From 78fd9e7f9d9934036b2f43392258b559886f3314 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sat, 24 Mar 2018 20:21:13 -0400 Subject: Adding in a patch to get lonestargpu to work --- libcuda/cuda_runtime_api.cc | 24 ++++++++++++++++++++---- 1 file changed, 20 insertions(+), 4 deletions(-) (limited to 'libcuda/cuda_runtime_api.cc') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index cbe8a11..d8e7832 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -336,10 +336,23 @@ class _cuda_device_id *GPGPUSim_Init() prop->minor = 2; prop->totalGlobalMem = 0x80000000 /* 2 GB */; prop->memPitch = 0; - prop->maxThreadsPerBlock = 512; - prop->maxThreadsDim[0] = 512; - prop->maxThreadsDim[1] = 512; - prop->maxThreadsDim[2] = 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->maxGridSize[0] = 0x40000000; prop->maxGridSize[1] = 0x40000000; prop->maxGridSize[2] = 0x40000000; @@ -351,6 +364,9 @@ class _cuda_device_id *GPGPUSim_Init() prop->clockRate = the_gpu->shader_clock(); #if (CUDART_VERSION >= 2010) prop->multiProcessorCount = the_gpu->get_config().num_shader(); +#endif +#if (CUDART_VERSION >= 4000) + prop->maxThreadsPerMultiProcessor = the_gpu->threads_per_core(); #endif the_gpu->set_prop(prop); the_device = new _cuda_device_id(the_gpu); -- 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 'libcuda/cuda_runtime_api.cc') 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 db4bd85339caee6a6bd470da1c5f24bc4e8cbe45 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sun, 25 Mar 2018 15:50:13 -0400 Subject: succeed if just remining the GPU that pinned memory can beused --- libcuda/cuda_runtime_api.cc | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-) (limited to 'libcuda/cuda_runtime_api.cc') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 4f5e0fb..7ee7dbe 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -2104,8 +2104,14 @@ cudaError_t CUDARTAPI cudaSetValidDevices(int *device_arr, int len) cudaError_t CUDARTAPI cudaSetDeviceFlags( int flags ) { - cuda_not_implemented(__my_func__,__LINE__); - return g_last_cudaError = cudaErrorUnknown; + // This flag is implicitly always on (unless you are using the driver API). It is safe for GPGPU-Sim to + // just ignore it. + if ( cudaDeviceMapHost == flags ) { + return g_last_cudaError = cudaSuccess; + } else { + cuda_not_implemented(__my_func__,__LINE__); + return g_last_cudaError = cudaErrorUnknown; + } } size_t getMaxThreadsPerBlock(struct cudaFuncAttributes *attr) { -- cgit v1.3 From 635366fe3e9b596318647b9c5bcdd546c522d52a Mon Sep 17 00:00:00 2001 From: Tor Aamodt Date: Wed, 28 Mar 2018 14:37:43 -0700 Subject: fix compile errors on Ubuntu LTS 16.04 --- libcuda/cuda_runtime_api.cc | 18 +++++++++++------- src/abstract_hardware_model.h | 2 +- src/cuda-sim/cuda-math.h | 2 +- src/cuda-sim/instructions.cc | 8 ++++---- 4 files changed, 17 insertions(+), 13 deletions(-) (limited to 'libcuda/cuda_runtime_api.cc') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index cbe8a11..b7f25bf 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1406,7 +1406,7 @@ void extract_code_using_cuobjdump(){ cmd << "ldd " << app_binary << " | grep $CUDA_INSTALL_PATH | awk \'{print $3}\' > _tempfile_.txt"; int result = system(cmd.str().c_str()); if(result){ - std::cout << "Failed to execute: " << cmd << std::endl; + std::cout << "Failed to execute: " << cmd.str() << std::endl; exit(1); } std::ifstream libsf; @@ -1438,7 +1438,7 @@ void extract_code_using_cuobjdump(){ if(result) {printf("ERROR: Failed to execute: %s\n", command); exit(1);} std::cout << "Done" << std::endl; - std::cout << "Trying to parse " << libcodfn << std::endl; + std::cout << "Trying to parse " << libcodfn.str() << std::endl; cuobjdump_in = fopen(libcodfn.str().c_str(), "r"); cuobjdump_parse(); fclose(cuobjdump_in); @@ -1540,7 +1540,7 @@ std::list pruneSectionList(std::list cuobj //! Merge all PTX sections that have a specific identifier into one file std::list mergeMatchingSections(std::list cuobjdumpSectionList, std::string identifier){ - char *ptxcode = ""; + const char *ptxcode = ""; std::list::iterator old_iter; cuobjdumpPTXSection* old_ptxsection = NULL; cuobjdumpPTXSection* ptxsection; @@ -1689,7 +1689,7 @@ std::mapfatbin_registered; std::map name_symtab; //! Keep track of the association between filename and cubin handle -void cuobjdumpRegisterFatBinary(unsigned int handle, char* filename){ +void cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename){ fatbinmap[handle] = filename; } @@ -1764,6 +1764,7 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) if (sizeof(void*) == 4) printf("GPGPU-Sim PTX: FatBin file name extraction has not been tested on 32-bit system.\n"); + #if (CUDART_VERSION <= 6000) // FatBin handle from the .fatbin.c file (one of the intermediate files generated by NVCC) typedef struct {int m; int v; const unsigned long long* d; char* f;} __fatDeviceText __attribute__ ((aligned (8))); __fatDeviceText * fatDeviceText = (__fatDeviceText *) fatCubin; @@ -1772,12 +1773,11 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) // - Obtains the pointer to the actual fatbin structure from the FatBin handle (fatCubin). // - An integer inside the fatbin structure contains the relative offset to the source code file name. // - This offset differs among different CUDA and GCC versions. - #if (CUDART_VERSION <= 6000) char * pfatbin = (char*) fatDeviceText->d; int offset = *((int*)(pfatbin+48)); char * filename = (pfatbin+16+offset); #else - char * filename = "default"; + const char * filename = "default"; #endif // The extracted file name is associated with a fat_cubin_handle passed // into cudaLaunch(). Inside cudaLaunch(), the associated file name is @@ -1798,7 +1798,7 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) return (void**)fat_cubin_handle; } - #if (CUDART_VERSION < 8000) +#if (CUDART_VERSION < 8000) else { static unsigned source_num=1; unsigned long long fat_cubin_handle = next_fat_bin_handle++; @@ -1857,6 +1857,10 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) return (void**)fat_cubin_handle; } #endif + else { + printf("ERROR ** __cudaRegisterFatBinary() needs to be updated\n"); + abort(); + } } void __cudaUnregisterFatBinary(void **fatCubinHandle) diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h index aaa4b00..7125b6b 100644 --- a/src/abstract_hardware_model.h +++ b/src/abstract_hardware_model.h @@ -383,7 +383,7 @@ protected: std::deque m_stack; }; -#define GLOBAL_HEAP_START 0x703E20000 +#define GLOBAL_HEAP_START 0xC0000000 // start allocating from this address (lower values used for allocating globals in .ptx file) #define SHARED_MEM_SIZE_MAX (64*1024) #define LOCAL_MEM_SIZE_MAX (8*1024) diff --git a/src/cuda-sim/cuda-math.h b/src/cuda-sim/cuda-math.h index 4721e8a..a3db0df 100644 --- a/src/cuda-sim/cuda-math.h +++ b/src/cuda-sim/cuda-math.h @@ -321,7 +321,7 @@ float __internal_accurate_fdividef(float a, float b) float __saturatef(float a) { float b; - if (isnan(a)) b = 0.0f; + if (std::isnan(a)) b = 0.0f; else if (a >= 1.0f) b = 1.0f; else if (a <= 0.0f) b = 0.0f; else b = a; diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 011c285..71286c9 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -1961,7 +1961,7 @@ ptx_reg_t d2d( ptx_reg_t x, unsigned from_width, unsigned to_width, int to_sign, y.f64 = x.f64; break; } - if (isnan(y.f64)) { + if (std::isnan(y.f64)) { y.u64 = 0xfff8000000000000ull; } else if (saturation_mode) { y.f64 = cuda_math::__saturatef(y.f64); @@ -2086,7 +2086,7 @@ void ptx_round(ptx_reg_t& data, int rounding_mode, int type) } } if ((type == F64_TYPE)||(type == FF64_TYPE)) { - if (isnan(data.f64)) { + if (std::isnan(data.f64)) { data.u64 = 0xfff8000000000000ull; } } @@ -2648,12 +2648,12 @@ void mad_def( const ptx_instruction *pI, ptx_thread_info *thread, bool use_carry bool isNaN(float x) { - return isnan(x); + return std::isnan(x); } bool isNaN(double x) { - return isnan(x); + return std::isnan(x); } void max_impl( const ptx_instruction *pI, ptx_thread_info *thread ) -- cgit v1.3 From c9cc4281bf84ad6cff77d20389b59d14a534ad6b Mon Sep 17 00:00:00 2001 From: Tor Aamodt Date: Wed, 28 Mar 2018 16:16:38 -0700 Subject: (attempt to) fix compile error in last checkin --- libcuda/cuda_runtime_api.cc | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) (limited to 'libcuda/cuda_runtime_api.cc') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index b7f25bf..9bdb993 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1856,11 +1856,12 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) } return (void**)fat_cubin_handle; } - #endif +#else else { printf("ERROR ** __cudaRegisterFatBinary() needs to be updated\n"); abort(); } +#endif } void __cudaUnregisterFatBinary(void **fatCubinHandle) -- cgit v1.3 From 373d64290239f3ed74d98b20494383f03fe189b6 Mon Sep 17 00:00:00 2001 From: Nathan Conrad Date: Sun, 22 Apr 2018 23:31:42 -0400 Subject: Some classes were referred to as a class and a struct (reported as clang warnings). This makes these consistent. --- libcuda/cuda_runtime_api.cc | 2 +- src/cuda-sim/cuda-sim.cc | 2 +- src/cuda-sim/ptx_ir.h | 2 +- src/debug.cc | 2 +- src/debug.h | 2 +- src/gpgpu-sim/dram.h | 4 ++-- src/gpgpu-sim/mem_fetch.cc | 2 +- src/gpgpu-sim/mem_fetch.h | 4 ++-- src/gpgpu-sim/shader.h | 4 ++-- src/stream_manager.h | 6 +++--- 10 files changed, 15 insertions(+), 15 deletions(-) (limited to 'libcuda/cuda_runtime_api.cc') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 9bdb993..5ef6115 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -324,7 +324,7 @@ private: gpgpu_ptx_sim_arg_list_t m_args; }; -class _cuda_device_id *GPGPUSim_Init() +struct _cuda_device_id *GPGPUSim_Init() { static _cuda_device_id *the_device = NULL; if( !the_device ) { diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index a34b99b..9f24c69 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -2067,7 +2067,7 @@ struct rec_pts { int s_num_recon; }; -struct std::map g_rpts; +class std::map g_rpts; struct rec_pts find_reconvergence_points( function_info *finfo ) { diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 9ad1571..8750187 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -105,7 +105,7 @@ private: int m_is_function; bool m_is_non_arch_reg; - friend class type_info_key_compare; + friend struct type_info_key_compare; }; class symbol_table; diff --git a/src/debug.cc b/src/debug.cc index cfd7bb0..ae15760 100644 --- a/src/debug.cc +++ b/src/debug.cc @@ -222,7 +222,7 @@ void gpgpu_sim::gpgpu_debug() } } -bool thread_at_brkpt( ptx_thread_info *thread, const struct brk_pt &b ) +bool thread_at_brkpt( ptx_thread_info *thread, const class brk_pt &b ) { return b.is_equal(thread->get_location(),thread->get_uid()); } diff --git a/src/debug.h b/src/debug.h index 7c79f1e..1277494 100644 --- a/src/debug.h +++ b/src/debug.h @@ -87,7 +87,7 @@ extern int gpgpu_ptx_instruction_classification ; class ptx_thread_info; class ptx_instruction; -bool thread_at_brkpt( ptx_thread_info *thd_info, const struct brk_pt &b ); +bool thread_at_brkpt( ptx_thread_info *thd_info, const class brk_pt &b ); void hit_watchpoint( unsigned watchpoint_num, ptx_thread_info *thd, const ptx_instruction *pI ); #endif diff --git a/src/gpgpu-sim/dram.h b/src/gpgpu-sim/dram.h index a8bff14..15c63e7 100644 --- a/src/gpgpu-sim/dram.h +++ b/src/gpgpu-sim/dram.h @@ -87,7 +87,7 @@ struct bank_t unsigned int bkgrpindex; }; -struct mem_fetch; +class mem_fetch; class dram_t { @@ -178,7 +178,7 @@ private: unsigned int ave_mrqs_partial; unsigned int bwutil_partial; - struct memory_stats_t *m_stats; + class memory_stats_t *m_stats; class Stats* mrqq_Dist; //memory request queue inside DRAM friend class frfcfs_scheduler; diff --git a/src/gpgpu-sim/mem_fetch.cc b/src/gpgpu-sim/mem_fetch.cc index 580c051..729636d 100644 --- a/src/gpgpu-sim/mem_fetch.cc +++ b/src/gpgpu-sim/mem_fetch.cc @@ -39,7 +39,7 @@ mem_fetch::mem_fetch( const mem_access_t &access, unsigned wid, unsigned sid, unsigned tpc, - const class memory_config *config ) + const struct memory_config *config ) { m_request_uid = sm_next_mf_request_uid++; m_access = access; diff --git a/src/gpgpu-sim/mem_fetch.h b/src/gpgpu-sim/mem_fetch.h index c89edbb..de98748 100644 --- a/src/gpgpu-sim/mem_fetch.h +++ b/src/gpgpu-sim/mem_fetch.h @@ -55,7 +55,7 @@ public: unsigned wid, unsigned sid, unsigned tpc, - const class memory_config *config ); + const struct memory_config *config ); ~mem_fetch(); void set_status( enum mem_fetch_status status, unsigned long long cycle ); @@ -141,7 +141,7 @@ private: static unsigned sm_next_mf_request_uid; - const class memory_config *m_mem_config; + const struct memory_config *m_mem_config; unsigned icnt_flit_size; }; diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h index bdd8dbe..ea8c019 100644 --- a/src/gpgpu-sim/shader.h +++ b/src/gpgpu-sim/shader.h @@ -286,7 +286,7 @@ typedef std::bitset warp_set_t; int register_bank(int regnum, int wid, unsigned num_banks, unsigned bank_warp_shift); class shader_core_ctx; -class shader_core_config; +struct shader_core_config; class shader_core_stats; enum scheduler_prioritization_type @@ -967,7 +967,7 @@ struct ifetch_buffer_t { unsigned m_warp_id; }; -class shader_core_config; +struct shader_core_config; class simd_function_unit { public: diff --git a/src/stream_manager.h b/src/stream_manager.h index 222a1b2..d3a804f 100644 --- a/src/stream_manager.h +++ b/src/stream_manager.h @@ -93,7 +93,7 @@ public: m_stream=stream; m_done=false; } - stream_operation( class CUevent_st *e, struct CUstream_st *stream ) + stream_operation( struct CUevent_st *e, struct CUstream_st *stream ) { m_kernel=NULL; m_type=stream_event; @@ -172,10 +172,10 @@ private: bool m_sim_mode; kernel_info_t *m_kernel; - class CUevent_st *m_event; + struct CUevent_st *m_event; }; -class CUevent_st { +struct CUevent_st { public: CUevent_st( bool blocking ) { -- cgit v1.3 From b71760a74c67039a8fc61d1d1418c03a0608ca5d Mon Sep 17 00:00:00 2001 From: tgrogers Date: Fri, 29 Jun 2018 10:16:33 -0400 Subject: Adding a print to indicate which PTX version we are actually using --- libcuda/cuda_runtime_api.cc | 1 + 1 file changed, 1 insertion(+) (limited to 'libcuda/cuda_runtime_api.cc') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 9bdb993..6941b26 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1714,6 +1714,7 @@ void cuobjdumpParseBinary(unsigned int handle){ unsigned capability = (*iter)->getArch(); if (capability > max_capability) max_capability = capability; } + printf("Using PTX version = %u\n", max_capability); if (max_capability > 20) printf("WARNING: No guarantee that PTX will be parsed for SM version %u\n", max_capability); cuobjdumpPTXSection* ptx = findPTXSection(fname); -- cgit v1.3 From 8334828738e3dec7221bd411d3c541bd940fde59 