToDevice ) - g_stream_manager->push( stream_operation(src,(size_t)dst,count,0) ); + g_stream_manager()->push( stream_operation(src,(size_t)dst,count,0) ); else if( kind == cudaMemcpyDeviceToHost ) - g_stream_manager->push( stream_operation((size_t)src,dst,count,0) ); + g_stream_manager()->push( stream_operation((size_t)src,dst,count,0) ); else if( kind == cudaMemcpyDeviceToDevice ) - g_stream_manager->push( stream_operation((size_t)src,(size_t)dst,count,0) ); + g_stream_manager()->push( stream_operation((size_t)src,(size_t)dst,count,0) ); else if ( kind == cudaMemcpyDefault ) { if ((size_t)src >= GLOBAL_HEAP_START) { if ((size_t)dst >= GLOBAL_HEAP_START) - g_stream_manager->push( stream_operation((size_t)src,(size_t)dst,count,0) ); // device to device + g_stream_manager()->push( stream_operation((size_t)src,(size_t)dst,count,0) ); // device to device else - g_stream_manager->push( stream_operation((size_t)src,dst,count,0) ); // device to host + g_stream_manager()->push( stream_operation((size_t)src,dst,count,0) ); // device to host } else { if ((size_t)dst >= GLOBAL_HEAP_START) - g_stream_manager->push( stream_operation(src,(size_t)dst,count,0) ); + g_stream_manager()->push( stream_operation(src,(size_t)dst,count,0) ); else { printf("GPGPU-Sim PTX: cudaMemcpy - ERROR : unsupported transfer: host to host\n"); abort(); @@ -855,7 +857,7 @@ __host__ cudaError_t CUDARTAPI cudaMemcpyToSymbol(const char *symbol, const void assert(kind == cudaMemcpyHostToDevice); printf("GPGPU-Sim PTX: cudaMemcpyToSymbol: symbol = %p\n", symbol); //stream_operation( const char *symbol, const void *src, size_t count, size_t offset ) - g_stream_manager->push( stream_operation(src,symbol,count,offset,0) ); + g_stream_manager()->push( stream_operation(src,symbol,count,offset,0) ); //gpgpu_ptx_sim_memcpy_symbol(symbol,src,count,offset,1,context->get_device()->get_gpgpu()); return g_last_cudaError = cudaSuccess; } @@ -869,7 +871,7 @@ __host__ cudaError_t CUDARTAPI cudaMemcpyFromSymbol(void *dst, const char *symbo //CUctx_st *context = GPGPUSim_Context(); assert(kind == cudaMemcpyDeviceToHost); printf("GPGPU-Sim PTX: cudaMemcpyFromSymbol: symbol = %p\n", symbol); - g_stream_manager->push( stream_operation(symbol,dst,count,offset,0) ); + g_stream_manager()->push( stream_operation(symbol,dst,count,offset,0) ); //gpgpu_ptx_sim_memcpy_symbol(symbol,dst,count,offset,0,context->get_device()->get_gpgpu()); return g_last_cudaError = cudaSuccess; } @@ -898,9 +900,9 @@ __host__ cudaError_t CUDARTAPI cudaMemcpyAsync(void *dst, const void *src, size_ } struct CUstream_st *s = (struct CUstream_st *)stream; switch( kind ) { - case cudaMemcpyHostToDevice: g_stream_manager->push( stream_operation(src,(size_t)dst,count,s) ); break; - case cudaMemcpyDeviceToHost: g_stream_manager->push( stream_operation((size_t)src,dst,count,s) ); break; - case cudaMemcpyDeviceToDevice: g_stream_manager->push( stream_operation((size_t)src,(size_t)dst,count,s) ); break; + case cudaMemcpyHostToDevice: g_stream_manager()->push( stream_operation(src,(size_t)dst,count,s) ); break; + case cudaMemcpyDeviceToHost: g_stream_manager()->push( stream_operation((size_t)src,dst,count,s) ); break; + case cudaMemcpyDeviceToDevice: g_stream_manager()->push( stream_operation((size_t)src,(size_t)dst,count,s) ); break; default: abort(); } @@ -1611,7 +1613,7 @@ __host__ cudaError_t CUDARTAPI cudaLaunch( const char *hostFun ) printf("GPGPU-Sim PTX: pushing kernel \'%s\' to stream %u, gridDim= (%u,%u,%u) blockDim = (%u,%u,%u) \n", kname.c_str(), stream?stream->get_uid():0, gridDim.x,gridDim.y,gridDim.z,blockDim.x,blockDim.y,blockDim.z ); stream_operation op(grid,g_ptx_sim_mode,stream); - g_stream_manager->push(op); + g_stream_manager()->push(op); context->g_cuda_launch_stack.pop_back(); return g_last_cudaError = cudaSuccess; } @@ -1650,7 +1652,7 @@ __host__ cudaError_t CUDARTAPI cudaStreamCreate(cudaStream_t *stream) printf("GPGPU-Sim PTX: cudaStreamCreate\n"); #if (CUDART_VERSION >= 3000) *stream = new struct CUstream_st(); - g_stream_manager->add_stream(*stream); + g_stream_manager()->add_stream(*stream); #else *stream = 0; printf("GPGPU-Sim PTX: WARNING: Asynchronous kernel execution not supported (%s)\n", __my_func__); @@ -1689,7 +1691,7 @@ __host__ cudaError_t CUDARTAPI cudaStreamDestroy(cudaStream_t stream) //per-stream synchronization required for application using external libraries without explicit synchronization in the code to //avoid the stream_manager from spinning forever to destroy non-empty streams without making any forward progress. stream->synchronize(); - g_stream_manager->destroy_stream(stream); + g_stream_manager()->destroy_stream(stream); #endif