玩转gpgpu-sim 04记—— __cudaRegisterBinary() of gpgpu-sim 到底做了什么

官方文档：

GPGPU-Sim 3.x Manual

__cudaRegisterBinary(void*) 被执行到的代码逻辑如下：

void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin )
{
#if (CUDART_VERSION < 2010)printf("GPGPU-Sim PTX: ERROR ** this version of GPGPU-Sim requires CUDA 2.1 or higher\n");exit(1);
#endifCUctx_st *context = GPGPUSim_Context();static unsigned next_fat_bin_handle = 1;if(context->get_device()->get_gpgpu()->get_config().use_cuobjdump()) {// The following workaround has only been verified on 64-bit systems. if (sizeof(void*) == 4) printf("GPGPU-Sim PTX: FatBin file name extraction has not been tested on 32-bit system.\n"); // 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); // 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// PTX/SASS code for the launched kernel function.  // This allows us to work around the fact that cuobjdump only outputs the// file name associated with each section. unsigned long long fat_cubin_handle = next_fat_bin_handle;next_fat_bin_handle++;printf("GPGPU-Sim PTX: __cudaRegisterFatBinary, fat_cubin_handle = %llu, filename=%s\n", fat_cubin_handle, filename);/*!* This function extracts all data from all files in first call* then for next calls, only returns the appropriate number*/assert(fat_cubin_handle >= 1);if (fat_cubin_handle==1) cuobjdumpInit();cuobjdumpRegisterFatBinary(fat_cubin_handle, filename);return (void**)fat_cubin_handle;}else{ ... }}

1. 调用关系

刚开始一波的调用关系如下：

代码方便索引，此处整理的整体关系为下面的函数调用上面首先定义的函数：

class gpgpu_functional_sim_config 
{	...int m_ptx_use_cuobjdump;...
}void gpgpu_functional_sim_config::reg_options(class OptionParser * opp)
{	...option_parser_register(opp, "-gpgpu_ptx_use_cuobjdump", OPT_BOOL,&m_ptx_use_cuobjdump,"Use cuobjdump to extract ptx and sass from binaries","1");//CUDART_VERSION >= 4000...
}gpgpu_sim *gpgpu_ptx_sim_init_perf()
{	...g_the_gpu_config.reg_options(opp);...
}class _cuda_device_id *GPGPUSim_Init()
{	...gpgpu_sim *the_gpu = gpgpu_ptx_sim_init_perf();the_gpu->set_prop(prop);the_device = new _cuda_device_id(the_gpu);start_sim_thread(1);...
}void** CUDARTAPI __cudaRegisterFatBinary( void *fatCubin )
{	...static CUctx_st* GPGPUSim_Context()class _cuda_device_id *GPGPUSim_Init()CUctx_st( _cuda_device_id *gpu ) { m_gpu = gpu; }//the_context = new CUctx_st(the_gpu);cuobjdumpInit();cuobjdumpRegisterFatBinary(fat_cubin_handle, filename);...
}