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Fri, 29 Jun 2018 10:18:11 -0400 Subject: Whitespace edits - removing tabs - replacing with spaces --- libcuda/cuda_runtime_api.cc | 106 ++++++++++++++++++++++---------------------- 1 file changed, 53 insertions(+), 53 deletions(-) (limited to 'libcuda/cuda_runtime_api.cc') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 6941b26..2b0f944 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1696,60 +1696,60 @@ void cuobjdumpRegisterFatBinary(unsigned int handle, const char* filename){ //! Either submit PTX for simulation or convert SASS to PTXPlus and submit it void cuobjdumpParseBinary(unsigned int handle){ - if(fatbin_registered[handle]) return; - fatbin_registered[handle] = true; - CUctx_st *context = GPGPUSim_Context(); - std::string fname = fatbinmap[handle]; - - if (name_symtab.find(fname) != name_symtab.end()) { - symbol_table *symtab = name_symtab[fname]; - context->add_binary(symtab, handle); - return; - } - - unsigned max_capability = 0; - for ( std::list::iterator iter = cuobjdumpSectionList.begin(); - iter != cuobjdumpSectionList.end(); - iter++){ - unsigned capability = (*iter)->getArch(); - if (capability > max_capability) max_capability = capability; - } + if(fatbin_registered[handle]) return; + fatbin_registered[handle] = true; + CUctx_st *context = GPGPUSim_Context(); + std::string fname = fatbinmap[handle]; + + if (name_symtab.find(fname) != name_symtab.end()) { + symbol_table *symtab = name_symtab[fname]; + context->add_binary(symtab, handle); + return; + } + + unsigned max_capability = 0; + for ( std::list::iterator iter = cuobjdumpSectionList.begin(); + iter != cuobjdumpSectionList.end(); + iter++ ){ + unsigned capability = (*iter)->getArch(); + if (capability > max_capability) max_capability = capability; + } printf("Using PTX version = %u\n", max_capability); - if (max_capability > 20) printf("WARNING: No guarantee that PTX will be parsed for SM version %u\n", max_capability); - - cuobjdumpPTXSection* ptx = findPTXSection(fname); - symbol_table *symtab; - char *ptxcode; - const char *override_ptx_name = getenv("PTX_SIM_KERNELFILE"); - if (override_ptx_name == NULL or getenv("PTX_SIM_USE_PTX_FILE") == NULL) { - ptxcode = readfile(ptx->getPTXfilename()); - } else { - printf("GPGPU-Sim PTX: overriding embedded ptx with '%s' (PTX_SIM_USE_PTX_FILE is set)\n", override_ptx_name); - ptxcode = readfile(override_ptx_name); - } - if(context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() ) { - cuobjdumpELFSection* elfsection = findELFSection(ptx->getIdentifier()); - assert (elfsection!= NULL); - char *ptxplus_str = gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus( - ptx->getPTXfilename(), - elfsection->getELFfilename(), - elfsection->getSASSfilename()); - symtab=gpgpu_ptx_sim_load_ptx_from_string(ptxplus_str, handle); - printf("Adding %s with cubin handle %u\n", ptx->getPTXfilename().c_str(), handle); - context->add_binary(symtab, handle); - gpgpu_ptxinfo_load_from_string( ptxcode, handle, max_capability ); - delete[] ptxplus_str; - } else { - symtab=gpgpu_ptx_sim_load_ptx_from_string(ptxcode, handle); - printf("Adding %s with cubin handle %u\n", ptx->getPTXfilename().c_str(), handle); - context->add_binary(symtab, handle); - gpgpu_ptxinfo_load_from_string( ptxcode, handle, max_capability ); - } - load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); - load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); - name_symtab[fname] = symtab; - - //TODO: Remove temporarily files as per configurations + if (max_capability > 20) printf("WARNING: No guarantee that PTX will be parsed for SM version %u\n", max_capability); + + cuobjdumpPTXSection* ptx = findPTXSection(fname); + symbol_table *symtab; + char *ptxcode; + const char *override_ptx_name = getenv("PTX_SIM_KERNELFILE"); + if (override_ptx_name == NULL or getenv("PTX_SIM_USE_PTX_FILE") == NULL) { + ptxcode = readfile(ptx->getPTXfilename()); + } else { + printf("GPGPU-Sim PTX: overriding embedded ptx with '%s' (PTX_SIM_USE_PTX_FILE is set)\n", override_ptx_name); + ptxcode = readfile(override_ptx_name); + } + if(context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() ) { + cuobjdumpELFSection* elfsection = findELFSection(ptx->getIdentifier()); + assert (elfsection!= NULL); + char *ptxplus_str = gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus( + ptx->getPTXfilename(), + elfsection->getELFfilename(), + elfsection->getSASSfilename()); + symtab=gpgpu_ptx_sim_load_ptx_from_string(ptxplus_str, handle); + printf("Adding %s with cubin handle %u\n", ptx->getPTXfilename().c_str(), handle); + context->add_binary(symtab, handle); + gpgpu_ptxinfo_load_from_string( ptxcode, handle, max_capability ); + delete[] ptxplus_str; + } else { + symtab=gpgpu_ptx_sim_load_ptx_from_string(ptxcode, handle); + printf("Adding %s with cubin handle %u\n", ptx->getPTXfilename().c_str(), handle); + context->add_binary(symtab, handle); + gpgpu_ptxinfo_load_from_string( ptxcode, handle, max_capability ); + } + load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); + load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); + name_symtab[fname] = symtab; + + //TODO: Remove temporarily files as per configurations } void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) -- cgit v1.3 From 112a3a2e3c9d17af420d2389f68a35407c692744 Mon Sep 17 00:00:00 2001 From: tgrogers Date: Fri, 29 Jun 2018 17:34:22 -0400 Subject: Allowing modern GPU configrations to properly run PTXPLUS. There are several problems with this: 1) To get proper occupancy, based on register file usage, you must querry ptxas with the sm version that corresponds to the register usage in the config file. To enable this, a new config option has been added that determines what sm version you pass to ptxas to compute reg-usage. This configuration option is always required in the gpgpusim.config file 2) If you are running PTXPLUS with a modern card (i.e. volta/pascal), you need ptxas from CUDA 9.1. However since PTXPLUS only supports sm_13 - you need a version of CUDA where cudaobjectdump supports sm_1x. This ended at CUDA 5 - and PTXPLUS requires CUDA 4.2. Therefore, to run PTXPLUS on a modern card, you need CUDA 4.2 + modern CUDA installed. To fascilitate this, a new envronment varaible is added and the setup envrionment script prints an appropraite warning if you are using a newer CUDA. We have tried to make this as fail-proof as possible - and die appropraitely when something is wrong. --- configs/GTX480/gpgpusim.config | 1 + configs/GeForceGTX1080Ti/gpgpusim.config | 3 ++- configs/GeForceGTX750Ti/gpgpusim.config | 1 + configs/QuadroFX5600/gpgpusim.config | 3 +++ configs/QuadroFX5800/gpgpusim.config | 2 ++ configs/TeslaC2050/gpgpusim.config | 1 + libcuda/cuda_runtime_api.cc | 43 +++++++++++++++++++++----------- libopencl/opencl_runtime_api.cc | 2 +- setup_environment | 13 ++++++++++ src/cuda-sim/ptx_loader.cc | 18 ++++++++++--- src/cuda-sim/ptx_loader.h | 2 +- 11 files changed, 67 insertions(+), 22 deletions(-) (limited to 'libcuda/cuda_runtime_api.cc') diff --git a/configs/GTX480/gpgpusim.config b/configs/GTX480/gpgpusim.config index 436cb41..ee90c12 100644 --- a/configs/GTX480/gpgpusim.config +++ b/configs/GTX480/gpgpusim.config @@ -22,6 +22,7 @@ # shader core pipeline config -gpgpu_shader_registers 32768 +-gpgpu_occupancy_sm_number 20 # This implies a maximum of 48 warps/SM -gpgpu_shader_core_pipeline 1536:32 diff --git a/configs/GeForceGTX1080Ti/gpgpusim.config b/configs/GeForceGTX1080Ti/gpgpusim.config index 47c2b6a..fb044c6 100644 --- a/configs/GeForceGTX1080Ti/gpgpusim.config +++ b/configs/GeForceGTX1080Ti/gpgpusim.config @@ -3,7 +3,7 @@ # functional simulator specification -gpgpu_ptx_instruction_classification 0 -gpgpu_ptx_sim_mode 0 --gpgpu_ptx_force_max_capability 20 +-gpgpu_ptx_force_max_capability 60 # SASS execution (only supported with CUDA >= 4.0) -gpgpu_ptx_convert_to_ptxplus 0 @@ -23,6 +23,7 @@ # shader core pipeline config -gpgpu_shader_registers 65536 +-gpgpu_occupancy_sm_number 60 # This implies a maximum of 64 warps/SM -gpgpu_shader_core_pipeline 2048:32 diff --git a/configs/GeForceGTX750Ti/gpgpusim.config b/configs/GeForceGTX750Ti/gpgpusim.config index 8b030b6..c675aab 100644 --- a/configs/GeForceGTX750Ti/gpgpusim.config +++ b/configs/GeForceGTX750Ti/gpgpusim.config @@ -21,6 +21,7 @@ # shader core pipeline config -gpgpu_shader_registers 65536 +-gpgpu_occupancy_sm_number 52 # This implies a maximum of 64 warps/SM -gpgpu_shader_core_pipeline 2048:32 diff --git a/configs/QuadroFX5600/gpgpusim.config b/configs/QuadroFX5600/gpgpusim.config index cb87b65..6f836ee 100644 --- a/configs/QuadroFX5600/gpgpusim.config +++ b/configs/QuadroFX5600/gpgpusim.config @@ -11,6 +11,9 @@ # shader core pipeline config -gpgpu_shader_registers 16384 +-gpgpu_occupancy_sm_number 12 + +-gpgpu_occupancy_sm_number 12 #8192 (registers per block as written by device Query and which used in this option in our other configurations but this break some benchmarks execution! it does not affect performance modeling though) -gpgpu_shader_core_pipeline 768:32 -gpgpu_shader_cta 8 diff --git a/configs/QuadroFX5800/gpgpusim.config b/configs/QuadroFX5800/gpgpusim.config index 82243c2..fef1110 100644 --- a/configs/QuadroFX5800/gpgpusim.config +++ b/configs/QuadroFX5800/gpgpusim.config @@ -11,6 +11,8 @@ # shader core pipeline config -gpgpu_shader_registers 16384 +-gpgpu_occupancy_sm_number 13 + -gpgpu_shader_core_pipeline 1024:32 -gpgpu_shader_cta 8 -gpgpu_simd_model 1 diff --git a/configs/TeslaC2050/gpgpusim.config b/configs/TeslaC2050/gpgpusim.config index 442ab8b..6ac2c12 100644 --- a/configs/TeslaC2050/gpgpusim.config +++ b/configs/TeslaC2050/gpgpusim.config @@ -25,6 +25,7 @@ # shader core pipeline config -gpgpu_shader_registers 32768 +-gpgpu_occupancy_sm_number 20 # This implies a maximum of 48 warps/SM -gpgpu_shader_core_pipeline 1536:32 diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 2b0f944..94705dc 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1350,39 +1350,52 @@ std::string get_app_binary(){ * enabled * */ void extract_code_using_cuobjdump(){ - CUctx_st *context = GPGPUSim_Context(); - char command[1000]; + CUctx_st *context = GPGPUSim_Context(); + std::string command; - std::string app_binary = get_app_binary(); + std::string app_binary = get_app_binary(); char fname[1024]; snprintf(fname,1024,"_cuobjdump_complete_output_XXXXXX"); int fd=mkstemp(fname); close(fd); // Running cuobjdump using dynamic link to current process - snprintf(command,1000,"md5sum %s ", app_binary.c_str()); - printf("Running md5sum using \"%s\"\n", command); - system(command); + command = "md5sum " + app_binary; + printf("Running md5sum using \"%s\"\n", command.c_str()); + system(command.c_str()); // Running cuobjdump using dynamic link to current process // Needs the option '-all' to extract PTX from CDP-enabled binary extern bool g_cdp_enabled; + const int current_cuda = atoi(getenv("CUDA_VERSION_NUMBER")); + if ( context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() && current_cuda > 42 ) { + const char* cuda_42_path = getenv("CUDA_42_INSTALL_PATH"); + if ( NULL == cuda_42_path ) { + printf("You are using a new version of CUDA and PTXPLUS. You are required to have CUDA SDK 4.2 and explicitly " + "point to it via the environment variable $CUDA_42_INSTALL_PATH."); + exit(1); + } else { + command = "$CUDA_42_INSTALL_PATH/bin/cuobjdump"; + } + } else { + command = "$CUDA_INSTALL_PATH/bin/cuobjdump"; + } if(!g_cdp_enabled) - snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass %s > %s", app_binary.c_str(), fname); + command += " -ptx -elf -sass " + app_binary + " > " + fname; else - snprintf(command,1000,"$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass -all %s > %s", app_binary.c_str(), fname); + command += " -ptx -elf -sass -all " + app_binary + " > " + fname; bool parse_output = true; - int result = system(command); + int result = system(command.c_str()); if(result) { if (context->get_device()->get_gpgpu()->get_config().experimental_lib_support() && (result == 65280)) { // Some CUDA application may exclusively use kernels provided by CUDA // libraries (e.g. CUBLAS). Skipping cuobjdump extraction from the // executable for this case. // 65280 is the return code from cuobjdump denoting the specific error (tested on CUDA 4.0/4.1/4.2) - printf("WARNING: Failed to execute: %s\n", command); + printf("WARNING: Failed to execute: %s\n", command.c_str()); printf(" Executable binary does not contain any GPU kernel.\n"); parse_output = false; } else { - printf("ERROR: Failed to execute: %s\n", command); + printf("ERROR: Failed to execute: %s\n", command.c_str()); exit(1); } } @@ -1435,7 +1448,7 @@ void extract_code_using_cuobjdump(){ std::cout << "Running cuobjdump on " << line << std::endl; std::cout << "Using command: " << cmd.str() << std::endl; result = system(cmd.str().c_str()); - if(result) {printf("ERROR: Failed to execute: %s\n", command); exit(1);} + if(result) {printf("ERROR: Failed to execute: %s\n", command.c_str()); exit(1);} std::cout << "Done" << std::endl; std::cout << "Trying to parse " << libcodfn.str() << std::endl; @@ -1737,13 +1750,13 @@ void cuobjdumpParseBinary(unsigned int handle){ symtab=gpgpu_ptx_sim_load_ptx_from_string(ptxplus_str, handle); printf("Adding %s with cubin handle %u\n", ptx->getPTXfilename().c_str(), handle); context->add_binary(symtab, handle); - gpgpu_ptxinfo_load_from_string( ptxcode, handle, max_capability ); + gpgpu_ptxinfo_load_from_string( ptxcode, handle ); delete[] ptxplus_str; } else { symtab=gpgpu_ptx_sim_load_ptx_from_string(ptxcode, handle); printf("Adding %s with cubin handle %u\n", ptx->getPTXfilename().c_str(), handle); context->add_binary(symtab, handle); - gpgpu_ptxinfo_load_from_string( ptxcode, handle, max_capability ); + gpgpu_ptxinfo_load_from_string( ptxcode, handle ); } load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); load_constants(symtab,STATIC_ALLOC_LIMIT,context->get_device()->get_gpgpu()); @@ -1847,7 +1860,7 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) } else { symtab=gpgpu_ptx_sim_load_ptx_from_string(ptx,source_num); context->add_binary(symtab,fat_cubin_handle); - gpgpu_ptxinfo_load_from_string( ptx, source_num, max_capability ); + gpgpu_ptxinfo_load_from_string( ptx, source_num ); } source_num++; load_static_globals(symtab,STATIC_ALLOC_LIMIT,0xFFFFFFFF,context->get_device()->get_gpgpu()); diff --git a/libopencl/opencl_runtime_api.cc b/libopencl/opencl_runtime_api.cc index 63f7c81..97a54d8 100644 --- a/libopencl/opencl_runtime_api.cc +++ b/libopencl/opencl_runtime_api.cc @@ -577,7 +577,7 @@ void _cl_program::Build(const char *options) } info.m_asm = tmp; info.m_symtab = gpgpu_ptx_sim_load_ptx_from_string( tmp, source_num ); - gpgpu_ptxinfo_load_from_string( tmp, source_num); + gpgpu_ptxinfo_load_from_string( tmp, source_num ); free(tmp); } printf("GPGPU-Sim OpenCL API: finished compiling OpenCL kernels.