return g_last_cudaError = cudaSuccess; } @@ -1769,7 +1771,7 @@ __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); - g_stream_manager->push(op); + g_stream_manager()->push(op); return g_last_cudaError = cudaSuccess; } @@ -1785,11 +1787,11 @@ __host__ cudaError_t CUDARTAPI cudaStreamWaitEvent(cudaStream_t stream, cudaEven return g_last_cudaError = cudaSuccess; } if (!stream){ - g_stream_manager->pushCudaStreamWaitEventToAllStreams(e, flags); + g_stream_manager()->pushCudaStreamWaitEventToAllStreams(e, flags); } else { struct CUstream_st *s = (struct CUstream_st *)stream; stream_operation op(s,e,flags); - g_stream_manager->push(op); + g_stream_manager()->push(op); } return g_last_cudaError = cudaSuccess; } diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index 63b139e..7755477 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -34,6 +34,7 @@ #include "cuda-sim/cuda-sim.h" #include "gpgpu-sim/gpu-sim.h" #include "option_parser.h" +#include "gpgpusim_entrypoint.h" #include #include #include @@ -771,14 +772,14 @@ void kernel_info_t::notify_parent_finished() { extern unsigned long long g_total_param_size; g_total_param_size -= ((m_kernel_entry->get_args_aligned_size() + 255)/256*256); m_parent_kernel->remove_child(this); - g_stream_manager->register_finished_kernel(m_parent_kernel->get_uid()); + g_stream_manager()->register_finished_kernel(m_parent_kernel->get_uid()); } } CUstream_st * kernel_info_t::create_stream_cta(dim3 ctaid) { assert(get_default_stream_cta(ctaid)); CUstream_st * stream = new CUstream_st(); - g_stream_manager->add_stream(stream); + g_stream_manager()->add_stream(stream); assert(m_cta_streams.find(ctaid) != m_cta_streams.end()); assert(m_cta_streams[ctaid].size() >= 1); //must have default stream m_cta_streams[ctaid].push_back(stream); @@ -794,7 +795,7 @@ CUstream_st * kernel_info_t::get_default_stream_cta(dim3 ctaid) { else { m_cta_streams[ctaid] = std::list(); CUstream_st * stream = new CUstream_st(); - g_stream_manager->add_stream(stream); + g_stream_manager()->add_stream(stream); m_cta_streams[ctaid].push_back(stream); return stream; } @@ -826,7 +827,7 @@ void kernel_info_t::destroy_cta_streams() { for(auto s = m_cta_streams.begin(); s != m_cta_streams.end(); s++) { stream_size += s->second.size(); for(auto ss = s->second.begin(); ss != s->second.end(); ss++) - g_stream_manager->destroy_stream(*ss); + g_stream_manager()->destroy_stream(*ss); s->second.clear(); } printf("size %lu\n", stream_size); diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 1735c2f..77d5f58 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -197,7 +197,7 @@ void increment_x_then_y_then_z( dim3 &i, const dim3 &bound); #include "stream_manager.h" class stream_manager; struct CUstream_st; -extern stream_manager * g_stream_manager; +//extern stream_manager * g_stream_manager; //support for pinned memories added extern std::map pinned_memory; extern std::map pinned_memory_size; diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index e733b7f..a456978 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -448,8 +448,8 @@ void gpgpu_t::memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t coun m_global_mem->write(dst_start_addr+n,1, src_data+n,NULL,NULL); // Copy into the performance model. - extern gpgpu_sim* g_the_gpu; - g_the_gpu->perf_memcpy_to_gpu(dst_start_addr, count); + //extern gpgpu_sim* g_the_gpu; + g_the_gpu()->perf_memcpy_to_gpu(dst_start_addr, count); if(g_debug_execution >= 3) { printf( " done.\n"); fflush(stdout); @@ -467,8 +467,8 @@ void gpgpu_t::memcpy_from_gpu( void *dst, size_t src_start_addr, size_t count ) m_global_mem->read(src_start_addr+n,1,dst_data+n); // Copy into the performance model. - extern gpgpu_sim* g_the_gpu; - g_the_gpu->perf_memcpy_to_gpu(src_start_addr, count); + //extern gpgpu_sim* g_the_gpu; + g_the_gpu()->perf_memcpy_to_gpu(src_start_addr, count); if(g_debug_execution >= 3) { printf( " done.\n"); fflush(stdout); @@ -1270,8 +1270,8 @@ void function_info::finalize( memory_space *param_mem ) void function_info::param_to_shared( memory_space *shared_mem, symbol_table *symtab ) { // TODO: call this only for PTXPlus with GT200 models - extern gpgpu_sim* g_the_gpu; - if (not g_the_gpu->get_config().convert_to_ptxplus()) return; + //extern gpgpu_sim* g_the_gpu; + if (not g_the_gpu()->get_config().convert_to_ptxplus()) return; // copies parameters into simulated shared memory for( std::map::iterator i=m_ptx_kernel_param_info.begin(); i!=m_ptx_kernel_param_info.end(); i++ ) { @@ -2150,7 +2150,7 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL ) printf("GPGPU-Sim: Performing Functional Simulation, executing kernel %s...