From 779a636d32914e30ca039a043d0f3abc88d2f1ea Mon Sep 17 00:00:00 2001 From: chris Date: Thu, 4 Jul 2019 00:18:58 +0000 Subject: added temp. solution to bypass function pointers in order for pytorch to run --- src/cuda-sim/ptx.y | 18 ++++++++++++++++++ 1 file changed, 18 insertions(+) (limited to 'src') diff --git a/src/cuda-sim/ptx.y b/src/cuda-sim/ptx.y index 45392fb..260564f 100644 --- a/src/cuda-sim/ptx.y +++ b/src/cuda-sim/ptx.y @@ -287,6 +287,7 @@ ptr_align_spec: ALIGN_DIRECTIVE INT_OPERAND statement_block: LEFT_BRACE statement_list RIGHT_BRACE statement_list: directive_statement { add_directive(); } + | statement_list prototype_block {printf("Prototype statement detected. WARNING: this is not supported yet on GPGPU-SIM\n"); } | instruction_statement { add_instruction(); } | statement_list directive_statement { add_directive(); } | statement_list instruction_statement { add_instruction(); } @@ -403,6 +404,23 @@ initializer_list: LEFT_BRACE literal_list RIGHT_BRACE { add_array_initializer(); literal_list: literal_operand | literal_list COMMA literal_operand; +// TODO: This is currently hardcoded to handle and ignore one specific case +// that all prototype statements follow in the PTX from Pytorch. As a +// workaround, this parses and ignores both the prototype declaration +// and calling of the prototype (which conveniently comes right after the +// declaration for all cases.) This should be changed to handle both +// declaring the prototype, and actually calling it. +prototype_block: prototype_decl prototype_call + +prototype_decl: IDENTIFIER COLON CALLPROTOTYPE_DIRECTIVE LEFT_PAREN prototype_param RIGHT_PAREN IDENTIFIER LEFT_PAREN prototype_param RIGHT_PAREN SEMI_COLON + +prototype_call: OPCODE LEFT_PAREN IDENTIFIER RIGHT_PAREN COMMA operand COMMA LEFT_PAREN IDENTIFIER RIGHT_PAREN COMMA IDENTIFIER SEMI_COLON + | OPCODE IDENTIFIER COMMA LEFT_PAREN IDENTIFIER RIGHT_PAREN COMMA IDENTIFIER SEMI_COLON + +prototype_param: /* empty */ + | PARAM_DIRECTIVE B64_TYPE IDENTIFIER + | PARAM_DIRECTIVE B32_TYPE IDENTIFIER + instruction_statement: instruction SEMI_COLON | IDENTIFIER COLON { add_label($1); } | pred_spec instruction SEMI_COLON; -- cgit v1.3 From d212d7e5fdcc9f8e10779d5cfb398a451f8ad033 Mon Sep 17 00:00:00 2001 From: boyealKim Date: Sun, 21 Jul 2019 02:28:59 +0900 Subject: take account of shfl latency --- configs/tested-cfgs/SM2_GTX480/gpgpusim.config | 6 +++--- configs/tested-cfgs/SM6_TITANX/gpgpusim.config | 6 +++--- configs/tested-cfgs/SM7_TITANV/gpgpusim.config | 6 +++--- src/cuda-sim/cuda-sim.cc | 29 +++++++++++++------------- src/gpgpu-sim/shader.cc | 9 ++++---- 5 files changed, 29 insertions(+), 27 deletions(-) (limited to 'src') diff --git a/configs/tested-cfgs/SM2_GTX480/gpgpusim.config b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config index cf3627b..7ff836a 100644 --- a/configs/tested-cfgs/SM2_GTX480/gpgpusim.config +++ b/configs/tested-cfgs/SM2_GTX480/gpgpusim.config @@ -46,9 +46,9 @@ -gpgpu_num_dp_units 0 # Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,2,2,1,8 +# "ADD,MAX,MUL,MAD,DIV,SHFL" +-ptx_opcode_latency_int 4,13,4,5,145,32 +-ptx_opcode_initiation_int 1,2,2,1,8,4 -ptx_opcode_latency_fp 4,13,4,5,39 -ptx_opcode_initiation_fp 1,2,1,1,4 -ptx_opcode_latency_dp 8,19,8,8,330 diff --git a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config index 2fe898a..da1af48 100644 --- a/configs/tested-cfgs/SM6_TITANX/gpgpusim.config +++ b/configs/tested-cfgs/SM6_TITANX/gpgpusim.config @@ -56,11 +56,11 @@ # Instruction latencies and initiation intervals -# "ADD,MAX,MUL,MAD,DIV" +# "ADD,MAX,MUL,MAD,DIV,SHFL" # All Div operations are executed on SFU unit # Throughput (initiation latency) are adopted from CUDA SDK document V8, section 5.4.1, Table 2 --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 1,1,1,1,4 +-ptx_opcode_latency_int 4,13,4,5,145,32 +-ptx_opcode_initiation_int 1,1,1,1,4,4 -ptx_opcode_latency_fp 4,13,4,5,39 -ptx_opcode_initiation_fp 1,2,1,1,4 -ptx_opcode_latency_dp 8,19,8,8,330 diff --git a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config index ebd442f..f2301e8 100644 --- a/configs/tested-cfgs/SM7_TITANV/gpgpusim.config +++ b/configs/tested-cfgs/SM7_TITANV/gpgpusim.config @@ -66,10 +66,10 @@ # Instruction latencies and initiation intervals # "ADD,MAX,MUL,MAD,DIV" # All Div operations are executed on SFU unit -# Throughput (initiation latency) are adopted from +# Throughput (initiation latency except shfl) are adopted from # http://on-demand.gputechconf.com/gtc/2018/presentation/s8122-dissecting-the-volta-gpu-architecture-through-microbenchmarking.pdf --ptx_opcode_latency_int 4,13,4,5,145 --ptx_opcode_initiation_int 2,2,2,2,8 +-ptx_opcode_latency_int 4,13,4,5,145,32 +-ptx_opcode_initiation_int 2,2,2,2,8,4 -ptx_opcode_latency_fp 4,13,4,5,39 -ptx_opcode_initiation_fp 2,2,2,2,4 -ptx_opcode_latency_dp 8,19,8,8,330 diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index f7bb9cc..7376cc6 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -71,9 +71,9 @@ unsigned cdp_latency[5]; void ptx_opcocde_latency_options (option_parser_t opp) { option_parser_register(opp, "-ptx_opcode_latency_int", OPT_CSTR, &opcode_latency_int, - "Opcode latencies for integers " - "Default 1,1,19,25,145", - "1,1,19,25,145"); + "Opcode latencies for integers " + "Default 1,1,19,25,145,32", + "1,1,19,25,145,32"); option_parser_register(opp, "-ptx_opcode_latency_fp", OPT_CSTR, &opcode_latency_fp, "Opcode latencies for single precision floating points " "Default 1,1,1,1,30", @@ -91,9 +91,9 @@ void ptx_opcocde_latency_options (option_parser_t opp) { "Default 64", "64"); option_parser_register(opp, "-ptx_opcode_initiation_int", OPT_CSTR, &opcode_initiation_int, - "Opcode initiation intervals for integers " - "Default 1,1,4,4,32", - "1,1,4,4,32"); + "Opcode initiation intervals for integers " + "Default 1,1,4,4,32,4", + "1,1,4,4,32,4"); option_parser_register(opp, "-ptx_opcode_initiation_fp", OPT_CSTR, &opcode_initiation_fp, "Opcode initiation intervals for single precision floating points " "Default 1,1,1,1,5", @@ -648,12 +648,12 @@ void ptx_instruction::set_bar_type() void ptx_instruction::set_opcode_and_latency() { - unsigned int_latency[5]; + unsigned int_latency[6]; unsigned fp_latency[5]; unsigned dp_latency[5]; unsigned sfu_latency; unsigned tensor_latency; - unsigned int_init[5]; + unsigned int_init[6]; unsigned fp_init[5]; unsigned dp_init[5]; unsigned sfu_init; @@ -664,10 +664,11 @@ void ptx_instruction::set_opcode_and_latency() * [2] MUL * [3] MAD * [4] DIV + * [5] SHFL */ - sscanf(opcode_latency_int, "%u,%u,%u,%u,%u", + sscanf(opcode_latency_int, "%u,%u,%u,%u,%u,%u", &int_latency[0],&int_latency[1],&int_latency[2], - &int_latency[3],&int_latency[4]); + &int_latency[3],&int_latency[4],&int_latency[5]); sscanf(opcode_latency_fp, "%u,%u,%u,%u,%u", &fp_latency[0],&fp_latency[1],&fp_latency[2], &fp_latency[3],&fp_latency[4]); @@ -678,9 +679,9 @@ void