\n"); diff --git a/setup_environment b/setup_environment index 3f2476e..939fb85 100644 --- a/setup_environment +++ b/setup_environment @@ -53,6 +53,8 @@ if [ $CUDA_VERSION_NUMBER -gt 9100 -o $CUDA_VERSION_NUMBER -lt 2030 ]; then return fi + + if [ $# = '1' ] ; then export GPGPUSIM_CONFIG=gcc-$CC_VERSION/cuda-$CUDA_VERSION_NUMBER/$1 @@ -134,4 +136,15 @@ fi echo "setup_environment succeeded"; +if [ $CUDA_VERSION_NUMBER -gt 4200 ]; then + echo "" + echo "----------------------------------------------------------------------------" + echo "WARNING - If you only care about PTX execution, ignore this warning."; + echo "If you are using PTXPLUS, only sm_13 is supported and it requires that the app binaries are compiled in CUDA 4.2 or less."; + echo "New versions of CUDA tools have dropped parsing support for sm_1x - therefore to use PTXPLUS with a modern card configuration, you must"; + echo "have both CUDA 4.2 and a modern CUDA installed. When running PTXPLUS under a CUDA version >= 4.2, the simulator will fail unless \$CUDA_42_INSTALL_PATH"; + echo "has been set properly. You must set this manually."; + echo "----------------------------------------------------------------------------" +fi + export GPGPUSIM_SETUP_ENVIRONMENT_WAS_RUN=1 diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc index 6c1b595..99174cf 100644 --- a/src/cuda-sim/ptx_loader.cc +++ b/src/cuda-sim/ptx_loader.cc @@ -54,6 +54,7 @@ extern int ptxinfo_debug; extern FILE *ptxinfo_in; static bool g_save_embedded_ptx; +static int g_occupancy_sm_number; bool g_keep_intermediate_files; bool m_ptx_save_converted_ptxplus; @@ -71,6 +72,10 @@ void ptx_reg_options(option_parser_t opp) &m_ptx_save_converted_ptxplus, "Saved converted ptxplus to a file", "0"); + option_parser_register(opp, "-gpgpu_occupancy_sm_number", OPT_INT32, &g_occupancy_sm_number, + "The SM number to pass to ptxas when getting register usage for computing GPU occupancy." + "This parameter is required in the config.", + "0"); } void print_ptx_file( const char *p, unsigned source_num, const char *filename ) @@ -287,7 +292,7 @@ void fix_duplicate_errors(char fname2[1024]) { } } -void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version ) +void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num ) { char fname[1024]; snprintf(fname,1024,"_ptx_XXXXXX"); @@ -321,12 +326,17 @@ void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num extra_flags[0]=0; #if CUDART_VERSION >= 3000 - if (sm_version == 0) sm_version = 20; + if ( g_occupancy_sm_number == 0 ) { + printf( "gpgpusim.config must specify the sm version for the GPU that you use to compute occupancy.\n" + "The register file size is specifically tied to the sm version used to querry ptxas for register usage." + "A register size/SM mismatch may result in occupancy differences." ); + exit(1); + } extern bool g_cdp_enabled; if(!g_cdp_enabled) - snprintf(extra_flags,1024,"--gpu-name=sm_%u",sm_version); + snprintf(extra_flags,1024,"--gpu-name=sm_%u", g_occupancy_sm_number); else - snprintf(extra_flags,1024,"--compile-only --gpu-name=sm_%u",sm_version); + snprintf(extra_flags,1024,"--compile-only --gpu-name=sm_%u",g_occupancy_sm_number); #endif snprintf(commandline,1024,"$CUDA_INSTALL_PATH/bin/ptxas %s -v %s --output-file /dev/null 2> %s", diff --git a/src/cuda-sim/ptx_loader.h b/src/cuda-sim/ptx_loader.h index d3d0c92..cb02eda 100644 --- a/src/cuda-sim/ptx_loader.h +++ b/src/cuda-sim/ptx_loader.h @@ -32,7 +32,7 @@ extern bool g_override_embedded_ptx; class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num ); -void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num, unsigned sm_version=20 ); +void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num ); char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const std::string ptx_str, const std::string sass_str, const std::string elf_str); bool keep_intermediate_files(); -- cgit v1.3 From d42cb5648f77cd2caed9c656f4d6ad75c23b6f5a Mon Sep 17 00:00:00 2001 From: tgrogers Date: Fri, 29 Jun 2018 23:43:31 -0400 Subject: Fixing the PTXPLUS + new CUDA execution case. --- libcuda/cuda_runtime_api.cc | 61 ++++++++++++++++++++++++++++++++------------- setup_environment | 16 ++++++------ src/cuda-sim/ptx_loader.cc | 6 ++--- 3 files changed, 55 insertions(+), 28 deletions(-) (limited to 'libcuda/cuda_runtime_api.cc') diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 94705dc..5964a4d 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1367,17 +1367,17 @@ void extract_code_using_cuobjdump(){ // Needs the option '-all' to extract PTX from CDP-enabled binary extern bool g_cdp_enabled; const int current_cuda = atoi(getenv("CUDA_VERSION_NUMBER")); - if ( context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() && current_cuda > 42 ) { + if ( context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() && current_cuda > 4020 ) { const char* cuda_42_path = getenv("CUDA_42_INSTALL_PATH"); if ( NULL == cuda_42_path ) { printf("You are using a new version of CUDA and PTXPLUS. You are required to have CUDA SDK 4.2 and explicitly " - "point to it via the environment variable $CUDA_42_INSTALL_PATH."); + "point to it via the environment variable $CUDA_42_INSTALL_PATH.\n\n"); exit(1); } else { command = "$CUDA_42_INSTALL_PATH/bin/cuobjdump"; } } else { - command = "$CUDA_INSTALL_PATH/bin/cuobjdump"; + command = "$CUDA_42_INSTALL_PATH/bin/cuobjdump"; } if(!g_cdp_enabled) command += " -ptx -elf -sass " + app_binary + " > " + fname; @@ -1778,21 +1778,46 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) if (sizeof(void*) == 4) printf("GPGPU-Sim PTX: FatBin file name extraction has not been tested on 32-bit system.\n"); - #if (CUDART_VERSION <= 6000) - // FatBin handle from the .fatbin.c file (one of the intermediate files generated by NVCC) - typedef struct {int m; int v; const unsigned long long* d; char* f;} __fatDeviceText __attribute__ ((aligned (8))); - __fatDeviceText * fatDeviceText = (__fatDeviceText *) fatCubin; - - // Extract the source code file name that generate the given FatBin. - // - Obtains the pointer to the actual fatbin structure from the FatBin handle (fatCubin). - // - An integer inside the fatbin structure contains the relative offset to the source code file name. - // - This offset differs among different CUDA and GCC versions. - char * pfatbin = (char*) fatDeviceText->d; - int offset = *((int*)(pfatbin+48)); - char * filename = (pfatbin+16+offset); - #else - const char * filename = "default"; - #endif + // This code will get the CUDA version the app was compiled with. + // We need this to determine how to handle the parsing of the binary. + // Making this a runtime variable based on the app, enables GPGPU-Sim compiled + // with a newer version of CUDA to run apps compiled with older versions of + // CUDA. This is especially useful for PTXPLUS execution. + char fname[1024]; + snprintf(fname,1024,"_app_cuda_version_XXXXXX"); + int fd=mkstemp(fname); + close(fd); + int app_cuda_version = 0; + std::string app_cuda_version_command = "ldd " + get_app_binary() + " | grep libcudart.so | sed 's/.*libcudart.so.\\(.*\\) =>.