diff options
| -rw-r--r-- | src/cuda-sim/cuda-sim.cc | 12 | ||||
| -rw-r--r-- | src/cuda-sim/instructions.cc | 16 | ||||
| -rw-r--r-- | src/cuda-sim/ptx_ir.cc | 25 | ||||
| -rw-r--r-- | src/cuda-sim/ptx_ir.h | 14 | ||||
| -rw-r--r-- | src/cuda-sim/ptx_sim.cc | 11 | ||||
| -rw-r--r-- | src/cuda-sim/ptx_sim.h | 6 |
6 files changed, 63 insertions, 21 deletions
diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index 41acd62..2f66d51 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -459,6 +459,7 @@ function_info::function_info(int entry_point ) m_kernel_info.lmem = 0; m_kernel_info.regs = 0; m_kernel_info.smem = 0; + m_local_mem_framesize = (unsigned)-1; } void function_info::print_insn( unsigned pc, FILE * fp ) const @@ -986,11 +987,6 @@ void function_info::add_param_name_type_size( unsigned index, std::string name, } } -void function_info::add_local_param_name_type_size( std::string name, int type, size_t size ) -{ - m_ptx_local_params[name] = param_info(name, type, size); -} - void function_info::add_param_data( unsigned argn, struct gpgpu_ptx_sim_arg *args ) { const void *data = args->m_start; @@ -1345,11 +1341,9 @@ unsigned ptx_sim_init_thread( ptx_thread_info** thread_info,int sid,unsigned tid unsigned max_cta_per_sm = num_threads/cta_size; // e.g., 256 / 48 = 5 assert( max_cta_per_sm > 0 ); - //following sm_idx calculation assumes shaders being filled up depth-first? - unsigned sm_idx = sid*max_cta_per_sm + sm_offset; /* Is this assuming non-even block distribution? */ + unsigned sm_idx = sid*max_cta_per_sm + sm_offset; sm_idx = max_cta_per_sm*sid + tid/cta_size; - if (!gpgpu_option_spread_blocks_across_cores) { // update offset... if ( (sm_offset + 1) >= max_cta_per_sm ) { @@ -1905,7 +1899,7 @@ void gpgpu_ptx_assemble( std::string kname, void *kinfo ) { function_info *func_info = (function_info *)kinfo; if( func_info->is_extern() ) { - printf("GPGPU-Sim: skipping assembly for extern declared function \'%s\'", func_info->get_name().c_str() ); + printf("GPGPU-Sim PTX: skipping assembly for extern declared function \'%s\'\n", func_info->get_name().c_str() ); return; } g_func_info = func_info; diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 2b4b410..7606686 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -141,6 +141,8 @@ ptx_reg_t ptx_thread_info::get_operand_value( const operand_info &op ) result.u64 = baseaddr.u64 + op.get_addr_offset(); } else if ( info.is_param_kernel() ) { result = sym->get_address() + op.get_addr_offset(); + } else if ( info.is_param_local() ) { + result = sym->get_address() + op.get_addr_offset(); } else if ( info.is_global() ) { assert( op.get_addr_offset() == 0 ); result = sym->get_address(); @@ -151,7 +153,9 @@ ptx_reg_t ptx_thread_info::get_operand_value( const operand_info &op ) } else if ( op.is_shared() ) { result = op.get_symbol()->get_address() + op.get_addr_offset(); } else { - assert(0); + name = op.name().c_str(); + printf("GPGPU-Sim PTX: ERROR ** get_operand_value : unknown memory operand type for %s\n", name ); + abort(); } } else if ( op.is_literal() ) { @@ -168,7 +172,7 @@ ptx_reg_t ptx_thread_info::get_operand_value( const operand_info &op ) result = op.get_symbol()->get_address(); } else { name = op.name().c_str(); - printf("ERROR! %s is neither local, global, const, shared, label, literal, memory operand, builtin, or register!\n",name); + printf("GPGPU-Sim PTX: ERROR ** get_operand_value : unknown operand type for %s\n", name ); assert(0); } @@ -1460,11 +1464,13 @@ void decode_space( memory_space_t &space, ptx_thread_info *thread, memory_space unsigned hwtid = thread->get_hw_tid(); switch ( space ) { case global_space: mem = g_global_mem; break; - case local_space: mem = thread->m_local_mem; break; + case param_space_local: + case local_space: + mem = thread->m_local_mem; + addr += thread->get_local_mem_stack_pointer(); + break; case tex_space: mem = g_tex_mem; break; case surf_space: mem = g_surf_mem; break; - case param_space_local: - abort(); // finish this case param_space_kernel: mem = g_param_mem; break; case shared_space: mem = thread->m_shared_mem; break; case const_space: mem = g_global_mem; break; diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index da1c49b..2597ac5 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -455,7 +455,7 @@ void add_identifier( const char *identifier, int array_dim, unsigned array_ident g_sym_name_to_symbol_table[ identifier ] = g_current_symbol_table; break; case local_space: - printf("GPGPU-Sim PTX: allocating local region for \"%s\" from 0x%x to 0x%lx (local memory space)\n", + printf("GPGPU-Sim PTX: allocating stack frame region for .local \"%s\" from 0x%x to 0x%lx\n", identifier, g_current_symbol_table->get_local_next(), g_current_symbol_table->get_local_next() + num_bits/8 ); @@ -468,6 +468,15 @@ void add_identifier( const char *identifier, int array_dim, unsigned array_ident printf("GPGPU-Sim PTX: encountered texture directive %s.