From 35cf76f383ec8de6de901bbbcd8fb478f69e46e4 Mon Sep 17 00:00:00 2001 From: sspenst Date: Wed, 6 Jul 2016 13:56:52 -0700 Subject: Added sstarr memory, which works the same as shared memory --- src/cuda-sim/ptx_sim.cc | 1 + 1 file changed, 1 insertion(+) (limited to 'src/cuda-sim/ptx_sim.cc') diff --git a/src/cuda-sim/ptx_sim.cc b/src/cuda-sim/ptx_sim.cc index 09844ae..511e8d6 100644 --- a/src/cuda-sim/ptx_sim.cc +++ b/src/cuda-sim/ptx_sim.cc @@ -153,6 +153,7 @@ ptx_thread_info::ptx_thread_info( kernel_info_t &kernel ) m_last_memory_space = undefined_space; m_branch_taken = 0; m_shared_mem = NULL; + m_sstarr_mem = NULL; m_cta_info = NULL; m_local_mem = NULL; m_symbol_table = NULL; -- cgit v1.3 From adc311951d67b0685ebf2fab4ce6410f96f0039a Mon Sep 17 00:00:00 2001 From: sspenst Date: Mon, 11 Jul 2016 11:16:46 -0700 Subject: Reverted the previous commit to add a cleaner way of getting NUM_THREADS. Now, sst_impl doesn't functionally execute on the last indexed element of an array, but instead on the actual last thread that executes --- libcuda/cuda_runtime_api.cc | 7 ++----- src/abstract_hardware_model.h | 3 --- src/cuda-sim/instructions.cc | 13 ++++++++----- src/cuda-sim/ptx_sim.cc | 16 ++++++++++++++++ src/cuda-sim/ptx_sim.h | 4 ++++ 5 files changed, 30 insertions(+), 13 deletions(-) (limited to 'src/cuda-sim/ptx_sim.cc') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 3eff4af..e8a0e91 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -180,9 +180,6 @@ cudaError_t g_last_cudaError = cudaSuccess; extern stream_manager *g_stream_manager; -dim3 gridDim = (dim3){1,1,1}; -dim3 blockDim = (dim3){1,1,1}; - void register_ptx_function( const char *name, function_info *impl ) { // no longer need this @@ -962,8 +959,8 @@ __host__ cudaError_t CUDARTAPI cudaLaunch( const char *hostFun ) g_ptx_sim_mode?"functional simulation":"performance simulation", stream?stream->get_uid():0 ); kernel_info_t *grid = gpgpu_cuda_ptx_sim_init_grid(hostFun,config.get_args(),config.grid_dim(),config.block_dim(),context); std::string kname = grid->name(); - gridDim = config.grid_dim(); - blockDim = config.block_dim(); + dim3 gridDim = config.grid_dim(); + dim3 blockDim = config.block_dim(); 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); diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 46c3279..6ed9b8e 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -162,9 +162,6 @@ struct dim3 { }; #endif -extern dim3 gridDim; -extern dim3 blockDim; - void increment_x_then_y_then_z( dim3 &i, const dim3 &bound); class kernel_info_t { diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index fd3b1fa..b5a3db4 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -3754,6 +3754,7 @@ void sst_impl( const ptx_instruction *pI, ptx_thread_info *thread ) memory_space_t space = pI->get_space(); memory_space *mem = NULL; addr_t addr = src2_data.u32 * 4; // this assumes sstarr memory starts at address 0 + ptx_cta_info *cta_info = thread->m_cta_info; decode_space(space,thread,src1,mem,addr); @@ -3765,18 +3766,19 @@ void sst_impl( const ptx_instruction *pI, ptx_thread_info *thread ) mem->write(addr,size/8,&src3_data.s64,thread,pI); // sync threads - cpI->set_bar_id(dst_data.u32); + cpI->set_bar_id(16); // use 16 for sst because bar uses an int from 0-15 thread->m_last_effective_address = addr; thread->m_last_memory_space = space; thread->m_last_dram_callback.function = bar_callback; thread->m_last_dram_callback.instruction = cpI; - int NUM_THREADS = blockDim.x * blockDim.y * blockDim.z; - if (src2_data.s64 == NUM_THREADS-1) { - // pick only one thread to load all of the data back from sstarr memory + // the last thread that executes loads all of the data back from sstarr memory + int NUM_THREADS = cta_info->num_threads(); + cta_info->inc_bar_threads(); + if (NUM_THREADS == cta_info->get_bar_threads()) { unsigned offset = 0; - addr -= (NUM_THREADS-1)*4; + addr = 0; ptx_reg_t data; float sstarr_fdata[NUM_THREADS]; signed long long sstarr_ldata[NUM_THREADS]; @@ -3812,6 +3814,7 @@ void sst_impl( const ptx_instruction *pI, ptx_thread_info *thread ) offset++; } + cta_info->reset_bar_threads(); thread->m_last_effective_address = addr+(NUM_THREADS-1)*4; thread->m_last_memory_space = space; } diff --git a/src/cuda-sim/ptx_sim.cc b/src/cuda-sim/ptx_sim.cc index 511e8d6..f48115b 100644 --- a/src/cuda-sim/ptx_sim.cc +++ b/src/cuda-sim/ptx_sim.cc @@ -44,6 +44,7 @@ ptx_cta_info::ptx_cta_info( unsigned sm_idx ) m_sm_idx = sm_idx; m_uid = g_ptx_cta_info_uid++; + m_bar_threads = 0; } void ptx_cta_info::add_thread( ptx_thread_info *thd ) @@ -128,6 +129,21 @@ unsigned ptx_cta_info::get_sm_idx() const return m_sm_idx; } +unsigned ptx_cta_info::get_bar_threads() const +{ + return m_bar_threads; +} + +void ptx_cta_info::inc_bar_threads() +{ + m_bar_threads++; +} + +void ptx_cta_info::reset_bar_threads() +{ + m_bar_threads = 0; +} + unsigned g_ptx_thread_info_uid_next=1; unsigned g_ptx_thread_info_delete_count=0; diff --git a/src/cuda-sim/ptx_sim.h b/src/cuda-sim/ptx_sim.h index c66b68c..4e748e9 100644 --- a/src/cuda-sim/ptx_sim.h +++ b/src/cuda-sim/ptx_sim.h @@ -158,8 +158,12 @@ public: void register_thread_exit( ptx_thread_info *thd ); void register_deleted_thread( ptx_thread_info *thd ); unsigned get_sm_idx() const; + unsigned get_bar_threads() const; + void inc_bar_threads(); + void reset_bar_threads(); private: + unsigned m_bar_threads; unsigned long long m_uid; unsigned m_sm_idx; std::set m_threads_in_cta; -- cgit v1.3 From d1b45cf53a39261663a3eff0d409d6c1220d923d Mon Sep 17 00:00:00 2001 From: sspenst Date: Fri, 5 Aug 2016 14:45:56 -0700 Subject: Added ptx_warp_info to know how many threads within a warp have executed --- src/abstract_hardware_model.h | 2 + src/cuda-sim/cuda-sim.cc | 10 +++++ src/cuda-sim/instructions.cc | 95 +++++++++++++++++++++++-------------------- src/cuda-sim/ptx_sim.cc | 21 ++++++++++ src/cuda-sim/ptx_sim.h | 12 ++++++ 5 files changed, 95 insertions(+), 45 deletions(-) (limited to 'src/cuda-sim/ptx_sim.cc') diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index 13dfce3..cfa8c9f 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -1028,6 +1028,7 @@ class core_t { m_warp_count += 1; } assert( m_warp_count * m_warp_size > 0 ); + //m_warp = ( ptx_warp_info** )calloc( m_warp_count, sizeof( ptx_warp_info* ) ); m_thread = ( ptx_thread_info** ) calloc( m_warp_count * m_warp_size, sizeof( ptx_thread_info* ) ); @@ -1063,6 +1064,7 @@ class core_t { class gpgpu_sim *m_gpu; kernel_info_t *m_kernel; simt_stack **m_simt_stack; // pdom based reconvergence context for each warp + //class ptx_warp_info ** m_warp; class ptx_thread_info ** m_thread; unsigned m_warp_size; unsigned m_warp_count; diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 337463b..ba0d00b 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -1417,6 +1417,7 @@ unsigned ptx_sim_init_thread( kernel_info_t &kernel, static std::map shared_memory_lookup; static std::map sstarr_memory_lookup; static std::map ptx_cta_lookup; + static std::map ptx_warp_lookup; static std::map > local_memory_lookup; if ( *thread_info != NULL ) { @@ -1501,6 +1502,15 @@ unsigned ptx_sim_init_thread( kernel_info_t &kernel, new_tid += tid; ptx_thread_info *thd = new ptx_thread_info(kernel); + ptx_warp_info *warp_info = NULL; + if ( ptx_warp_lookup.find(hw_warp_id) == ptx_warp_lookup.end() ) { + warp_info = new ptx_warp_info(); // num_threads should be threads in the warp + ptx_warp_lookup[hw_warp_id] = warp_info; + } else { + warp_info = ptx_warp_lookup[hw_warp_id]; + } + thd->m_warp_info = warp_info; + memory_space *local_mem = NULL; std::map::iterator l = local_mem_lookup.find(new_tid); if ( l != local_mem_lookup.end() ) { diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index