*/\\1/' > " + fname; + system(app_cuda_version_command.c_str()); + FILE * cmd = fopen(fname, "r"); + char buf[256]; + while (fgets(buf, sizeof(buf), cmd) != 0) { + std::cout << buf; + app_cuda_version = atoi(buf); + } + fclose(cmd); + if ( app_cuda_version == 0 ) { + printf( "Error - Cannot detect the app's CUDA version.\n" ); + exit(1); + } + const char* filename; + if ( app_cuda_version < 6 ) { + // FatBin handle from the .fatbin.c file (one of the intermediate files generated by NVCC) + typedef struct {int m; int v; const unsigned long long* d; char* f;} __fatDeviceText __attribute__ ((aligned (8))); + __fatDeviceText * fatDeviceText = (__fatDeviceText *) fatCubin; + + // Extract the source code file name that generate the given FatBin. + // - Obtains the pointer to the actual fatbin structure from the FatBin handle (fatCubin). + // - An integer inside the fatbin structure contains the relative offset to the source code file name. + // - This offset differs among different CUDA and GCC versions. + char * pfatbin = (char*) fatDeviceText->d; + int offset = *((int*)(pfatbin+48)); + filename = (pfatbin+16+offset); + } else { + filename = "default"; + } + // The extracted file name is associated with a fat_cubin_handle passed // into cudaLaunch(). Inside cudaLaunch(), the associated file name is // used to find the PTX/SASS section from cuobjdump, which contains the diff --git a/setup_environment b/setup_environment index 939fb85..8bcb9aa 100644 --- a/setup_environment +++ b/setup_environment @@ -47,7 +47,7 @@ fi CC_VERSION=`gcc --version | head -1 | awk '{for(i=1;i<=NF;i++){ if(match($i,/^[0-9]\.[0-9]\.[0-9]$/)) {print $i; exit 0}}}'` CUDA_VERSION_STRING=`$CUDA_INSTALL_PATH/bin/nvcc --version | awk '/release/ {print $5;}' | sed 's/,//'`; -CUDA_VERSION_NUMBER=`echo $CUDA_VERSION_STRING | sed 's/\./ /' | awk '{printf("%02u%02u", 10*int($1), 10*$2);}'` +export CUDA_VERSION_NUMBER=`echo $CUDA_VERSION_STRING | sed 's/\./ /' | awk '{printf("%02u%02u", 10*int($1), 10*$2);}'` if [ $CUDA_VERSION_NUMBER -gt 9100 -o $CUDA_VERSION_NUMBER -lt 2030 ]; then echo "ERROR ** GPGPU-Sim version $GPGPUSIM_VERSION_STRING not tested with CUDA version $CUDA_VERSION_STRING (please see README)"; return @@ -134,17 +134,19 @@ else echo "configured without a power model."; fi -echo "setup_environment succeeded"; - if [ $CUDA_VERSION_NUMBER -gt 4200 ]; then - echo "" - echo "----------------------------------------------------------------------------" - echo "WARNING - If you only care about PTX execution, ignore this warning."; - echo "If you are using PTXPLUS, only sm_13 is supported and it requires that the app binaries are compiled in CUDA 4.2 or less."; + echo ""; + echo "----------------------------------------------------------------------------"; + echo "WARNING - If you only care about PTX execution, ignore this warning. GPGPU-Sim supports PTX execution in modern CUDA." + echo "If you want to run PTXPLUS, download SDK 4.2 (in addition to your modern CUDA) and set \$CUDA_42_INSTALL_PATH." + echo "The following text describes why:"; + echo "If you are using PTXPLUS, only sm_1x is supported and it requires that the app binaries are compiled in CUDA 4.2 or less."; echo "New versions of CUDA tools have dropped parsing support for sm_1x - therefore to use PTXPLUS with a modern card configuration, you must"; echo "have both CUDA 4.2 and a modern CUDA installed. When running PTXPLUS under a CUDA version >= 4.2, the simulator will fail unless \$CUDA_42_INSTALL_PATH"; echo "has been set properly. You must set this manually."; echo "----------------------------------------------------------------------------" fi +echo "setup_environment succeeded"; + export GPGPUSIM_SETUP_ENVIRONMENT_WAS_RUN=1 diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc index 99174cf..22f03b9 100644 --- a/src/cuda-sim/ptx_loader.cc +++ b/src/cuda-sim/ptx_loader.cc @@ -73,7 +73,7 @@ void ptx_reg_options(option_parser_t opp) "Saved converted ptxplus to a file", "0"); option_parser_register(opp, "-gpgpu_occupancy_sm_number", OPT_INT32, &g_occupancy_sm_number, - "The SM number to pass to ptxas when getting register usage for computing GPU occupancy." + "The SM number to pass to ptxas when getting register usage for computing GPU occupancy. " "This parameter is required in the config.", "0"); } @@ -327,8 +327,8 @@ void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num #if CUDART_VERSION >= 3000 if ( g_occupancy_sm_number == 0 ) { - printf( "gpgpusim.config must specify the sm version for the GPU that you use to compute occupancy.\n" - "The register file size is specifically tied to the sm version used to querry ptxas for register usage." + printf( "gpgpusim.config must specify the sm version for the GPU that you use to compute occupancy \"-gpgpu_occupancy_sm_number XX\".\n" + "The register file size is specifically tied to the sm version used to querry ptxas for register usage.\n" "A register size/SM mismatch may result in occupancy differences." ); exit(1); } -- cgit v1.3 From 691c57ea8596da35da3ce90a7b58796f3791036c Mon Sep 17 00:00:00 2001 From: tgrogers Date: Sat, 30 Jun 2018 10:12:20 -0400 Subject: changing the flow s.t. you cannot compile the simulator with a different version of CUDA from the app. --- diff.diff | 173 ++++++++++++++++++++++++++++++++++++++++++++ libcuda/cuda_runtime_api.cc | 69 +++++++++--------- setup_environment | 28 ++++--- src/cuda-sim/ptx_loader.cc | 2 +- 4 files changed, 225 insertions(+), 47 deletions(-) create mode 100644 diff.diff (limited to 'libcuda/cuda_runtime_api.cc') diff --git a/diff.diff b/diff.diff new file mode 100644 index 0000000..8ab27fe --- /dev/null +++ b/diff.diff @@ -0,0 +1,173 @@ +diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc +index 5964a4d..53436d3 100644 +--- a/libcuda/cuda_runtime_api.cc ++++ b/libcuda/cuda_runtime_api.cc +@@ -182,6 +182,8 @@ cudaError_t g_last_cudaError = cudaSuccess; + + extern stream_manager *g_stream_manager; + ++ ++ + void register_ptx_function( const char *name, function_info *impl ) + { + // no longer need this +@@ -1341,6 +1343,29 @@ std::string get_app_binary(){ + return self_exe_path; + } + ++static int get_app_cuda_version() { ++ int app_cuda_version = 0; ++ char fname[1024]; ++ snprintf(fname,1024,"_app_cuda_version_XXXXXX"); ++ int fd=mkstemp(fname); ++ close(fd); ++ std::string app_cuda_version_command = "ldd " + get_app_binary() + " | grep libcudart.so | sed 's/.*libcudart.so.\\(.*\\) =>.*/\\1/' > " + fname; ++ system(app_cuda_version_command.c_str()); ++ FILE * cmd = fopen(fname, "r"); ++ char buf[256]; ++ while (fgets(buf, sizeof(buf), cmd) != 0) { ++ std::cout << buf; ++ app_cuda_version = atoi(buf); ++ } ++ fclose(cmd); ++ if ( app_cuda_version == 0 ) { ++ printf( "Error - Cannot detect the app's CUDA version.\n" ); ++ exit(1); ++ } ++ return app_cuda_version; ++} ++ ++ + //! Call cuobjdump to extract everything (-elf -sass -ptx) + /*! + * This Function extract the whole PTX (for all the files) using cuobjdump +@@ -1366,23 +1391,10 @@ void extract_code_using_cuobjdump(){ + // Running cuobjdump using dynamic link to current process + // Needs the option '-all' to extract PTX from CDP-enabled binary + extern bool g_cdp_enabled; +- const int current_cuda = atoi(getenv("CUDA_VERSION_NUMBER")); +- if ( context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() && current_cuda > 4020 ) { +- const char* cuda_42_path = getenv("CUDA_42_INSTALL_PATH"); +- if ( NULL == cuda_42_path ) { +- printf("You are using a new version of CUDA and PTXPLUS. You are required to have CUDA SDK 4.2 and explicitly " +- "point to it via the environment variable $CUDA_42_INSTALL_PATH.