\n",kernel.name().c_str()); //using a shader core object for book keeping, it is not needed but as most function built for performance simulation need it we use it here - extern gpgpu_sim *g_the_gpu; + //extern gpgpu_sim *g_the_gpu; //before we execute, we should do PDOM analysis for functional simulation scenario. function_info *kernel_func_info = kernel.entry(); const struct gpgpu_ptx_sim_info *kernel_info = ptx_sim_kernel_info(kernel_func_info); @@ -2165,7 +2165,7 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL ) kernel_func_info->set_pdom(); } - unsigned max_cta_tot = max_cta(kernel_info,kernel.threads_per_cta(), g_the_gpu->getShaderCoreConfig()->warp_size, g_the_gpu->getShaderCoreConfig()->n_thread_per_shader, g_the_gpu->getShaderCoreConfig()->gpgpu_shmem_size, g_the_gpu->getShaderCoreConfig()->gpgpu_shader_registers, g_the_gpu->getShaderCoreConfig()->max_cta_per_core); + unsigned max_cta_tot = max_cta(kernel_info,kernel.threads_per_cta(), g_the_gpu()->getShaderCoreConfig()->warp_size, g_the_gpu()->getShaderCoreConfig()->n_thread_per_shader, g_the_gpu()->getShaderCoreConfig()->gpgpu_shmem_size, g_the_gpu()->getShaderCoreConfig()->gpgpu_shader_registers, g_the_gpu()->getShaderCoreConfig()->max_cta_per_core); printf("Max CTA : %d\n",max_cta_tot); @@ -2173,11 +2173,11 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL ) int inst_count=50; - int cp_op= g_the_gpu->checkpoint_option; - int cp_CTA = g_the_gpu->checkpoint_CTA; - int cp_kernel= g_the_gpu->checkpoint_kernel; - cp_count= g_the_gpu->checkpoint_insn_Y; - cp_cta_resume= g_the_gpu->checkpoint_CTA_t; + int cp_op= g_the_gpu()->checkpoint_option; + int cp_CTA = g_the_gpu()->checkpoint_CTA; + int cp_kernel= g_the_gpu()->checkpoint_kernel; + cp_count= g_the_gpu()->checkpoint_insn_Y; + cp_cta_resume= g_the_gpu()->checkpoint_CTA_t; int cta_launched =0; //we excute the kernel one CTA (Block) at the time, as synchronization functions work block wise @@ -2189,8 +2189,8 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL ) { functionalCoreSim cta( &kernel, - g_the_gpu, - g_the_gpu->getShaderCoreConfig()->warp_size + g_the_gpu(), + g_the_gpu()->getShaderCoreConfig()->warp_size ); cta.execute(cp_count,temp); @@ -2211,7 +2211,7 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL ) { char f1name[2048]; snprintf(f1name,2048,"checkpoint_files/global_mem_%d.txt", kernel.get_uid() ); - g_checkpoint->store_global_mem(g_the_gpu->get_global_memory(), f1name , "%08x"); + g_checkpoint->store_global_mem(g_the_gpu()->get_global_memory(), f1name , "%08x"); } @@ -2221,8 +2221,8 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL ) //openCL kernel simulation calls don't register the kernel so we don't register its exit if(!openCL) { - extern stream_manager *g_stream_manager; - g_stream_manager->register_finished_kernel(kernel.get_uid()); + //extern stream_manager *g_stream_manager; + g_stream_manager()->register_finished_kernel(kernel.get_uid()); } //******PRINTING******* diff --git a/src/cuda-sim/cuda_device_runtime.cc b/src/cuda-sim/cuda_device_runtime.cc index 86e8147..be8369f 100644 --- a/src/cuda-sim/cuda_device_runtime.cc +++ b/src/cuda-sim/cuda_device_runtime.cc @@ -18,6 +18,7 @@ unsigned long long g_max_total_param_size = 0; #include "cuda-sim.h" #include "ptx_ir.h" #include "../stream_manager.h" +#include "../gpgpusim_entrypoint.h" #include "cuda_device_runtime.h" #define DEV_RUNTIME_REPORT(a) \ @@ -64,7 +65,7 @@ public: std::map g_cuda_device_launch_param_map; std::list g_cuda_device_launch_op; -extern stream_manager *g_stream_manager; +//extern stream_manager *g_stream_manager(); //Handling device runtime api: //void * cudaGetParameterBufferV2(void *func, dim3 gridDimension, dim3 blockDimension, unsigned int sharedMemSize) @@ -322,7 +323,7 @@ void launch_one_device_kernel() { device_launch_operation_t &op = g_cuda_device_launch_op.front(); stream_operation stream_op = stream_operation(op.grid, g_ptx_sim_mode, op.stream); - g_stream_manager->push(stream_op); + g_stream_manager()->push(stream_op); g_cuda_device_launch_op.pop_front(); } } diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index c8770e2..de937b0 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -41,8 +41,6 @@ struct GPGPUsim_ctx* the_gpgpusim = NULL; -static void print_simulation_time(); - struct GPGPUsim_ctx* GPGPUsim_ctx_ptr(){ if(the_gpgpusim == NULL) the_gpgpusim = new GPGPUsim_ctx(); @@ -50,6 +48,16 @@ struct GPGPUsim_ctx* GPGPUsim_ctx_ptr(){ return the_gpgpusim; } +class gpgpu_sim* g_the_gpu() { + return GPGPUsim_ctx_ptr()->g_the_gpu; +} + +class stream_manager* g_stream_manager() { + return