From 6dfb3ba656b09acafdb274d7ab04fb081ca06e57 Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 8 Nov 2017 20:13:28 -0500 Subject: Fixing the break limit bug --- src/gpgpusim_entrypoint.cc | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) (limited to 'src') diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index 04845e7..ad4587a 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -91,6 +91,7 @@ void *gpgpu_sim_thread_sequential(void*) pthread_mutex_t g_sim_lock = PTHREAD_MUTEX_INITIALIZER; bool g_sim_active = false; bool g_sim_done = true; +bool break_limit = false; void *gpgpu_sim_thread_concurrent(void*) { @@ -144,11 +145,13 @@ void *gpgpu_sim_thread_concurrent(void*) if(g_the_gpu->cycle_insn_cta_max_hit()){ g_stream_manager->stop_all_running_kernels(); g_sim_done = true; + break_limit = true; } } active=g_the_gpu->active() || !g_stream_manager->empty_protected(); - } while( active ); + + } while( active && !g_sim_done); if(g_debug_execution >= 3) { printf("GPGPU-Sim: ** STOP simulation thread (no work) **\n"); fflush(stdout); @@ -166,6 +169,11 @@ void *gpgpu_sim_thread_concurrent(void*) printf("GPGPU-Sim: *** simulation thread exiting ***\n"); fflush(stdout); } + if(break_limit) { + printf("GPGPU-Sim: ** break due to reaching the maximum cycles (or instructions) **\n"); + exit(1); + } + sem_post(&g_sim_signal_exit); return NULL; } @@ -179,7 +187,7 @@ void synchronize() bool done = false; do { pthread_mutex_lock(&g_sim_lock); - done = g_stream_manager->empty() && !g_sim_active; + done = ( g_stream_manager->empty() && !g_sim_active ) || g_sim_done; pthread_mutex_unlock(&g_sim_lock); } while (!done); printf("GPGPU-Sim: detected inactive GPU simulation thread\n"); -- cgit v1.3 From b4cf66026f66291293c6e11e9c07976f773732fa Mon Sep 17 00:00:00 2001 From: Mengchi Zhang Date: Tue, 3 Oct 2017 11:36:52 -0400 Subject: Add lonestar tick support Signed-off-by: Mengchi Zhang --- cuobjdump_to_ptxplus/ptx_parser.h | 4 ++++ libcuda/cuda_runtime_api.cc | 49 +++++++++++++++++++++++++-------------- src/abstract_hardware_model.h | 10 +++++++- src/cuda-sim/cuda-sim.cc | 5 ++-- src/cuda-sim/ptx.y | 3 ++- src/cuda-sim/ptx_ir.cc | 1 + src/cuda-sim/ptx_ir.h | 7 ++++++ src/cuda-sim/ptx_parser.cc | 4 ++++ src/cuda-sim/ptx_parser.h | 1 + 9 files changed, 62 insertions(+), 22 deletions(-) (limited to 'src') diff --git a/cuobjdump_to_ptxplus/ptx_parser.h b/cuobjdump_to_ptxplus/ptx_parser.h index 418a733..729eaec 100644 --- a/cuobjdump_to_ptxplus/ptx_parser.h +++ b/cuobjdump_to_ptxplus/ptx_parser.h @@ -373,4 +373,8 @@ void func_header_info_int(const char* s, int i) g_headerList->getListEnd().addOperand(buff); } } + +void maxnt_id(int x, int y, int z) { + +} #endif //_PTX_PARSER_H_ diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index d8e7832..4f5e0fb 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -336,23 +336,19 @@ class _cuda_device_id *GPGPUSim_Init() prop->minor = 2; prop->totalGlobalMem = 0x80000000 /* 2 GB */; prop->memPitch = 0; + if(prop->major >= 2) { + prop->maxThreadsPerBlock = 1024; + prop->maxThreadsDim[0] = 1024; + prop->maxThreadsDim[1] = 1024; + } + else + { + prop->maxThreadsPerBlock = 512; + prop->maxThreadsDim[0] = 512; + prop->maxThreadsDim[1] = 512; + } - if(prop->major >= 2) { - prop->maxThreadsPerBlock = 1024; - prop->maxThreadsDim[0] = 1024; - prop->maxThreadsDim[1] = 1024; - } - else - { - prop->maxThreadsPerBlock = 512; - prop->maxThreadsDim[0] = 512; - prop->maxThreadsDim[1] = 512; - } - - prop->maxThreadsDim[2] = 64; - prop->maxGridSize[0] = 0x40000000; - prop->maxGridSize[1] = 0x40000000; - prop->maxGridSize[2] = 0x40000000; + prop->maxThreadsDim[2] = 64; prop->maxGridSize[0] = 0x40000000; prop->maxGridSize[1] = 0x40000000; prop->maxGridSize[2] = 0x40000000; @@ -366,7 +362,7 @@ class _cuda_device_id *GPGPUSim_Init() prop->multiProcessorCount = the_gpu->get_config().num_shader(); #endif #if (CUDART_VERSION >= 4000) - prop->maxThreadsPerMultiProcessor = the_gpu->threads_per_core(); + prop->maxThreadsPerMultiProcessor = the_gpu->threads_per_core(); #endif the_gpu->set_prop(prop); the_device = new _cuda_device_id(the_gpu); @@ -2112,6 +2108,20 @@ cudaError_t CUDARTAPI cudaSetDeviceFlags( int flags ) return g_last_cudaError = cudaErrorUnknown; } +size_t getMaxThreadsPerBlock(struct