\n\n"); +- exit(1); +- } else { +- command = "$CUDA_42_INSTALL_PATH/bin/cuobjdump"; +- } +- } else { +- command = "$CUDA_42_INSTALL_PATH/bin/cuobjdump"; +- } + if(!g_cdp_enabled) +- command += " -ptx -elf -sass " + app_binary + " > " + fname; ++ command = "$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass " + app_binary + " > " + fname; + else +- command += " -ptx -elf -sass -all " + app_binary + " > " + fname; ++ command = "$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass -all " + app_binary + " > " + fname; + bool parse_output = true; + int result = system(command.c_str()); + if(result) { +@@ -1765,6 +1777,7 @@ void cuobjdumpParseBinary(unsigned int handle){ + //TODO: Remove temporarily files as per configurations + } + ++ + void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) + { + #if (CUDART_VERSION < 2010) +@@ -1783,26 +1796,10 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) + // Making this a runtime variable based on the app, enables GPGPU-Sim compiled + // with a newer version of CUDA to run apps compiled with older versions of + // CUDA. This is especially useful for PTXPLUS execution. +- char fname[1024]; +- snprintf(fname,1024,"_app_cuda_version_XXXXXX"); +- int fd=mkstemp(fname); +- close(fd); +- int app_cuda_version = 0; +- std::string app_cuda_version_command = "ldd " + get_app_binary() + " | grep libcudart.so | sed 's/.*libcudart.so.\\(.*\\) =>.*/\\1/' > " + fname; +- system(app_cuda_version_command.c_str()); +- FILE * cmd = fopen(fname, "r"); +- char buf[256]; +- while (fgets(buf, sizeof(buf), cmd) != 0) { +- std::cout << buf; +- app_cuda_version = atoi(buf); +- } +- fclose(cmd); +- if ( app_cuda_version == 0 ) { +- printf( "Error - Cannot detect the app's CUDA version.\n" ); +- exit(1); +- } ++ int app_cuda_version = get_app_cuda_version(); ++ assert( app_cuda_version == CUDART_VERSION / 1000 && "The app must be compiled with same major version as the simulator." ); + const char* filename; +- if ( app_cuda_version < 6 ) { ++#if CUDART_VERSION < 6000 + // FatBin handle from the .fatbin.c file (one of the intermediate files generated by NVCC) + typedef struct {int m; int v; const unsigned long long* d; char* f;} __fatDeviceText __attribute__ ((aligned (8))); + __fatDeviceText * fatDeviceText = (__fatDeviceText *) fatCubin; +@@ -1814,9 +1811,9 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) + char * pfatbin = (char*) fatDeviceText->d; + int offset = *((int*)(pfatbin+48)); + filename = (pfatbin+16+offset); +- } else { ++#else + filename = "default"; +- } ++#endif + + // The extracted file name is associated with a fat_cubin_handle passed + // into cudaLaunch(). Inside cudaLaunch(), the associated file name is +diff --git a/setup_environment b/setup_environment +index 8bcb9aa..e5cd4fa 100644 +--- a/setup_environment ++++ b/setup_environment +@@ -134,18 +134,26 @@ else + echo "configured without a power model."; + fi + ++export PTXAS_CUDA_INSTALL_PATH=$CUDA_INSTALL_PATH ++echo ""; ++echo "----------------------------------------------------------------------------"; ++echo "WARNING - If you only care about PTX execution, ignore this warning. GPGPU-Sim supports PTX execution in modern CUDA." + if [ $CUDA_VERSION_NUMBER -gt 4200 ]; then +- echo ""; +- echo "----------------------------------------------------------------------------"; +- echo "WARNING - If you only care about PTX execution, ignore this warning. GPGPU-Sim supports PTX execution in modern CUDA." +- echo "If you want to run PTXPLUS, download SDK 4.2 (in addition to your modern CUDA) and set \$CUDA_42_INSTALL_PATH." +- echo "The following text describes why:"; +- echo "If you are using PTXPLUS, only sm_1x is supported and it requires that the app binaries are compiled in CUDA 4.2 or less."; +- echo "New versions of CUDA tools have dropped parsing support for sm_1x - therefore to use PTXPLUS with a modern card configuration, you must"; +- echo "have both CUDA 4.2 and a modern CUDA installed. When running PTXPLUS under a CUDA version >= 4.2, the simulator will fail unless \$CUDA_42_INSTALL_PATH"; +- echo "has been set properly. You must set this manually."; +- echo "----------------------------------------------------------------------------" ++ echo "If you want to run PTXPLUS (sm_1x SASS) with a modern card configuration, the apps and simulator must be compiled with CUDA 4.2." ++ echo "You can still run a PASCAL configuration when compiling with 4.2 by setting the \$PTXAS_CUDA_INSTALL_PATH directory environment variable." ++else ++ echo "If you want to run PTXPLUS (sm_1x SASS) with a modern card configuration - set the envronment variable" ++ echo "\$PTXAS_CUDA_INSTALL_PATH to point a CUDA version compabible with your card configurations (i.e. 8+ for PASCAL, 9+ for VOLTA etc..)" ++ echo "For example: \"export \$PTXAS_CUDA_INSTALL_PATH=/usr/local/cuda-9.1\"" + fi ++echo "The following text describes why:"; ++echo "If you are using PTXPLUS, only sm_1x is supported and it requires that the app and simulator binaries are compiled in CUDA 4.2 or less."; ++echo "The simulator requires it since CUDA headers desribe struct sizes in the exec which change from gen to gen."; ++echo "The apps require 4.2 because new versions of CUDA tools have dropped parsing support for generating sm_1x"; ++echo "When running using modern config (i.e. pascal) and PTXPLUS with CUDA 4.2, the \$PTXAS_CUDA_INSTALL_PATH env variable is required to get proper register usage" ++echo "(and hence occupancy) using a version of CUDA that knows the register usage on the real card." ++echo ""; ++echo "----------------------------------------------------------------------------"; + + echo "setup_environment succeeded"; + +diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc +index 22f03b9..d8e04d4 100644 +--- a/src/cuda-sim/ptx_loader.cc ++++ b/src/cuda-sim/ptx_loader.cc +@@ -339,7 +339,7 @@ void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num + snprintf(extra_flags,1024,"--compile-only --gpu-name=sm_%u",g_occupancy_sm_number); + #endif + +- snprintf(commandline,1024,"$CUDA_INSTALL_PATH/bin/ptxas %s -v %s --output-file /dev/null 2> %s", ++ snprintf(commandline,1024,"$PTXAS_CUDA_INSTALL_PATH/bin/ptxas %s -v %s --output-file /dev/null 2> %s", + extra_flags, fname2, tempfile_ptxinfo); + printf("GPGPU-Sim PTX: generating ptxinfo using \"%s\"\n", commandline); + result = system(commandline); diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index 5964a4d..53436d3 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -182,6 +182,8 @@ cudaError_t g_last_cudaError = cudaSuccess; extern stream_manager *g_stream_manager; + + void register_ptx_function( const char *name, function_info *impl ) { // no longer need this @@ -1341,6 +1343,29 @@ std::string get_app_binary(){ return self_exe_path; } +static int get_app_cuda_version() { + int app_cuda_version = 0; + char fname[1024]; + snprintf(fname,1024,"_app_cuda_version_XXXXXX"); + int fd=mkstemp(fname); + close(fd); + std::string app_cuda_version_command = "ldd " + get_app_binary() + " | grep libcudart.so | sed 's/.*libcudart.so.\\(.*\\) =>.