GPGPUsim_ctx_ptr()->g_stream_manager; +} + +static void print_simulation_time(); + void *gpgpu_sim_thread_sequential(void*) { // at most one kernel running at a time diff --git a/src/gpgpusim_entrypoint.h b/src/gpgpusim_entrypoint.h index e29159b..2ad0fdf 100644 --- a/src/gpgpusim_entrypoint.h +++ b/src/gpgpusim_entrypoint.h @@ -46,9 +46,15 @@ struct GPGPUsim_ctx { sg_argc = 3; sg_argv = {"", "-config","gpgpusim.config"}; + + g_the_gpu_config=NULL; + g_the_gpu=NULL; + g_stream_manager=NULL; + the_cude_device=NULL; + the_context=NULL; } - struct gpgpu_ptx_sim_arg *grid_params; + //struct gpgpu_ptx_sim_arg *grid_params; sem_t g_sim_signal_start; sem_t g_sim_signal_finish; @@ -60,19 +66,27 @@ struct GPGPUsim_ctx { class gpgpu_sim *g_the_gpu; class stream_manager *g_stream_manager; + struct _cuda_device_id *the_cude_device; + struct CUctx_st* the_context; + + int sg_argc; const char *sg_argv[3]; - pthread_mutex_t g_sim_lock; - bool g_sim_active; - bool g_sim_done; - bool break_limit; + pthread_mutex_t g_sim_lock; + bool g_sim_active; + bool g_sim_done; + bool break_limit; }; class gpgpu_sim *gpgpu_ptx_sim_init_perf(); void start_sim_thread(int api); +class gpgpu_sim* g_the_gpu(); +struct GPGPUsim_ctx* GPGPUsim_ctx_ptr(); +class stream_manager* g_stream_manager(); + int gpgpu_opencl_ptx_sim_main_perf( kernel_info_t *grid ); int gpgpu_opencl_ptx_sim_main_func( kernel_info_t *grid ); -- cgit v1.3 From 4d4d5938d715d2b79a617c32583184426b4a642d Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Tue, 9 Jul 2019 23:16:17 -0400 Subject: Move g_ptx_sim_mode Signed-off-by: Mengchi Zhang --- libcuda/cuda_runtime_api.cc | 6 +++--- libcuda/gpgpu_context.h | 9 ++++++++- libopencl/opencl_runtime_api.cc | 6 +++--- src/cuda-sim/cuda-sim.cc | 6 ++---- src/cuda-sim/cuda-sim.h | 10 +++++++--- src/cuda-sim/cuda_device_runtime.cc | 7 ++++--- src/cuda-sim/cuda_device_runtime.h | 18 ++++++++++++++++-- src/gpgpu-sim/gpu-sim.cc | 4 ++-- src/gpgpusim_entrypoint.cc | 2 +- 9 files changed, 46 insertions(+), 22 deletions(-) (limited to 'src/cuda-sim/cuda_device_runtime.cc') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index bbbaf23..59d2a60 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1076,7 +1076,7 @@ cudaError_t cudaLaunchInternal( const char *hostFun, gpgpu_context* gpgpu_ctx = CUctx_st* context = GPGPUSim_Context(); char *mode = getenv("PTX_SIM_MODE_FUNC"); if( mode ) - sscanf(mode,"%u", &g_ptx_sim_mode); + sscanf(mode,"%u", &(ctx->func_sim->g_ptx_sim_mode)); gpgpusim_ptx_assert( !ctx->api->g_cuda_launch_stack.empty(), "empty launch stack" ); kernel_config config = ctx->api->g_cuda_launch_stack.back(); { @@ -1092,7 +1092,7 @@ cudaError_t cudaLaunchInternal( const char *hostFun, gpgpu_context* gpgpu_ctx = } struct CUstream_st *stream = config.get_stream(); printf("\nGPGPU-Sim PTX: cudaLaunch for 0x%p (mode=%s) on stream %u\n", hostFun, - g_ptx_sim_mode?"functional simulation":"performance simulation", stream?stream->get_uid():0 ); + (ctx->func_sim->g_ptx_sim_mode)?"functional simulation":"performance simulation", stream?stream->get_uid():0 ); kernel_info_t *grid = ctx->api->gpgpu_cuda_ptx_sim_init_grid(hostFun,config.get_args(),config.grid_dim(),config.block_dim(),context); //do dynamic PDOM analysis for performance simulation scenario std::string kname = grid->name(); @@ -1143,7 +1143,7 @@ cudaError_t cudaLaunchInternal( const char *hostFun, gpgpu_context* gpgpu_ctx = } printf("GPGPU-Sim PTX: pushing kernel \'%s\' to stream %u, gridDim= (%u,%u,%u) blockDim = (%u,%u,%u) \n", kname.c_str(), stream?stream->get_uid():0, gridDim.x,gridDim.y,gridDim.z,blockDim.x,blockDim.y,blockDim.z ); - stream_operation op(grid,g_ptx_sim_mode,stream); + stream_operation op(grid,ctx->func_sim->g_ptx_sim_mode,stream); g_stream_manager()->push(op); ctx->api->g_cuda_launch_stack.pop_back(); return g_last_cudaError = cudaSuccess; diff --git a/libcuda/gpgpu_context.h b/libcuda/gpgpu_context.h index a2ae7b6..3c9f87c 100644 --- a/libcuda/gpgpu_context.h +++ b/libcuda/gpgpu_context.h @@ -5,6 +5,7 @@ #include "../src/cuda-sim/ptx_parser.h" #include "../src/gpgpusim_entrypoint.h" #include "../src/cuda-sim/cuda-sim.h" +#include "../src/cuda-sim/cuda_device_runtime.h" class gpgpu_context { public: @@ -14,7 +15,10 @@ class gpgpu_context { ptxinfo = new ptxinfo_data(this); ptx_parser = new ptx_recognizer(this); the_gpgpusim = new GPGPUsim_ctx(this); - func_sim = new cuda_sim(); + func_sim = new cuda_sim(this); +#if (CUDART_VERSION >= 5000) + device_runtime = new cuda_device_runtime(this); +#endif } // global list symbol_table *g_global_allfiles_symbol_table; @@ -25,6 +29,9 @@ class gpgpu_context { ptx_recognizer* ptx_parser; GPGPUsim_ctx* the_gpgpusim; cuda_sim* func_sim; +#if (CUDART_VERSION >= 5000) + cuda_device_runtime* device_runtime; +#endif // member function list void cuobjdumpParseBinary(unsigned int handle); class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num ); diff --git a/libopencl/opencl_runtime_api.cc b/libopencl/opencl_runtime_api.cc index 0a6eb3e..aaaec4f 100644 --- a/libopencl/opencl_runtime_api.cc +++ b/libopencl/opencl_runtime_api.cc @@ -884,9 +884,9 @@ clEnqueueNDRangeKernel(cl_command_queue command_queue, printf("\n\n\n"); char *mode = getenv("PTX_SIM_MODE_FUNC"); if ( mode ) - sscanf(mode,"%u", &g_ptx_sim_mode); + sscanf(mode,"%u", &(ctx->func_sim->g_ptx_sim_mode)); printf("GPGPU-Sim OpenCL API: clEnqueueNDRangeKernel '%s' (mode=%s)\n", kernel->name().c_str(), - g_ptx_sim_mode?"functional simulation":"performance simulation"); + (ctx->func_sim->g_ptx_sim_mode)?"functional simulation":"performance simulation"); if ( !work_dim || work_dim > 3 ) return CL_INVALID_WORK_DIMENSION; size_t _local_size[3]; if( local_work_size != NULL ) { @@ -957,7 +957,7 @@ clEnqueueNDRangeKernel(cl_command_queue command_queue, gpgpu_ptx_sim_memcpy_symbol( "%_global_block_offset", zeros, 3 * sizeof(int), 0, 1, gpu ); } kernel_info_t *grid = ctx->func_sim->gpgpu_opencl_ptx_sim_init_grid(kernel->get_implementation(),params,GridDim,BlockDim,gpu); - if ( g_ptx_sim_mode ) + if ( ctx->func_sim->g_ptx_sim_mode ) ctx->func_sim->gpgpu_opencl_ptx_sim_main_func( grid ); else gpgpu_opencl_ptx_sim_main_perf( grid ); diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index b3e2965..7a7d205 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -2033,13 +2033,11 @@ void gpgpu_ptx_sim_memcpy_symbol(const char *hostVar, const void *src, size_t co fflush(stdout); } -int g_ptx_sim_mode; // if non-zero run functional simulation only (i.e., no notion of a clock cycle) - extern int ptx_debug; bool g_cuda_launch_blocking = false; -void read_sim_environment_variables() +void cuda_sim::read_sim_environment_variables() { ptx_debug = 0; g_debug_execution = 0; @@ -2185,7 +2183,7 @@ void cuda_sim::gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL cta.execute(cp_count,temp); #if (CUDART_VERSION >= 5000) - launch_all_device_kernels(); + gpgpu_ctx->device_runtime->launch_all_device_kernels(); #endif } else diff --git a/src/cuda-sim/cuda-sim.h b/src/cuda-sim/cuda-sim.h index 25ebf7b..3c4336d 100644 --- a/src/cuda-sim/cuda-sim.h +++ b/src/cuda-sim/cuda-sim.h @@ -36,12 +36,12 @@ #include #include"ptx_sim.h" +class gpgpu_context; class memory_space; class function_info; class symbol_table; extern const char *g_gpgpusim_version_string; -extern int g_ptx_sim_mode; extern int g_debug_execution; extern int g_debug_thread_uid; extern void ** g_inst_classification_stat; @@ -50,7 +50,6 @@ extern void ** g_inst_op_classification_stat; extern void print_splash(); extern void gpgpu_ptx_sim_memcpy_symbol(const char *hostVar, const void *src, size_t count, size_t offset, int to, gpgpu_t *gpu ); -extern void read_sim_environment_variables(); extern void ptxinfo_opencl_addinfo( std::map &kernels ); unsigned ptx_sim_init_thread( kernel_info_t &kernel, class ptx_thread_info** thread_info, @@ -124,9 +123,10 @@ struct gpgpu_ptx_sim_info get_ptxinfo(); class cuda_sim { public: - cuda_sim() { + cuda_sim( gpgpu_context* ctx ) { g_ptx_sim_num_insn = 0; g_ptx_kernel_count = -1; // used for classification stat collection purposes + gpgpu_ctx = ctx; } //global variables char *opcode_latency_int; @@ -147,6 +147,9 @@ class cuda_sim { int g_ptx_kernel_count; // used for classification stat collection purposes std::map g_global_name_lookup; // indexed by hostVar std::map g_const_name_lookup; // indexed by hostVar + int g_ptx_sim_mode; // if non-zero run functional simulation only (i.e., no notion of a clock cycle) + // backward pointer + class gpgpu_context* gpgpu_ctx; //global functions void ptx_opcocde_latency_options (option_parser_t opp); void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL = false ); @@ -159,6 +162,7 @@ class cuda_sim { gpgpu_t *gpu ); void gpgpu_ptx_sim_register_global_variable(void *hostVar, const char *deviceName, size_t size ); void gpgpu_ptx_sim_register_const_variable(void*, const char *deviceName, size_t size ); + void read_sim_environment_variables(); }; #endif diff --git a/src/cuda-sim/cuda_device_runtime.cc b/src/cuda-sim/cuda_device_runtime.cc index be8369f..354fa79 100644 --- a/src/cuda-sim/cuda_device_runtime.cc +++ b/src/cuda-sim/cuda_device_runtime.cc @@ -20,6 +20,7 @@ unsigned long long g_max_total_param_size = 0; #include "../stream_manager.h" #include "../gpgpusim_entrypoint.h" #include "cuda_device_runtime.h" +#include "../../libcuda/gpgpu_context.h" #define DEV_RUNTIME_REPORT(a) \ if( g_debug_execution ) { \ @@ -318,17 +319,17 @@ void gpgpusim_cuda_streamCreateWithFlags(const ptx_instruction * pI, ptx_thread_ } -void launch_one_device_kernel() { +void cuda_device_runtime::launch_one_device_kernel() { if(!g_cuda_device_launch_op.empty()) { device_launch_operation_t &op = g_cuda_device_launch_op.front(); - stream_operation stream_op = stream_operation(op.grid, g_ptx_sim_mode, op.stream); + stream_operation stream_op = stream_operation(op.grid, gpgpu_ctx->func_sim->g_ptx_sim_mode, op.stream); g_stream_manager()->push(stream_op); g_cuda_device_launch_op.pop_front(); } } -void launch_all_device_kernels() { +void cuda_device_runtime::launch_all_device_kernels() { while(!g_cuda_device_launch_op.empty()) { launch_one_device_kernel(); } diff --git a/src/cuda-sim/cuda_device_runtime.h b/src/cuda-sim/cuda_device_runtime.h index 6dbcd71..851fed2 100644 --- a/src/cuda-sim/cuda_device_runtime.h +++ b/src/cuda-sim/cuda_device_runtime.h @@ -6,6 +6,20 @@ void gpgpusim_cuda_getParameterBufferV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); void gpgpusim_cuda_launchDeviceV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); void gpgpusim_cuda_streamCreateWithFlags(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); -void launch_all_device_kernels(); -void launch_one_device_kernel(); +#endif +#if (CUDART_VERSION >= 5000) + +class gpgpu_context; + +class cuda_device_runtime { + public: + cuda_device_runtime( gpgpu_context* ctx ) { + gpgpu_ctx = ctx; + } + // backward pointer + class gpgpu_context* gpgpu_ctx; + void launch_all_device_kernels(); + void launch_one_device_kernel(); +}; + #endif diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 9f47067..bbcc078 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -503,7 +503,7 @@ void gpgpu_sim_config::reg_options(option_parser_t opp) &gpgpu_ptx_instruction_classification, "if enabled will classify ptx instruction types per kernel (Max 255 kernels now)", "0"); - option_parser_register(opp, "-gpgpu_ptx_sim_mode", OPT_INT32, &g_ptx_sim_mode, + option_parser_register(opp, "-gpgpu_ptx_sim_mode", OPT_INT32, &(gpgpu_ctx->func_sim->g_ptx_sim_mode), "Select between Performance (default) or Functional simulation (1)", "0"); option_parser_register(opp, "-gpgpu_clock_domains", OPT_CSTR, &gpgpu_clock_domains, @@ -1753,7 +1753,7 @@ void gpgpu_sim::cycle() #if (CUDART_VERSION >= 5000) //launch device kernel - launch_one_device_kernel(); + gpgpu_ctx->device_runtime->launch_one_device_kernel(); #endif } } diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index d9d1023..683a695 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -213,7 +213,7 @@ gpgpu_sim *gpgpu_context::gpgpu_ptx_sim_init_perf() { srand(1); print_splash(); - read_sim_environment_variables(); + func_sim->read_sim_environment_variables(); ptx_parser->read_parser_environment_variables(); option_parser_t opp = option_parser_create(); -- cgit v1.3 From 8019833f6c9d9552c70739dc2bf45e53655666ec Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Thu, 11 Jul 2019 11:01:27 -0400 Subject: Move g_total_param_size Signed-off-by: Mengchi Zhang --- src/abstract_hardware_model.cc | 3 +-- src/cuda-sim/cuda_device_runtime.cc | 3 +-- src/cuda-sim/cuda_device_runtime.h | 4 +++- src/cuda-sim/instructions.cc | 4 +++- 4 files changed, 8 insertions(+), 6 deletions(-) (limited to 'src/cuda-sim/cuda_device_runtime.cc') diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc index ef09051..0f4291d 100644 --- a/src/abstract_hardware_model.cc +++ b/src/abstract_hardware_model.cc @@ -771,8 +771,7 @@ bool kernel_info_t::children_all_finished() { void kernel_info_t::notify_parent_finished() { if(m_parent_kernel) { - extern unsigned long long g_total_param_size; - g_total_param_size -= ((m_kernel_entry->get_args_aligned_size() + 255)/256*256); + m_kernel_entry->gpgpu_ctx->device_runtime->g_total_param_size -= ((m_kernel_entry->get_args_aligned_size() + 255)/256*256); m_parent_kernel->remove_child(this); g_stream_manager()->register_finished_kernel(m_parent_kernel->get_uid()); } diff --git a/src/cuda-sim/cuda_device_runtime.cc b/src/cuda-sim/cuda_device_runtime.cc index 354fa79..22304db 100644 --- a/src/cuda-sim/cuda_device_runtime.cc +++ b/src/cuda-sim/cuda_device_runtime.cc @@ -5,7 +5,6 @@ #include #include -unsigned long long g_total_param_size = 0; unsigned long long g_max_total_param_size = 0; @@ -70,7 +69,7 @@ std::list g_cuda_device_launch_op; //Handling device runtime api: //void * cudaGetParameterBufferV2(void *func, dim3 gridDimension, dim3 blockDimension, unsigned int sharedMemSize) -void gpgpusim_cuda_getParameterBufferV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func) +void cuda_device_runtime::gpgpusim_cuda_getParameterBufferV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func) { DEV_RUNTIME_REPORT("Calling cudaGetParameterBufferV2"); diff --git a/src/cuda-sim/cuda_device_runtime.h b/src/cuda-sim/cuda_device_runtime.h index 851fed2..dd908ae 100644 --- a/src/cuda-sim/cuda_device_runtime.h +++ b/src/cuda-sim/cuda_device_runtime.h @@ -3,7 +3,6 @@ #if (CUDART_VERSION >= 5000) #pragma once -void gpgpusim_cuda_getParameterBufferV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); void gpgpusim_cuda_launchDeviceV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); void gpgpusim_cuda_streamCreateWithFlags(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); #endif @@ -14,10 +13,13 @@ class gpgpu_context; class cuda_device_runtime { public: cuda_device_runtime( gpgpu_context* ctx ) { + g_total_param_size = 0; gpgpu_ctx = ctx; } + unsigned long long g_total_param_size; // backward pointer class gpgpu_context* gpgpu_ctx; + void gpgpusim_cuda_getParameterBufferV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); void launch_all_device_kernels(); void launch_one_device_kernel(); }; diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 46a262d..a72bc0c 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -56,6 +56,8 @@ class ptx_recognizer; #include "cuda_device_runtime.h" #include +#include "../../libcuda/gpgpu_context.h" + using half_float::half; unsigned ptx_instruction::g_num_ptx_inst_uid=0; @@ -2095,7 +2097,7 @@ void call_impl( const ptx_instruction *pI, ptx_thread_info *thread ) #if (CUDART_VERSION >= 5000) //Jin: handle device runtime apis for CDP else if(fname == "cudaGetParameterBufferV2") { - gpgpusim_cuda_getParameterBufferV2(pI, thread, target_func); + target_func->gpgpu_ctx->device_runtime->gpgpusim_cuda_getParameterBufferV2(pI, thread, target_func); return; } else if(fname == "cudaLaunchDeviceV2") { -- cgit v1.3 From b1385157251803aa45eaff154637218c9f6b15b1 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Thu, 11 Jul 2019 14:23:13 -0400 Subject: Move g_cuda_device_launch_param_map and g_cuda_device_launch_op Signed-off-by: Mengchi Zhang --- src/cuda-sim/cuda_device_runtime.cc | 41 ++---------------------------------- src/cuda-sim/cuda_device_runtime.h | 42 +++++++++++++++++++++++++++++++++---- src/cuda-sim/instructions.cc | 4 ++-- 3 files changed, 42 insertions(+), 45 deletions(-) (limited to 'src/cuda-sim/cuda_device_runtime.cc') diff --git a/src/cuda-sim/cuda_device_runtime.cc b/src/cuda-sim/cuda_device_runtime.cc index 22304db..326a1c5 100644 --- a/src/cuda-sim/cuda_device_runtime.cc +++ b/src/cuda-sim/cuda_device_runtime.cc @@ -27,44 +27,7 @@ unsigned long long g_max_total_param_size = 0; std::cout.flush(); \ } -class device_launch_config_t { - -public: - device_launch_config_t() {} - - device_launch_config_t(dim3 _grid_dim, - dim3 _block_dim, - unsigned int _shared_mem, - function_info * _entry): - grid_dim(_grid_dim), - block_dim(_block_dim), - shared_mem(_shared_mem), - entry(_entry) {} - - dim3 grid_dim; - dim3 block_dim; - unsigned int shared_mem; - function_info * entry; - -}; - -class device_launch_operation_t { - -public: - device_launch_operation_t() {} - device_launch_operation_t(kernel_info_t *_grid, - CUstream_st * _stream) : - grid(_grid), stream(_stream) {} - - kernel_info_t * grid; //a new child grid - - CUstream_st * stream; - -}; - -std::map g_cuda_device_launch_param_map; -std::list g_cuda_device_launch_op; //extern stream_manager *g_stream_manager(); //Handling device runtime api: @@ -142,7 +105,7 @@ void cuda_device_runtime::gpgpusim_cuda_getParameterBufferV2(const ptx_instructi //Handling device runtime api: //cudaError_t cudaLaunchDeviceV2(void *parameterBuffer, cudaStream_t stream) -void gpgpusim_cuda_launchDeviceV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func) { +void cuda_device_runtime::gpgpusim_cuda_launchDeviceV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func) { DEV_RUNTIME_REPORT("Calling cudaLaunchDeviceV2"); unsigned n_return = target_func->has_return(); @@ -263,7 +226,7 @@ void gpgpusim_cuda_launchDeviceV2(const ptx_instruction * pI, ptx_thread_info * //Handling device runtime api: //cudaError_t cudaStreamCreateWithFlags ( cudaStream_t* pStream, unsigned int flags) //flags can only be cudaStreamNonBlocking -void gpgpusim_cuda_streamCreateWithFlags(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func) { +void