ptx_instruction::set_opcode_and_latency() &sfu_latency); sscanf(opcode_latency_tensor, "%u", &tensor_latency); - sscanf(opcode_initiation_int, "%u,%u,%u,%u,%u", + sscanf(opcode_initiation_int, "%u,%u,%u,%u,%u,%u", &int_init[0],&int_init[1],&int_init[2], - &int_init[3],&int_init[4]); + &int_init[3],&int_init[4],&int_init[5]); sscanf(opcode_initiation_fp, "%u,%u,%u,%u,%u", &fp_init[0],&fp_init[1],&fp_init[2], &fp_init[3],&fp_init[4]); @@ -873,8 +874,8 @@ void ptx_instruction::set_opcode_and_latency() op=TENSOR_CORE_OP; break; case SHFL_OP: - latency = 32; - initiation_interval = 4; + latency = int_latency[5]; + initiation_interval = int_init[5]; break; default: break; diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index 007ad42..96ba385 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -3002,7 +3002,7 @@ void shader_core_config::set_pipeline_latency() { //calculate the max latency based on the input - unsigned int_latency[5]; + unsigned int_latency[6]; unsigned fp_latency[5]; unsigned dp_latency[5]; unsigned sfu_latency; @@ -3014,10 +3014,11 @@ void shader_core_config::set_pipeline_latency() { * [2] MUL * [3] MAD * [4] DIV + * [5] SHFL */ - sscanf(opcode_latency_int, "%u,%u,%u,%u,%u", + sscanf(opcode_latency_int, "%u,%u,%u,%u,%u,%u", &int_latency[0],&int_latency[1],&int_latency[2], - &int_latency[3],&int_latency[4]); + &int_latency[3],&int_latency[4],&int_latency[5]); sscanf(opcode_latency_fp, "%u,%u,%u,%u,%u", &fp_latency[0],&fp_latency[1],&fp_latency[2], &fp_latency[3],&fp_latency[4]); @@ -3034,7 +3035,7 @@ void shader_core_config::set_pipeline_latency() { max_sfu_latency = std::max(dp_latency[4],sfu_latency); //assume that the max operation has the max latency max_sp_latency = fp_latency[1]; - max_int_latency = int_latency[1]; + max_int_latency = std::max(int_latency[1],int_latency[5]); max_dp_latency = dp_latency[1]; max_tensor_core_latency = tensor_latency; -- cgit v1.3 From f829a12add380c78e50ecc86357d4a1b513f21f0 Mon Sep 17 00:00:00 2001 From: boyealkim Date: Thu, 25 Jul 2019 20:42:13 +0900 Subject: fix stream_merge front algo --- src/stream_manager.cc | 50 +++++++++++++++++++++++++++++++------------------- src/stream_manager.h | 12 ++++++++++-- 2 files changed, 41 insertions(+), 21 deletions(-) (limited to 'src') diff --git a/src/stream_manager.cc b/src/stream_manager.cc index 6cd62a2..6e02dcd 100644 --- a/src/stream_manager.cc +++ b/src/stream_manager.cc @@ -194,14 +194,17 @@ bool stream_operation::do_operation( gpgpu_sim *gpu ) m_stream->record_next_done(); } break; - case stream_wait_event: { + case stream_wait_event: //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; + } break; default: abort(); @@ -232,6 +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(); } bool stream_manager::operation( bool * sim) @@ -330,22 +334,31 @@ stream_operation stream_manager::front() m_service_stream_zero = false; } } - if(!m_service_stream_zero) { - std::list::iterator s; - for( s=m_streams.begin(); s != m_streams.end(); 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; } @@ -370,7 +383,8 @@ void stream_manager::destroy_stream( CUstream_st *stream ) break; } } - delete stream; + delete stream; + m_last_stream = m_streams.begin(); pthread_mutex_unlock(&m_lock); } @@ -384,7 +398,6 @@ 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; } @@ -437,7 +450,6 @@ 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 91d1b36..ae1af9f 100644 --- a/src/stream_manager.h +++ b/src/stream_manager.h @@ -163,7 +163,6 @@ 