*/\\1/' > " + fname; + system(app_cuda_version_command.c_str()); + FILE * cmd = fopen(fname, "r"); + char buf[256]; + while (fgets(buf, sizeof(buf), cmd) != 0) { + std::cout << buf; + app_cuda_version = atoi(buf); + } + fclose(cmd); + if ( app_cuda_version == 0 ) { + printf( "Error - Cannot detect the app's CUDA version.\n" ); + exit(1); + } + return app_cuda_version; +} + + //! Call cuobjdump to extract everything (-elf -sass -ptx) /*! * This Function extract the whole PTX (for all the files) using cuobjdump @@ -1366,23 +1391,10 @@ void extract_code_using_cuobjdump(){ // Running cuobjdump using dynamic link to current process // Needs the option '-all' to extract PTX from CDP-enabled binary extern bool g_cdp_enabled; - const int current_cuda = atoi(getenv("CUDA_VERSION_NUMBER")); - if ( context->get_device()->get_gpgpu()->get_config().convert_to_ptxplus() && current_cuda > 4020 ) { - const char* cuda_42_path = getenv("CUDA_42_INSTALL_PATH"); - if ( NULL == cuda_42_path ) { - printf("You are using a new version of CUDA and PTXPLUS. You are required to have CUDA SDK 4.2 and explicitly " - "point to it via the environment variable $CUDA_42_INSTALL_PATH.\n\n"); - exit(1); - } else { - command = "$CUDA_42_INSTALL_PATH/bin/cuobjdump"; - } - } else { - command = "$CUDA_42_INSTALL_PATH/bin/cuobjdump"; - } if(!g_cdp_enabled) - command += " -ptx -elf -sass " + app_binary + " > " + fname; + command = "$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass " + app_binary + " > " + fname; else - command += " -ptx -elf -sass -all " + app_binary + " > " + fname; + command = "$CUDA_INSTALL_PATH/bin/cuobjdump -ptx -elf -sass -all " + app_binary + " > " + fname; bool parse_output = true; int result = system(command.c_str()); if(result) { @@ -1765,6 +1777,7 @@ void cuobjdumpParseBinary(unsigned int handle){ //TODO: Remove temporarily files as per configurations } + void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) { #if (CUDART_VERSION < 2010) @@ -1783,26 +1796,10 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) // Making this a runtime variable based on the app, enables GPGPU-Sim compiled // with a newer version of CUDA to run apps compiled with older versions of // CUDA. This is especially useful for PTXPLUS execution. - char fname[1024]; - snprintf(fname,1024,"_app_cuda_version_XXXXXX"); - int fd=mkstemp(fname); - close(fd); - int app_cuda_version = 0; - std::string app_cuda_version_command = "ldd " + get_app_binary() + " | grep libcudart.so | sed 's/.*libcudart.so.\\(.*\\) =>.*/\\1/' > " + fname; - system(app_cuda_version_command.c_str()); - FILE * cmd = fopen(fname, "r"); - char buf[256]; - while (fgets(buf, sizeof(buf), cmd) != 0) { - std::cout << buf; - app_cuda_version = atoi(buf); - } - fclose(cmd); - if ( app_cuda_version == 0 ) { - printf( "Error - Cannot detect the app's CUDA version.\n" ); - exit(1); - } + int app_cuda_version = get_app_cuda_version(); + assert( app_cuda_version == CUDART_VERSION / 1000 && "The app must be compiled with same major version as the simulator." ); const char* filename; - if ( app_cuda_version < 6 ) { +#if CUDART_VERSION < 6000 // FatBin handle from the .fatbin.c file (one of the intermediate files generated by NVCC) typedef struct {int m; int v; const unsigned long long* d; char* f;} __fatDeviceText __attribute__ ((aligned (8))); __fatDeviceText * fatDeviceText = (__fatDeviceText *) fatCubin; @@ -1814,9 +1811,9 @@ void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin ) char * pfatbin = (char*) fatDeviceText->d; int offset = *((int*)(pfatbin+48)); filename = (pfatbin+16+offset); - } else { +#else filename = "default"; - } +#endif // The extracted file name is associated with a fat_cubin_handle passed // into cudaLaunch(). Inside cudaLaunch(), the associated file name is diff --git a/setup_environment b/setup_environment index 8bcb9aa..e5cd4fa 100644 --- a/setup_environment +++ b/setup_environment @@ -134,18 +134,26 @@ else echo "configured without a power model."; fi +export PTXAS_CUDA_INSTALL_PATH=$CUDA_INSTALL_PATH +echo ""; +echo "----------------------------------------------------------------------------"; +echo "WARNING - If you only care about PTX execution, ignore this warning. GPGPU-Sim supports PTX execution in modern CUDA." if [ $CUDA_VERSION_NUMBER -gt 4200 ]; then - echo ""; - echo "----------------------------------------------------------------------------"; - echo "WARNING - If you only care about PTX execution, ignore this warning. GPGPU-Sim supports PTX execution in modern CUDA." - echo "If you want to run PTXPLUS, download SDK 4.2 (in addition to your modern CUDA) and set \$CUDA_42_INSTALL_PATH." - echo "The following text describes why:"; - echo "If you are using PTXPLUS, only sm_1x is supported and it requires that the app binaries are compiled in CUDA 4.2 or less."; - echo "New versions of CUDA tools have dropped parsing support for sm_1x - therefore to use PTXPLUS with a modern card configuration, you must"; - echo "have both CUDA 4.2 and a modern CUDA installed. When running PTXPLUS under a CUDA version >= 4.2, the simulator will fail unless \$CUDA_42_INSTALL_PATH"; - echo "has been set properly. You must set this manually."; - echo "----------------------------------------------------------------------------" + echo "If you want to run PTXPLUS (sm_1x SASS) with a modern card configuration, the apps and simulator must be compiled with CUDA 4.2." + echo "You can still run a PASCAL configuration when compiling with 4.2 by setting the \$PTXAS_CUDA_INSTALL_PATH directory environment variable." +else + echo "If you want to run PTXPLUS (sm_1x SASS) with a modern card configuration - set the envronment variable" + echo "\$PTXAS_CUDA_INSTALL_PATH to point a CUDA version compabible with your card configurations (i.e. 8+ for PASCAL, 9+ for VOLTA etc..)" + echo "For example: \"export \$PTXAS_CUDA_INSTALL_PATH=/usr/local/cuda-9.1\"" fi +echo "The following text describes why:"; +echo "If you are using PTXPLUS, only sm_1x is supported and it requires that the app and simulator binaries are compiled in CUDA 4.2 or less."; +echo "The simulator requires it since CUDA headers desribe struct sizes in the exec which change from gen to gen."; +echo "The apps require 4.2 because new versions of CUDA tools have dropped parsing support for generating sm_1x"; +echo "When running using modern config (i.e. pascal) and PTXPLUS with CUDA 4.2, the \$PTXAS_CUDA_INSTALL_PATH env variable is required to get proper register usage" +echo "(and hence occupancy) using a version of CUDA that knows the register usage on the real card." +echo ""; +echo "----------------------------------------------------------------------------"; echo "setup_environment succeeded"; diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc index 22f03b9..d8e04d4 100644 --- a/src/cuda-sim/ptx_loader.cc +++ b/src/cuda-sim/ptx_loader.cc @@ -339,7 +339,7 @@ void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num snprintf(extra_flags,1024,"--compile-only --gpu-name=sm_%u",g_occupancy_sm_number); #endif - snprintf(commandline,1024,"$CUDA_INSTALL_PATH/bin/ptxas %s -v %s --output-file /dev/null 2> %s", + snprintf(commandline,1024,"$PTXAS_CUDA_INSTALL_PATH/bin/ptxas %s -v %s --output-file /dev/null 2> %s", extra_flags, fname2, tempfile_ptxinfo); printf("GPGPU-Sim PTX: generating ptxinfo using \"%s\"\n", commandline); result = system(commandline); -- cgit v1.3