diff options
Diffstat (limited to 'src/cuda-sim/cuda-sim.cc')
| -rw-r--r-- | src/cuda-sim/cuda-sim.cc | 27 |
1 files changed, 23 insertions, 4 deletions
diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index a34b99b..946043a 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -212,7 +212,9 @@ void function_info::ptx_assemble() m_start_PC = PC; addr_t n=0; // offset in m_instr_mem - s_g_pc_to_insn.reserve(s_g_pc_to_insn.size() + MAX_INST_SIZE*m_instructions.size()); + //Why s_g_pc_to_insn.size() is needed to reserve additional memory for insts? reserve is cumulative. + //s_g_pc_to_insn.reserve(s_g_pc_to_insn.size() + MAX_INST_SIZE*m_instructions.size()); + s_g_pc_to_insn.reserve(MAX_INST_SIZE*m_instructions.size()); for ( i=m_instructions.begin(); i != m_instructions.end(); i++ ) { ptx_instruction *pI = *i; if ( pI->is_label() ) { @@ -250,11 +252,13 @@ void function_info::ptx_assemble() target.set_type(label_t); } } - + m_n = n; printf(" done.\n"); fflush(stdout); - printf("GPGPU-Sim PTX: finding reconvergence points for \'%s\'...\n", m_name.c_str() ); + //disable pdom analysis here and do it at runtime +#if 0 + printf("GPGPU-Sim PTX: finding reconvergence points for \'%s\'...\n", m_name.c_str() ); create_basic_blocks(); connect_basic_blocks(); bool modified = false; @@ -288,6 +292,7 @@ void function_info::ptx_assemble() fflush(stdout); m_assembled = true; +#endif } addr_t shared_to_generic( unsigned smid, addr_t addr ) @@ -1134,8 +1139,13 @@ 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 type_info *paramtype = param->type(); + int align_amount = paramtype->get_key().get_alignment_spec(); + align_amount = (align_amount == -1) ? size : align_amount; + param_address = (param_address + align_amount - 1) / align_amount * align_amount; //aligned + 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 for (size_t idx = 0; idx < size; idx += word_size) { const char *pdata = reinterpret_cast<const char*>(param_value.pdata) + idx; // cast to char * for ptr arithmetic param_mem->write(param_address + idx, word_size, pdata,NULL,NULL); @@ -1799,6 +1809,15 @@ void gpgpu_cuda_ptx_sim_main_func( kernel_info_t &kernel, bool openCL ) //using a shader core object for book keeping, it is not needed but as most function built for performance simulation need it we use it here extern gpgpu_sim *g_the_gpu; + //before we execute, we should do PDOM analysis for functional simulation scenario. + function_info *kernel_func_info = kernel.entry(); + if (kernel_func_info->is_pdom_set()) { + printf("GPGPU-Sim PTX: PDOM analysis already done for %s \n", kernel.name().c_str() ); + } else { + printf("GPGPU-Sim PTX: finding reconvergence points for \'%s\'...\n", kernel.name().c_str() ); + kernel_func_info->do_pdom(); + kernel_func_info->set_pdom(); + } //we excute the kernel one CTA (Block) at the time, as synchronization functions work block wise while(!kernel.no_more_ctas_to_run()){ |