f58c4f5..9dcc25c 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -1471,74 +1471,81 @@ void bsmad_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) // 7 = synapse value // 8 = output value - // TODO: what should happen when the output precision is larger than the input precision? - // TODO: create a ptx_warp_info that can do the same thing that ptx_cta_info does here - ptx_cta_info *cta_info = core->get_thread_info()[inst.warp_id() * core->get_warp_size()]->m_cta_info; - const int NUM_THREADS = cta_info->num_threads(); - const int NUM_BUFFERS = 4; - cta_info->inc_bar_threads(); + const operand_info &dst = pI->dst(); + const operand_info &src1 = pI->src1(); + const operand_info &src2 = pI->src2(); + unsigned type = pI->get_type(); + int tid = inst.warp_id() * core->get_warp_size(); + ptx_thread_info *thread = core->get_thread_info()[tid]; + const int ip = (thread->get_operand_value(src1, dst, type, thread, 1)).u32; + const int op = (thread->get_operand_value(src2, dst, type, thread, 1)).u32; + const int THREADS = inst.active_count(); + const int INBUFFERS = 4; + const int OUTBUFFERS = (((32/ip)*INBUFFERS) / (32/op)) + ((((32/ip)*INBUFFERS) % (32/op)) != 0); + if (OUTBUFFERS > THREADS) { + printf("GPGPU-Sim PTX: BSMAD ERROR - Number of output registers required (%d) is greater than the number available (%d)\n", OUTBUFFERS, THREADS); + abort(); + } + ptx_warp_info *warp_info = core->get_thread_info()[inst.warp_id() * core->get_warp_size()]->m_warp_info; + warp_info->inc_done_threads(); // threads within the warp are executed sequentially by the simulator, store output in first four registers - if (cta_info->get_bar_threads() <= NUM_BUFFERS) { - unsigned ip, op; // only get these when i = 0 - unsigned buffer[inst.active_count()][NUM_BUFFERS]; + if (warp_info->get_done_threads() <= OUTBUFFERS) { + unsigned buffer[inst.active_count()][INBUFFERS]; unsigned synapse[inst.active_count()]; - unsigned output[NUM_BUFFERS]; + unsigned output; // loop through all threads in the warp and get all data for (unsigned i = 0, j = 0; i < core->get_warp_size(); i++) { if (inst.active(i)) { - const operand_info dst = pI->dst(); - const operand_info src1 = pI->operand_lookup(1); - const operand_info src2 = pI->operand_lookup(2); - const operand_info src3 = pI->operand_lookup(3); - const operand_info src4 = pI->operand_lookup(4); - const operand_info src5 = pI->operand_lookup(5); - const operand_info src6 = pI->operand_lookup(6); - const operand_info src7 = pI->operand_lookup(7); - const operand_info src8 = pI->operand_lookup(8); - unsigned type = pI->get_type(); - - int tid = inst.warp_id() * core->get_warp_size() + i; - ptx_thread_info *thread = core->get_thread_info()[tid]; - - // only get precision data once - if (j == 0) { - ip = (thread->get_operand_value(src1, dst, type, thread, 1)).u32; - op = (thread->get_operand_value(src2, dst, type, thread, 1)).u32; - } + const operand_info &src3 = pI->operand_lookup(3); + const operand_info &src4 = pI->operand_lookup(4); + const operand_info &src5 = pI->operand_lookup(5); + const operand_info &src6 = pI->operand_lookup(6); + const operand_info &src7 = pI->operand_lookup(7); + const operand_info &src8 = pI->operand_lookup(8); + + thread = core->get_thread_info()[tid+i]; // get buffer data and synapse data from each thread buffer[j][0] = (thread->get_operand_value(src3, dst, type, thread, 1)).u32; buffer[j][1] = (thread->get_operand_value(src4, dst, type, thread, 1)).u32; buffer[j][2] = (thread->get_operand_value(src5, dst, type, thread, 1)).u32; buffer[j][3] = (thread->get_operand_value(src6, dst, type, thread, 1)).u32; synapse[j] = (thread->get_operand_value(src7, dst, type, thread, 1)).u32; + j++; // get output data from the first 4 threads - if (j < NUM_BUFFERS) { - output[j] = (thread->get_operand_value(src8, dst, type, thread, 