cuda_device_runtime::gpgpusim_cuda_streamCreateWithFlags(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func) { DEV_RUNTIME_REPORT("Calling cudaStreamCreateWithFlags"); unsigned n_return = target_func->has_return(); diff --git a/src/cuda-sim/cuda_device_runtime.h b/src/cuda-sim/cuda_device_runtime.h index dd908ae..b49d0eb 100644 --- a/src/cuda-sim/cuda_device_runtime.h +++ b/src/cuda-sim/cuda_device_runtime.h @@ -3,10 +3,40 @@ #if (CUDART_VERSION >= 5000) #pragma once -void gpgpusim_cuda_launchDeviceV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); -void gpgpusim_cuda_streamCreateWithFlags(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); -#endif -#if (CUDART_VERSION >= 5000) +class device_launch_config_t { + +public: + device_launch_config_t() {} + + device_launch_config_t(dim3 _grid_dim, + dim3 _block_dim, + unsigned int _shared_mem, + function_info * _entry): + grid_dim(_grid_dim), + block_dim(_block_dim), + shared_mem(_shared_mem), + entry(_entry) {} + + dim3 grid_dim; + dim3 block_dim; + unsigned int shared_mem; + function_info * entry; + +}; + +class device_launch_operation_t { + +public: + device_launch_operation_t() {} + device_launch_operation_t(kernel_info_t *_grid, + CUstream_st * _stream) : + grid(_grid), stream(_stream) {} + + kernel_info_t * grid; //a new child grid + + CUstream_st * stream; + +}; class gpgpu_context; @@ -17,8 +47,12 @@ class cuda_device_runtime { gpgpu_ctx = ctx; } unsigned long long g_total_param_size; + std::map g_cuda_device_launch_param_map; + std::list g_cuda_device_launch_op; // backward pointer class gpgpu_context* gpgpu_ctx; + void gpgpusim_cuda_launchDeviceV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); + void gpgpusim_cuda_streamCreateWithFlags(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); void gpgpusim_cuda_getParameterBufferV2(const ptx_instruction * pI, ptx_thread_info * thread, const function_info * target_func); void launch_all_device_kernels(); void launch_one_device_kernel(); diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index dfe5af6..565340c 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -2101,11 +2101,11 @@ void call_impl( const ptx_instruction *pI, ptx_thread_info *thread ) return; } else if(fname == "cudaLaunchDeviceV2") { - gpgpusim_cuda_launchDeviceV2(pI, thread, target_func); + target_func->gpgpu_ctx->device_runtime->gpgpusim_cuda_launchDeviceV2(pI, thread, target_func); return; } else if(fname == "cudaStreamCreateWithFlags") { - gpgpusim_cuda_streamCreateWithFlags(pI, thread, target_func); + target_func->gpgpu_ctx->device_runtime->gpgpusim_cuda_streamCreateWithFlags(pI, thread, target_func); return; } #endif -- cgit v1.3 From 524c7e2ff37bba3ec60706605b34c6e354f928f1 Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Fri, 12 Jul 2019 00:02:57 -0400 Subject: Move g_max_total_param_size Signed-off-by: Mengchi Zhang --- src/cuda-sim/cuda_device_runtime.cc | 1 - src/cuda-sim/cuda_device_runtime.h | 2 ++ src/gpgpu-sim/gpu-sim.cc | 3 +-- 3 files changed, 3 insertions(+), 3 deletions(-) (limited to 'src/cuda-sim/cuda_device_runtime.cc') diff --git a/src/cuda-sim/cuda_device_runtime.cc b/src/cuda-sim/cuda_device_runtime.cc index 326a1c5..dc3adc3 100644 --- a/src/cuda-sim/cuda_device_runtime.cc +++ b/src/cuda-sim/cuda_device_runtime.cc @@ -5,7 +5,6 @@ #include #include -unsigned long long g_max_total_param_size = 0; #if (CUDART_VERSION >= 5000) diff --git a/src/cuda-sim/cuda_device_runtime.h b/src/cuda-sim/cuda_device_runtime.h index 5927b54..f37849b 100644 --- a/src/cuda-sim/cuda_device_runtime.h +++ b/src/cuda-sim/cuda_device_runtime.h @@ -43,12 +43,14 @@ class cuda_device_runtime { public: cuda_device_runtime( gpgpu_context* ctx ) { g_total_param_size = 0; + g_max_total_param_size = 0; gpgpu_ctx = ctx; } unsigned long long g_total_param_size; std::map g_cuda_device_launch_param_map; std::list g_cuda_device_launch_op; unsigned g_kernel_launch_latency; + unsigned long long g_max_total_param_size; // backward pointer class gpgpu_context* gpgpu_ctx; #if (CUDART_VERSION >= 5000) diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 9726cf5..a3d6a8a 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -1106,8 +1106,7 @@ void gpgpu_sim::gpu_print_stat() printf("gpu_tot_occupancy = %.4f\% \n", (gpu_occupancy + gpu_tot_occupancy).get_occ_fraction() * 100); - extern unsigned long long g_max_total_param_size; - fprintf(statfout, "max_total_param_size = %llu\n", g_max_total_param_size); + fprintf(statfout, "max_total_param_size = %llu\n", gpgpu_ctx->device_runtime->g_max_total_param_size); // performance counter for stalls due to congestion. printf("gpu_stall_dramfull = %d\n", gpu_stall_dramfull); -- cgit v1.3