cudaFuncAttributes *attr) { + _cuda_device_id *dev = GPGPUSim_Init(); + struct cudaDeviceProp prop; + + prop = *dev->get_prop(); + + size_t max = prop.maxThreadsPerBlock; + + if ((prop.regsPerBlock / attr->numRegs) < max) + max = prop.regsPerBlock / attr->numRegs; + + return max; +} + cudaError_t CUDARTAPI cudaFuncGetAttributes(struct cudaFuncAttributes *attr, const char *hostFun ) { CUctx_st *context = GPGPUSim_Context(); @@ -2122,7 +2132,10 @@ cudaError_t CUDARTAPI cudaFuncGetAttributes(struct cudaFuncAttributes *attr, con attr->constSizeBytes = kinfo->cmem; attr->localSizeBytes = kinfo->lmem; attr->numRegs = kinfo->regs; - attr->maxThreadsPerBlock = 0; // from pragmas? + if(kinfo->maxthreads > 0) + attr->maxThreadsPerBlock = kinfo->maxthreads; + else + attr->maxThreadsPerBlock = getMaxThreadsPerBlock(attr); #if CUDART_VERSION >= 3000 attr->ptxVersion = kinfo->ptx_version; attr->binaryVersion = kinfo->sm_target; diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index cec75f9..1b764e2 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -517,7 +517,14 @@ public: const struct textureReference* get_texref(const std::string &texname) const { std::map::const_iterator t=m_NameToTextureRef.find(texname); - assert( t != m_NameToTextureRef.end() ); + if( t == m_NameToTextureRef.end() ) { + // search for :: prefixed names + std::string temp("::" + texname); + t=m_NameToTextureRef.find(temp); + } + + assert(t != m_NameToTextureRef.end()); + return t->second; } const struct cudaArray* get_texarray( const struct textureReference *texref ) const @@ -568,6 +575,7 @@ struct gpgpu_ptx_sim_info int cmem; int gmem; int regs; + unsigned maxthreads; unsigned ptx_version; unsigned sm_target; }; diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 2f166aa..f143d35 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -1159,13 +1159,14 @@ void function_info::finalize( memory_space *param_mem ) // copy the parameter over word-by-word so that parameter that crosses a memory page can be copied over //Jin: copy parameter using aligned rules const size_t word_size = 4; - param_address = (param_address + size - 1) / size * size; //aligned with size + //param_address = (param_address + size - 1) / size * size; //aligned with size TODO: align not correct for (size_t idx = 0; idx < size; idx += word_size) { const char *pdata = reinterpret_cast(param_value.pdata) + idx; // cast to char * for ptr arithmetic param_mem->write(param_address + idx, word_size, pdata,NULL,NULL); + param_mem->read(param_address + idx, word_size, &data); } unsigned offset = p.get_offset(); - assert(offset == param_address); + //assert(offset == param_address); param->set_address(param_address); param_address += size; } diff --git a/src/cuda-sim/ptx.y b/src/cuda-sim/ptx.y index e00aa4b..c0c58a6 100644 --- a/src/cuda-sim/ptx.y +++ b/src/cuda-sim/ptx.y @@ -228,7 +228,8 @@ function_defn: function_decl { set_symtab($1); func_header(".skip"); } statement block_spec: MAXNTID_DIRECTIVE INT_OPERAND COMMA INT_OPERAND COMMA INT_OPERAND {func_header_info_int(".maxntid", $2); func_header_info_int(",", $4); - func_header_info_int(",", $6); } + func_header_info_int(",", $6); + maxnt_id($2, $4, $6);} | MINNCTAPERSM_DIRECTIVE INT_OPERAND { func_header_info_int(".minnctapersm", $2); printf("GPGPU-Sim: Warning: .minnctapersm ignored. \n"); } | MAXNCTAPERSM_DIRECTIVE INT_OPERAND { func_header_info_int(".maxnctapersm", $2); printf("GPGPU-Sim: Warning: .maxnctapersm ignored. \n"); } ; diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index 8ebdcf8..ee36957 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -222,6 +222,7 @@ bool symbol_table::add_function_decl( const char *name, int entry_point, functio } else { *func_info = new function_info(entry_point); (*func_info)->set_name(name); + (*func_info)->set_maxnt_id(0); m_function_info_lookup[key] = *func_info; } diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 9ad1571..36ef3d5 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -1245,6 +1245,7 @@ public: const struct gpgpu_ptx_sim_info* get_kernel_info () const { + assert (m_kernel_info.maxthreads == maxnt_id); return &m_kernel_info; } @@ -1252,6 +1253,8 @@ public: m_kernel_info = info; m_kernel_info.ptx_version = 10*get_ptx_version().ver(); m_kernel_info.sm_target = get_ptx_version().target(); + // THIS DEPENDS ON ptxas being called after the PTX is parsed. + m_kernel_info.maxthreads = maxnt_id; } symbol_table *get_symtab() { @@ -1275,7 +1278,11 @@ public: } bool is_entry_point() const { return m_entry_point; } + void set_maxnt_id(unsigned maxthreads) { maxnt_id = maxthreads;} + unsigned get_maxnt_id() { return maxnt_id;} + private: + unsigned maxnt_id; unsigned m_uid; unsigned m_local_mem_framesize; bool m_entry_point; diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc index baa3bcd..e5731a8 100644 --- a/src/cuda-sim/ptx_parser.cc +++ b/src/cuda-sim/ptx_parser.cc @@ -969,6 +969,10 @@ void target_header3(char* a, char* b, char* c) g_global_symbol_table->set_sm_target(a,b,c); } +void maxnt_id(int x, int y, int z) { + g_func_info->set_maxnt_id(x * y * z); +} + void func_header(const char* a) {} //intentional dummy function void func_header_info(const char* a) {} //intentional dummy function void func_header_info_int(const char* a, int b) {} //intentional dummy function diff --git a/src/cuda-sim/ptx_parser.h b/src/cuda-sim/ptx_parser.h index 32f3903..13042e1 100644 --- a/src/cuda-sim/ptx_parser.h +++ b/src/cuda-sim/ptx_parser.h @@ -93,6 +93,7 @@ void change_double_operand_type( int addr_type ); void change_operand_neg( ); void set_immediate_operand_type( ); void version_header(double a); +void maxnt_id(int x, int y, int z); //Jin: handle instructino group for cdp void start_inst_group(); -- cgit v1.3 From 78a971019b7eac438521c9a77fdf9de30340f29e Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sun, 25 Mar 2018 15:26:38 -0400 Subject: Fixing a bug in mengchi's commit --- src/cuda-sim/cuda-sim.cc | 1 - 1 file changed, 1 deletion(-) (limited to 'src') diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index f143d35..3d3a820 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -1163,7 +1163,6 @@ void function_info::finalize( memory_space *param_mem ) for (size_t idx = 0; idx < size; idx += word_size) { const char *pdata = reinterpret_cast(param_value.pdata) + idx; // cast to char * for ptr arithmetic param_mem->write(param_address + idx, word_size, pdata,NULL,NULL); - param_mem->read(param_address + idx, word_size, &data); } unsigned offset = p.get_offset(); //assert(offset == param_address); -- cgit v1.3 From 4e91a60a48b07f41f4bfb4d59fa2355024a3914b Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sun, 25 Mar 2018 16:51:21 -0400 Subject: Need to make sure we don't kill the L1 too --- configs/Pascal-P100-HBM/gpgpusim.config | 2 ++ configs/Pascal-P102-GDDR5X/gpgpusim.config | 2 ++ configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config | 2 ++ configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config | 2 ++ src/gpgpu-sim/gpu-sim.cc | 2 +- 5 files changed, 9 insertions(+), 1 deletion(-) (limited to 'src') diff --git a/configs/Pascal-P100-HBM/gpgpusim.config b/configs/Pascal-P100-HBM/gpgpusim.config index 31fbecb..857a0ff 100644 --- a/configs/Pascal-P100-HBM/gpgpusim.config +++ b/configs/Pascal-P100-HBM/gpgpusim.config @@ -55,6 +55,8 @@ # Note: Hashing set index function (H) only applies to a set size of 32 or 64. # Pascal GP100 has 64KB Shared memory -gpgpu_cache:dl1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:64:128:6,L:L:f:N:H,A:256:8,16:0,32 -gpgpu_shmem_size 65536 -gpgpu_shmem_size_PrefL1 65536 -gpgpu_shmem_size_PrefShared 65536 diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config index 8e1e56f..feb8c0b 100644 --- a/configs/Pascal-P102-GDDR5X/gpgpusim.config +++ b/configs/Pascal-P102-GDDR5X/gpgpusim.config @@ -62,6 +62,8 @@ # Pascal GP102 has 24KB L1 cache # The defulat is to disable the L1 cache, unless cache modifieres is used -gpgpu_cache:dl1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefL1 S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 +-gpgpu_cache:dl1PrefShared S:32:128:6,L:L:f:N:H,A:256:8,16:0,32 -gpgpu_shmem_size 49152 -gpgpu_shmem_size_PrefL1 49152 -gpgpu_shmem_size_PrefShared 49152 diff --git a/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config index b87d958..a5e6736 100644 --- a/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config +++ b/configs/PascalP100X/gpgpu-sim-3.x/gpgpusim.config @@ -55,6 +55,8 @@ # Pascal GP100 has 64KB Shared memory # Pascal GP100 has 48KB