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; @@ -179,9 +178,9 @@ 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 { @@ -193,6 +192,7 @@ public: m_updates = 0; m_wallclock = 0; m_gpu_tot_sim_cycle = 0; + m_issued = 0; m_done = false; } void update( double cycle, time_t clk ) @@ -207,11 +207,18 @@ 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; + } private: int m_uid; bool m_blocking; bool m_done; int m_updates; + unsigned int m_issued; time_t m_wallclock; double m_gpu_tot_sim_cycle; @@ -268,6 +275,7 @@ private: CUstream_st m_stream_zero; bool m_service_stream_zero; pthread_mutex_t m_lock; + std::list::iterator m_last_stream; }; #endif -- 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 'src') 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 f3ec23390a0798eab1426adf962487680ea89e93 Mon Sep 17 00:00:00 2001 From: boyealkim Date: Sat, 27 Jul 2019 03:54:44 +0900 Subject: fix event done impl --- src/stream_manager.cc | 2 +- src/stream_manager.h | 80 +++++++++++++++++++++++++-------------------------- 2 files changed, 41 insertions(+), 41 deletions(-) (limited to 'src') diff --git a/src/stream_manager.cc b/src/stream_manager.cc index d0eecdd..0d67e10 100644 --- a/src/stream_manager.cc +++ b/src/stream_manager.cc @@ -198,7 +198,7 @@ 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->num_updates()>=m_cnt){ printf("stream wait event done\n"); m_stream->record_next_done(); } diff --git a/src/stream_manager.h b/src/stream_manager.h index eb4c6ff..e73d71a 100644 --- a/src/stream_manager.h +++ b/src/stream_manager.h @@ -43,6 +43,45 @@ // unsigned m_pending_streams; //}; +struct CUevent_st { +public: + CUevent_st( bool blocking ) + { + m_uid = ++m_next_event_uid; + m_blocking = blocking; + m_updates = 0; + m_wallclock = 0; + m_gpu_tot_sim_cycle = 0; + m_issued = 0; + m_done = false; + } + void update( double cycle, time_t clk ) + { + m_updates++; + m_wallclock=clk; + m_gpu_tot_sim_cycle=cycle; + m_done = true; + } + //void set_done() { assert(!m_done); m_done=true; } + int get_uid() const { return m_uid; } + unsigned num_updates() const { return m_updates; } + bool done() const { return m_updates==m_issued; } + time_t clock() const { return m_wallclock; } + 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; + time_t m_wallclock; + double m_gpu_tot_sim_cycle; + + static int m_next_event_uid; +}; + + enum stream_operation_type { stream_no_op, stream_memcpy_host_to_device, @@ -107,6 +146,7 @@ public: m_kernel=NULL; m_type=stream_wait_event; m_event=e; + m_cnt = m_event->num_issued(); m_stream=stream; m_done=false; } @@ -163,7 +203,6 @@ 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; @@ -183,45 +222,6 @@ private: kernel_info_t *m_kernel; struct CUevent_st *m_event; }; - -struct CUevent_st { -public: - CUevent_st( bool blocking ) - { - m_uid = ++m_next_event_uid; - m_blocking = blocking; - m_updates = 0; - m_wallclock = 0; - m_gpu_tot_sim_cycle = 0; - m_issued = 0; - m_done = false; - } - void update( double cycle, time_t clk ) - { - m_updates++; - m_wallclock=clk; - m_gpu_tot_sim_cycle=cycle; - m_done = true; - } - //void set_done() { assert(!m_done); m_done=true; } - int get_uid() const { return m_uid; } - 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; } -private: - int m_uid; - bool m_blocking; - bool m_done; - int m_updates; - unsigned int m_issued; - time_t m_wallclock; - double m_gpu_tot_sim_cycle; - - static int m_next_event_uid; -}; - struct CUstream_st { public: CUstream_st(); -- cgit v1.3