1)).u32; + if (j == warp_info->get_done_threads()) { + output = (thread->get_operand_value(src8, dst, type, thread, 1)).u32; } - j++; } } // unpack registers, compute enough outputs to fill an output register unsigned *unpacked_output = (unsigned*)calloc(32/op,sizeof(unsigned)); - unsigned buffer_data_start = (32/op)*(cta_info->get_bar_threads()-1); - for (unsigned i = buffer_data_start; i < (32/op + buffer_data_start) && i < (32/ip)*NUM_BUFFERS; i++) { + unsigned buffer_data_start = (32/op)*(warp_info->get_done_threads()-1); + for (unsigned i = buffer_data_start; i < (32/op + buffer_data_start) && i < (32/ip)*INBUFFERS; i++) { unsigned buf = i/(32/ip); unsigned pos = i%(32/ip); - unsigned mask = 0; + int sum = 0; + // sum values from the buffers for (int b = 0; b < ip; b++) { mask |= (1 << b); } mask <<= (pos*ip); - int sum = 0; - for (int j = 0; j < NUM_THREADS; j++) { + for (int j = 0; j < THREADS; j++) { sum += (mask & buffer[j][buf]) >> (pos*ip); } - unpacked_output[i - buffer_data_start] = sum; + // get the previous output + mask = 0; + for (int b = 0; b < op; b++) { + mask |= (1 << b); + } + mask <<= (op*(i-buffer_data_start)); + int past_output = (mask & output) >> (op*(i-buffer_data_start)); + + unpacked_output[i-buffer_data_start] = sum + past_output; } // truncate output @@ -1575,11 +1582,8 @@ void bsmad_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) // store the result in the correct thread's output register for (unsigned i = 0, j = 0; i < core->get_warp_size(); i++) { if (inst.active(i)) j++; - if (j == cta_info->get_bar_threads()) { - const operand_info &dst = pI->dst(); - unsigned type = pI->get_type(); - int tid = inst.warp_id() * core->get_warp_size() + i; - ptx_thread_info *thread = core->get_thread_info()[tid]; + if (j == warp_info->get_done_threads()) { + thread = core->get_thread_info()[tid+i]; ptx_reg_t data; data.u32 = output_data; thread->set_operand_value(dst, data, type, thread, pI); @@ -1588,8 +1592,9 @@ void bsmad_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst ) } } - if (cta_info->get_bar_threads() == NUM_THREADS) { - cta_info->reset_bar_threads(); + // once the warp has finished, set the number of completed threads back to 0 for the next warp + if (warp_info->get_done_threads() == THREADS) { + warp_info->reset_done_threads(); } } diff --git a/src/cuda-sim/ptx_sim.cc b/src/cuda-sim/ptx_sim.cc index f48115b..820287d 100644 --- a/src/cuda-sim/ptx_sim.cc +++ b/src/cuda-sim/ptx_sim.cc @@ -144,6 +144,26 @@ void ptx_cta_info::reset_bar_threads() m_bar_threads = 0; } +ptx_warp_info::ptx_warp_info() +{ + reset_done_threads(); +} + +unsigned ptx_warp_info::get_done_threads() const +{ + return m_done_threads; +} + +void ptx_warp_info::inc_done_threads() +{ + m_done_threads++; +} + +void ptx_warp_info::reset_done_threads() +{ + m_done_threads = 0; +} + unsigned g_ptx_thread_info_uid_next=1; unsigned g_ptx_thread_info_delete_count=0; @@ -170,6 +190,7 @@ ptx_thread_info::ptx_thread_info( kernel_info_t &kernel ) m_branch_taken = 0; m_shared_mem = NULL; m_sstarr_mem = NULL; + m_warp_info = NULL; m_cta_info = NULL; m_local_mem = NULL; m_symbol_table = NULL; diff --git a/src/cuda-sim/ptx_sim.h b/src/cuda-sim/ptx_sim.h index 4e748e9..c62fa57 100644 --- a/src/cuda-sim/ptx_sim.h +++ b/src/cuda-sim/ptx_sim.h @@ -171,6 +171,17 @@ private: std::set m_dangling_pointers; }; +class ptx_warp_info { +public: + ptx_warp_info(); // add get_core or something, or threads? + unsigned get_done_threads() const; + void inc_done_threads(); + void reset_done_threads(); + +private: + unsigned m_done_threads; +}; + class symbol; struct stack_entry { @@ -430,6 +441,7 @@ public: memory_space *m_shared_mem; memory_space *m_sstarr_mem; memory_space *m_local_mem; + ptx_warp_info *m_warp_info; ptx_cta_info *m_cta_info; ptx_reg_t m_last_set_operand_value; -- cgit v1.3