L1 cache -gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 -gpgpu_shmem_size 65536 -gpgpu_shmem_size_PrefL1 65536 -gpgpu_shmem_size_PrefShared 65536 diff --git a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config index 2df24e1..f78bd02 100644 --- a/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config +++ b/configs/PascalTitanX/gpgpu-sim-3.x/gpgpusim.config @@ -56,6 +56,8 @@ # Pascal GP102 has 64KB L1 cache # The defulat is to disable the L1 cache, unless cache modifieres is used -gpgpu_cache:dl1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefL1 64:128:6,L:L:m:N:H,A:128:8,16 +-gpgpu_cache:dl1PrefShared 64:128:6,L:L:m:N:H,A:128:8,16 -gpgpu_shmem_size 98304 -gpgpu_shmem_size_PrefL1 98304 -gpgpu_shmem_size_PrefShared 98304 diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 17f1714..bb448b3 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -245,7 +245,7 @@ void shader_core_config::reg_options(class OptionParser * opp) "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); - option_parser_register(opp, "-gpgpu_cache:dl1PreShared", OPT_CSTR, &m_L1D_config.m_config_stringPrefShared, + option_parser_register(opp, "-gpgpu_cache:dl1PrefShared", OPT_CSTR, &m_L1D_config.m_config_stringPrefShared, "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); -- cgit v1.3 From 1af82fbe97428654b06b7cd8d40c2d2ce4592aae Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Mon, 26 Mar 2018 11:35:25 -0400 Subject: Fixing resources limitation assertion for compute cab 61 --- configs/Pascal-P102-GDDR5X/gpgpusim.config | 1 + src/gpgpu-sim/gpu-sim.cc | 3 +++ src/gpgpu-sim/shader.cc | 4 ++++ src/gpgpu-sim/shader.h | 1 + 4 files changed, 9 insertions(+) (limited to 'src') diff --git a/configs/Pascal-P102-GDDR5X/gpgpusim.config b/configs/Pascal-P102-GDDR5X/gpgpusim.config index 0c6c126..b7bcdaf 100644 --- a/configs/Pascal-P102-GDDR5X/gpgpusim.config +++ b/configs/Pascal-P102-GDDR5X/gpgpusim.config @@ -6,6 +6,7 @@ -gpgpu_ptx_instruction_classification 0 -gpgpu_ptx_sim_mode 0 -gpgpu_ptx_force_max_capability 61 +-gpgpu_ignore_resources_limitation 1 # SASS execution (only supported with CUDA >= 4.0) -gpgpu_ptx_convert_to_ptxplus 0 diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 17f1714..48af41e 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -268,6 +268,9 @@ void shader_core_config::reg_options(class OptionParser * opp) option_parser_register(opp, "-gpgpu_shader_registers", OPT_UINT32, &gpgpu_shader_registers, "Number of registers per shader core. Limits number of concurrent CTAs. (default 8192)", "8192"); + option_parser_register(opp, "-gpgpu_ignore_resources_limitation", OPT_BOOL, &gpgpu_ignore_resources_limitation, + "gpgpu_ignore_resources_limitation (default 0)", + "0"); option_parser_register(opp, "-gpgpu_shader_cta", OPT_UINT32, &max_cta_per_core, "Maximum number of concurrent CTAs in shader (default 8)", "8"); diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index bf482fb..8173191 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -2614,6 +2614,10 @@ unsigned int shader_core_config::max_cta( const kernel_info_t &k ) const assert( result <= MAX_CTA_PER_SHADER ); if (result < 1) { printf ("GPGPU-Sim uArch: ERROR ** Kernel requires more resources than shader has.\n"); + if(gpgpu_ignore_resources_limitation) { + printf ("GPGPU-Sim uArch: gpgpu_ignore_resources_limitation is set, ignore the ERROR!\n"); + return 1; + } abort(); } diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index 5b41c06..ae22eaa 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -1386,6 +1386,7 @@ struct shader_core_config : public core_config unsigned gpgpu_num_reg_banks; bool gpgpu_reg_bank_use_warp_id; bool gpgpu_local_mem_map; + bool gpgpu_ignore_resources_limitation; unsigned max_sp_latency; unsigned max_sfu_latency; -- cgit v1.3 From 67a257442f46a9e8e02f63a9d71fcc7ca54f3f5e Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 28 Mar 2018 14:00:16 -0400 Subject: adding new counters for parition level para and L2 BW --- src/gpgpu-sim/gpu-sim.cc | 36 ++++++++++++++++++++++++++++++++++++ 1 file changed, 36 insertions(+) (limited to 'src') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 48af41e..228d4be 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -87,6 +87,12 @@ unsigned long long gpu_tot_sim_cycle = 0; // performance counter for stalls due to congestion. unsigned int gpu_stall_dramfull = 0; unsigned int gpu_stall_icnt2sh = 0; +unsigned long long partiton_reqs_in_parallel = 0; +unsigned long long partiton_reqs_in_parallel_total = 0; +unsigned long long partiton_reqs_in_parallel_util = 0; +unsigned long long partiton_reqs_in_parallel_util_total = 0; +unsigned long long partiton_replys_in_parallel = 0; +unsigned long long partiton_replys_in_parallel_total = 0; /* Clock Domains */ @@ -786,6 +792,9 @@ void gpgpu_sim::init() gpu_sim_insn = 0; last_gpu_sim_insn = 0; m_total_cta_launched=0; + partiton_reqs_in_parallel = 0; + partiton_replys_in_parallel = 0; + partiton_reqs_in_parallel_util = 0; reinit_clock_domains(); set_param_gpgpu_num_shaders(m_config.num_shader()); @@ -822,8 +831,14 @@ void gpgpu_sim::update_stats() { gpu_tot_sim_cycle += gpu_sim_cycle; gpu_tot_sim_insn += gpu_sim_insn; gpu_tot_issued_cta += m_total_cta_launched; + partiton_reqs_in_parallel_total += partiton_reqs_in_parallel; + partiton_replys_in_parallel_total += partiton_replys_in_parallel; + partiton_reqs_in_parallel_util_total += partiton_reqs_in_parallel_util; gpu_sim_cycle = 0; + partiton_reqs_in_parallel = 0; + partiton_replys_in_parallel = 0; + partiton_reqs_in_parallel_util = 0; gpu_sim_insn = 0; m_total_cta_launched = 0; } @@ -1007,6 +1022,19 @@ void gpgpu_sim::gpu_print_stat() printf("gpu_stall_dramfull = %d\n", gpu_stall_dramfull); printf("gpu_stall_icnt2sh = %d\n", gpu_stall_icnt2sh ); + printf("partiton_reqs_in_parallel = %lld\n", partiton_reqs_in_parallel); + printf("partiton_reqs_in_parallel_total = %lld\n", partiton_reqs_in_parallel_total ); + printf("partiton_level_parallism = %12.4f\n", (float)partiton_reqs_in_parallel / gpu_sim_cycle); + printf("partiton_level_parallism_total = %12.4f\n", (float)(partiton_reqs_in_parallel+partiton_reqs_in_parallel_total) / (gpu_tot_sim_cycle+gpu_sim_cycle) ); + printf("partiton_reqs_in_parallel_util = %lld\n", partiton_reqs_in_parallel_util); + printf("partiton_reqs_in_parallel_util_total = %lld\n", partiton_reqs_in_parallel_util_total ); + printf("partiton_level_parallism_util = %12.4f\n", (float)partiton_reqs_in_parallel_util / gpu_sim_cycle); + printf("partiton_level_parallism_util_total = %12.4f\n", (float)(partiton_reqs_in_parallel_util+partiton_reqs_in_parallel_util_total) / (gpu_tot_sim_cycle+gpu_sim_cycle) ); + printf("partiton_replys_in_parallel = %lld\n", partiton_replys_in_parallel); + printf("partiton_replys_in_parallel_total = %lld\n", partiton_replys_in_parallel_total ); + printf("L2_BW = %12.4f GB/Sec\n", ((float)(partiton_replys_in_parallel * 32) / (gpu_sim_cycle * m_config.icnt_period)) / 1000000000); + printf("L2_BW_total = %12.4f GB/Sec\n", ((float)((partiton_replys_in_parallel+partiton_replys_in_parallel_total) * 32) / ((gpu_tot_sim_cycle+gpu_sim_cycle) * m_config.icnt_period)) / 1000000000 ); + time_t curr_time; time(&curr_time); unsigned long long elapsed_time = MAX( curr_time - g_simulation_starttime, 1 ); @@ -1412,6 +1440,7 @@ void gpgpu_sim::cycle() for (unsigned i=0;in_simt_clusters;i++) m_cluster[i]->icnt_cycle(); } + unsigned partiton_replys_in_parallel_per_cycle = 0; if (clock_mask & ICNT) { // pop from memory controller to interconnect for (unsigned i=0;im_n_mem_sub_partition;i++) { @@ -1424,6 +1453,7 @@ void gpgpu_sim::cycle() mf->set_status(IN_ICNT_TO_SHADER,gpu_sim_cycle+gpu_tot_sim_cycle); ::icnt_push( m_shader_config->mem2device(i), mf->get_tpc(), mf, response_size ); m_memory_sub_partition[i]->pop(); + partiton_replys_in_parallel_per_cycle++; } else { gpu_stall_icnt2sh++; } @@ -1432,6 +1462,7 @@ void gpgpu_sim::cycle() } } } + partiton_replys_in_parallel += partiton_replys_in_parallel_per_cycle; if (clock_mask & DRAM) { for (unsigned i=0;im_n_mem;i++){ @@ -1444,6 +1475,7 @@ void gpgpu_sim::cycle() } // L2 operations follow L2 clock domain + unsigned partiton_reqs_in_parallel_per_cycle = 0; if (clock_mask & L2) { m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX].clear(); for (unsigned i=0;im_n_mem_sub_partition;i++) { @@ -1455,11 +1487,15 @@ void gpgpu_sim::cycle() } else { mem_fetch* mf = (mem_fetch*) icnt_pop( m_shader_config->mem2device(i) ); m_memory_sub_partition[i]->push( mf, gpu_sim_cycle + gpu_tot_sim_cycle ); + partiton_reqs_in_parallel_per_cycle++; } m_memory_sub_partition[i]->cache_cycle(gpu_sim_cycle+gpu_tot_sim_cycle); m_memory_sub_partition[i]->accumulate_L2cache_stats(m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX]); } } + partiton_reqs_in_parallel += partiton_reqs_in_parallel_per_cycle; + if(partiton_reqs_in_parallel_per_cycle > 0) + partiton_reqs_in_parallel_util += partiton_reqs_in_parallel_per_cycle; if (clock_mask & ICNT) { icnt_transfer(); -- cgit v1.3 From 93cdd2e217d7903ccf715a00b6dae8265bb38fca Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Wed, 28 Mar 2018 15:22:33 -0400 Subject: adding new stats counter --- src/gpgpu-sim/gpu-sim.cc | 16 +++++++++++++--- 1 file changed, 13 insertions(+), 3 deletions(-) (limited to 'src') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index f216e7c..006d243 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -91,6 +91,8 @@ unsigned long long partiton_reqs_in_parallel = 0; unsigned long long partiton_reqs_in_parallel_total = 0; unsigned long long partiton_reqs_in_parallel_util = 0; unsigned long long partiton_reqs_in_parallel_util_total = 0; +unsigned long long gpu_sim_cycle_parition_util = 0; +unsigned long long gpu_tot_sim_cycle_parition_util = 0; unsigned long long partiton_replys_in_parallel = 0; unsigned long long partiton_replys_in_parallel_total = 0; @@ -795,6 +797,7 @@ void gpgpu_sim::init() partiton_reqs_in_parallel = 0; partiton_replys_in_parallel = 0; partiton_reqs_in_parallel_util = 0; + gpu_sim_cycle_parition_util = 0; reinit_clock_domains(); set_param_gpgpu_num_shaders(m_config.num_shader()); @@ -834,11 +837,14 @@ void gpgpu_sim::update_stats() { partiton_reqs_in_parallel_total += partiton_reqs_in_parallel; partiton_replys_in_parallel_total += partiton_replys_in_parallel; partiton_reqs_in_parallel_util_total += partiton_reqs_in_parallel_util; + gpu_tot_sim_cycle_parition_util += gpu_sim_cycle_parition_util ; gpu_sim_cycle = 0; partiton_reqs_in_parallel = 0; partiton_replys_in_parallel = 0; partiton_reqs_in_parallel_util = 0; + gpu_sim_cycle_parition_util = 0; + gpu_tot_sim_cycle = 0; gpu_sim_insn = 0; m_total_cta_launched = 0; } @@ -1028,8 +1034,10 @@ void gpgpu_sim::gpu_print_stat() printf("partiton_level_parallism_total = %12.4f\n", (float)(partiton_reqs_in_parallel+partiton_reqs_in_parallel_total) / (gpu_tot_sim_cycle+gpu_sim_cycle) ); printf("partiton_reqs_in_parallel_util = %lld\n", partiton_reqs_in_parallel_util); printf("partiton_reqs_in_parallel_util_total = %lld\n", partiton_reqs_in_parallel_util_total ); - printf("partiton_level_parallism_util = %12.4f\n", (float)partiton_reqs_in_parallel_util / gpu_sim_cycle); - printf("partiton_level_parallism_util_total = %12.4f\n", (float)(partiton_reqs_in_parallel_util+partiton_reqs_in_parallel_util_total) / (gpu_tot_sim_cycle+gpu_sim_cycle) ); + printf("gpu_sim_cycle_parition_util = %lld\n", gpu_sim_cycle_parition_util); + printf("gpu_tot_sim_cycle_parition_util = %lld\n", gpu_tot_sim_cycle_parition_util ); + printf("partiton_level_parallism_util = %12.4f\n", (float)partiton_reqs_in_parallel_util / gpu_sim_cycle_parition_util); + printf("partiton_level_parallism_util_total = %12.4f\n", (float)(partiton_reqs_in_parallel_util+partiton_reqs_in_parallel_util_total) / (gpu_sim_cycle_parition_util+gpu_tot_sim_cycle_parition_util) ); printf("partiton_replys_in_parallel = %lld\n", partiton_replys_in_parallel); printf("partiton_replys_in_parallel_total = %lld\n", partiton_replys_in_parallel_total ); printf("L2_BW = %12.4f GB/Sec\n", ((float)(partiton_replys_in_parallel * 32) / (gpu_sim_cycle * m_config.icnt_period)) / 1000000000); @@ -1494,8 +1502,10 @@ void gpgpu_sim::cycle() } } partiton_reqs_in_parallel += partiton_reqs_in_parallel_per_cycle; - if(partiton_reqs_in_parallel_per_cycle > 0) + if(partiton_reqs_in_parallel_per_cycle > 0){ partiton_reqs_in_parallel_util += partiton_reqs_in_parallel_per_cycle; + gpu_sim_cycle_parition_util++; + } if (clock_mask & ICNT) { icnt_transfer(); -- cgit v1.3 From c22eb8a8f6e41921e4e1a5e705cca8d5bbed67e0 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Wed, 28 Mar 2018 15:29:46 -0400 Subject: gpuwaatch build does not parallelize nicely without -p --- src/gpuwattch/makefile | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'src') diff --git a/src/gpuwattch/makefile b/src/gpuwattch/makefile index ab718cc..354c9ec 100644 --- a/src/gpuwattch/makefile +++ b/src/gpuwattch/makefile @@ -11,10 +11,10 @@ opt: $(TAR).mk obj_opt @$(MAKE) TAG=opt -C . -f $(TAR).mk obj_dbg: - mkdir $@ + mkdir -p $@ obj_opt: - mkdir $@ + mkdir -p $@ depend: @$(MAKE) TAG=opt -C . -f $(TAR).mk depend -- cgit v1.3 From 21ac2041365dc8521a6926244b96e85fd19cc2d3 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Wed, 28 Mar 2018 22:52:12 -0400 Subject: Fixing a minor typo in an old config --- src/gpgpu-sim/gpu-sim.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'src') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 58a5d16..d452888 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -223,7 +223,7 @@ void shader_core_config::reg_options(class OptionParser * opp) "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); - option_parser_register(opp, "-gpgpu_cache:dl1PreShared", OPT_CSTR, &m_L1D_config.m_config_stringPrefShared, + option_parser_register(opp, "-gpgpu_cache:dl1PrefShared", OPT_CSTR, &m_L1D_config.m_config_stringPrefShared, "per-shader L1 data cache config " " {::,:::,::, | none}", "none" ); -- cgit v1.3 From 3b9cc9285b5394320873e42edee03d8f23d6d17e Mon Sep 17 00:00:00 2001 From: Mahmoud Date: Fri, 30 Mar 2018 11:27:57 -0400 Subject: fixing gpu-tot-cycle bug --- src/gpgpu-sim/gpu-sim.cc | 1 - 1 file changed, 1 deletion(-) (limited to 'src') diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index 006d243..7835496 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -844,7 +844,6 @@ void gpgpu_sim::update_stats() { partiton_replys_in_parallel = 0; partiton_reqs_in_parallel_util = 0; gpu_sim_cycle_parition_util = 0; - gpu_tot_sim_cycle = 0; gpu_sim_insn = 0; m_total_cta_launched = 0; } -- cgit v1.3 From 6aba79083ec47a78ad2a9572a03cd52f42141def Mon Sep 17 00:00:00 2001 From: tgrogers Date: Fri, 30 Mar 2018 16:10:27 -0400 Subject: When CDP was added - we started printing garbage stats after every memcpy. I am stopping this by gaurding against it in the same way we do for updating the stat counts (checking if we have executed any cycles before printing). I am not sure what this does to CDP stat prints - as I am not sure what CDP kernels so to sim_cycles.... However EVERY other app except CDP is spewing junk to the output and it really messes up any per-kernel stat collection scripts. --- src/gpgpusim_entrypoint.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'src') diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index ad4587a..dac9fd8 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -156,8 +156,8 @@ void *gpgpu_sim_thread_concurrent(void*) printf("GPGPU-Sim: ** STOP simulation thread (no work) **\n"); fflush(stdout); } - g_the_gpu->print_stats(); if(sim_cycles) { + g_the_gpu->print_stats(); g_the_gpu->update_stats(); print_simulation_time(); } -- cgit v1.3 From 9b586f9a61cac16b147e6dfea62abecc0a2a8eb7 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Fri, 30 Mar 2018 16:20:36 -0400 Subject: Moving the print to avoid spurrious statistics --- src/gpgpusim_entrypoint.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'src') diff --git a/src/gpgpusim_entrypoint.cc b/src/gpgpusim_entrypoint.cc index ad4587a..dac9fd8 100644 --- a/src/gpgpusim_entrypoint.cc +++ b/src/gpgpusim_entrypoint.cc @@ -156,8 +156,8 @@ void *gpgpu_sim_thread_concurrent(void*) printf("GPGPU-Sim: ** STOP simulation thread (no work) **\n"); fflush(stdout); } - g_the_gpu->print_stats(); if(sim_cycles) { + g_the_gpu->print_stats(); g_the_gpu->update_stats(); print_simulation_time(); } -- cgit v1.3 From 89db73061e043c26df22c7f18d9adb106d8078ac Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sat, 31 Mar 2018 17:56:50 -0400 Subject: Getting rid of our constant, annoying prints. Running workloads of any size causes ridiculous output file sizes --- src/cuda-sim/cuda-sim.cc | 2 +- src/gpgpu-sim/gpu-sim.cc | 12 +++++++----- src/gpgpu-sim/shader.cc | 14 ++++++++------ src/trace_streams.tup | 1 + 4 files changed, 17 insertions(+), 12 deletions(-) (limited to 'src') diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index d4ace76..a34b99b 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -1408,7 +1408,7 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id) if ( (g_ptx_sim_num_insn % 100000) == 0 ) { dim3 ctaid = get_ctaid(); dim3 tid = get_tid(); - printf("GPGPU-Sim PTX: %u instructions simulated : ctaid=(%u,%u,%u) tid=(%u,%u,%u)\n", + DPRINTF(LIVENESS, "GPGPU-Sim PTX: %u instructions simulated : ctaid=(%u,%u,%u) tid=(%u,%u,%u)\n", g_ptx_sim_num_insn, ctaid.x,ctaid.y,ctaid.z,tid.x,tid.y,tid.z ); fflush(stdout); } diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc index d452888..0ef267d 100644 --- a/src/gpgpu-sim/gpu-sim.cc +++ b/src/gpgpu-sim/gpu-sim.cc @@ -36,6 +36,7 @@ #include "shader.h" +#include "shader_trace.h" #include "dram.h" #include "mem_fetch.h" @@ -1175,8 +1176,8 @@ bool shader_core_ctx::occupy_shader_resource_1block(kernel_info_t & k, bool occu m_occupied_regs += (padded_cta_size * ((kernel_info->regs+3)&~3)); m_occupied_ctas++; - printf("GPGPU-Sim uArch: Shader %d occupied %d threads, %d shared mem, %d registers, %d ctas\n", - m_sid, m_occupied_n_threads, m_occupied_shmem, m_occupied_regs, m_occupied_ctas); + SHADER_DPRINTF(LIVENESS, "GPGPU-Sim uArch: Occupied %d threads, %d shared mem, %d registers, %d ctas\n", + m_occupied_n_threads, m_occupied_shmem, m_occupied_regs, m_occupied_ctas); } return true; @@ -1301,8 +1302,8 @@ void shader_core_ctx::issue_block2core( kernel_info_t &kernel ) m_n_active_cta++; shader_CTA_count_log(m_sid, 1); - printf("GPGPU-Sim uArch: core:%3d, cta:%2u, start_tid:%4u, end_tid:%4u, initialized @(%lld,%lld)\n", - m_sid, free_cta_hw_id, start_thread, end_thread, gpu_sim_cycle, gpu_tot_sim_cycle ); + SHADER_DPRINTF(LIVENESS, "GPGPU-Sim uArch: cta:%2u, start_tid:%4u, end_tid:%4u, initialized @(%lld,%lld)\n", + free_cta_hw_id, start_thread, end_thread, gpu_sim_cycle, gpu_tot_sim_cycle ); } @@ -1499,7 +1500,8 @@ void gpgpu_sim::cycle() hrs = elapsed_time/3600 - 24*days; minutes = elapsed_time/60 - 60*(hrs + 24*days); sec = elapsed_time - 60*(minutes + 60*(hrs + 24*days)); - printf("GPGPU-Sim uArch: cycles simulated: %lld inst.: %lld (ipc=%4.1f) sim_rate=%u (inst/sec) elapsed = %u:%u:%02u:%02u / %s", + + DPRINTF(LIVENESS, "GPGPU-Sim uArch: cycles simulated: %lld inst.: %lld (ipc=%4.1f) sim_rate=%u (inst/sec) elapsed = %u:%u:%02u:%02u / %s", gpu_tot_sim_cycle + gpu_sim_cycle, gpu_tot_sim_insn + gpu_sim_insn, (double)gpu_sim_insn/(double)gpu_sim_cycle, (unsigned)((gpu_tot_sim_insn+gpu_sim_insn) / elapsed_time), diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc index d17e51d..4640d65 100644 --- a/src/gpgpu-sim/shader.cc +++ b/src/gpgpu-sim/shader.cc @@ -1969,12 +1969,12 @@ void shader_core_ctx::register_cta_thread_exit( unsigned cta_num, kernel_info_t m_barriers.deallocate_barrier(cta_num); shader_CTA_count_unlog(m_sid, 1); - printf("GPGPU-Sim uArch: Shader %d finished CTA #%d (%lld,%lld), %u CTAs running\n", m_sid, cta_num, gpu_sim_cycle, gpu_tot_sim_cycle, - m_n_active_cta ); + SHADER_DPRINTF(LIVENESS, "GPGPU-Sim uArch: Finished CTA #%d (%lld,%lld), %u CTAs running\n", + cta_num, gpu_sim_cycle, gpu_tot_sim_cycle, m_n_active_cta); if( m_n_active_cta == 0 ) { - printf("GPGPU-Sim uArch: Shader %u empty (last released kernel %u \'%s\').\n", m_sid, kernel->get_uid(), - kernel->name().c_str() ); + SHADER_DPRINTF(LIVENESS, "GPGPU-Sim uArch: Empty (last released kernel %u \'%s\').\n", + kernel->get_uid(), kernel->name().c_str()); fflush(stdout); //Shader can only be empty when no more cta are dispatched @@ -1989,8 +1989,10 @@ void shader_core_ctx::register_cta_thread_exit( unsigned cta_num, kernel_info_t kernel->dec_running(); if( !m_gpu->kernel_more_cta_left(kernel) ) { if( !kernel->running() ) { - printf("GPGPU-Sim uArch: GPU detected kernel %u \'%s\' finished on shader %u.\n", kernel->get_uid(), - kernel->name().c_str(), m_sid ); + SHADER_DPRINTF(LIVENESS, + "GPGPU-Sim uArch: GPU detected kernel %u \'%s\' finished on shader %u.\n", kernel->get_uid(), + kernel->name().c_str(), m_sid); + if(m_kernel == kernel) m_kernel = NULL; m_gpu->set_kernel_done( kernel ); diff --git a/src/trace_streams.tup b/src/trace_streams.tup index c41690e..fd68a16 100644 --- a/src/trace_streams.tup +++ b/src/trace_streams.tup @@ -29,5 +29,6 @@ TS_TUP_BEGIN( trace_streams_type ) TS_TUP( WARP_SCHEDULER ), TS_TUP( SCOREBOARD ), TS_TUP( MEMORY_PARTITION_UNIT ), + TS_TUP( LIVENESS ), TS_TUP( NUM_TRACE_STREAMS ) TS_TUP_END( trace_streams_type ) -- cgit v1.3