\n", identifier); break; case param_space_local: + printf("GPGPU-Sim PTX: allocating stack frame region for .param \"%s\" from 0x%x to 0x%lx\n", + identifier, + g_current_symbol_table->get_local_next(), + g_current_symbol_table->get_local_next() + num_bits/8 ); + fflush(stdout); + assert( (num_bits%8) == 0 ); + g_last_symbol->set_address( g_current_symbol_table->get_local_next() ); + g_current_symbol_table->alloc_local( num_bits/8 ); + break; case param_space_kernel: break; default: @@ -475,12 +484,20 @@ void add_identifier( const char *identifier, int array_dim, unsigned array_ident break; } - + assert( !ti.is_param_unclassified() ); if ( ti.is_param_kernel() ) { g_func_info->add_param_name_type_size(g_entry_func_param_index,identifier, ti.scalar_type(), num_bits ); g_entry_func_param_index++; - } else if ( ti.is_param_local() || ti.is_param_unclassified() ) { - g_func_info->add_local_param_name_type_size( identifier, ti.scalar_type(), num_bits ); + } else if ( ti.is_param_local() ) { + printf("GPGPU-Sim PTX: allocating stack frame region for input/output .param \"%s\" from 0x%x to 0x%lx\n", + identifier, + g_current_symbol_table->get_local_next(), + g_current_symbol_table->get_local_next() + num_bits/8 ); + fflush(stdout); + assert( (num_bits%8) == 0 ); + addr_t addr = g_current_symbol_table->get_local_next(); + g_last_symbol->set_address( addr ); + g_current_symbol_table->alloc_local( num_bits/8 ); } } diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index b143697..8f3d48e 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -196,6 +196,7 @@ public: m_is_const = false; m_is_global = false; m_is_local = false; + m_is_param_local = false; m_is_tex = false; m_is_func_addr = false; m_reg_num_valid = false; @@ -204,6 +205,7 @@ public: if ( type ) m_is_const = type->get_key().is_const(); if ( type ) m_is_global = type->get_key().is_global(); if ( type ) m_is_local = type->get_key().is_local(); + if ( type ) m_is_param_local = type->get_key().is_param_local(); if ( type ) m_is_tex = type->get_key().is_tex(); if ( type ) m_is_func_addr = type->get_key().is_func_addr(); } @@ -249,6 +251,7 @@ public: bool is_const() const { return m_is_const;} bool is_global() const { return m_is_global;} bool is_local() const { return m_is_local;} + bool is_param_local() const { return m_is_param_local; } bool is_tex() const { return m_is_tex;} bool is_func_addr() const { return m_is_func_addr; } @@ -289,6 +292,7 @@ private: bool m_is_const; bool m_is_global; bool m_is_local; + bool m_is_param_local; bool m_is_tex; bool m_is_func_addr; unsigned m_reg_num; @@ -374,6 +378,8 @@ public: m_type = symbolic_t; } else if ( addr->is_local() ) { m_type = symbolic_t; + } else if ( addr->is_param_local() ) { + m_type = symbolic_t; } else if ( addr->is_tex() ) { m_type = symbolic_t; } else if ( addr->is_func_addr() ) { @@ -944,7 +950,6 @@ public: m_kernel_params[ name ] = value; } void add_param_name_type_size( unsigned index, std::string name, int type, size_t size ); - void add_local_param_name_type_size( std::string name, int type, size_t size ); void add_param_data( unsigned argn, struct gpgpu_ptx_sim_arg *args ); void add_return_var( const symbol *rv ) { @@ -1009,9 +1014,15 @@ public: assert(pc <= s_g_pc_to_insn.size()); return s_g_pc_to_insn[pc - 1]; } + unsigned local_mem_framesize() const + { + assert( m_local_mem_framesize != (unsigned) -1 ); + return m_local_mem_framesize; + } private: unsigned m_uid; + unsigned m_local_mem_framesize; bool m_entry_point; bool m_extern; bool m_assembled; @@ -1021,7 +1032,6 @@ private: unsigned m_instr_mem_size; std::map<std::string,param_t> m_kernel_params; std::map<unsigned,param_info> m_ptx_kernel_param_info; - std::map<std::string,param_info> m_ptx_local_params; const symbol *m_return_var_sym; std::vector<const symbol*> m_args; std::list<ptx_instruction*> m_instructions; diff --git a/src/cuda-sim/ptx_sim.cc b/src/cuda-sim/ptx_sim.cc index 269b858..85e6185 100644 --- a/src/cuda-sim/ptx_sim.cc +++ b/src/cuda-sim/ptx_sim.cc @@ -241,6 +241,7 @@ ptx_thread_info::ptx_thread_info() m_RPC_updated = false; m_last_was_call = false; m_enable_debug_trace = false; + m_local_mem_stack_pointer = 0; } const ptx_version &ptx_thread_info::get_ptx_version() const @@ -447,6 +448,16 @@ void ptx_thread_info::dump_modifiedregs() } } +void ptx_thread_info::push_local_mem_stack() +{ + m_local_mem_stack_pointer += m_func_info->local_mem_framesize(); +} + +void ptx_thread_info::pop_local_mem_stack() +{ + m_local_mem_stack_pointer -= m_func_info->local_mem_framesize(); +} + void ptx_thread_info::set_npc( const function_info *f ) { m_NPC = f->get_start_PC(); diff --git a/src/cuda-sim/ptx_sim.h b/src/cuda-sim/ptx_sim.h index 8a74168..d4bdf48 100644 --- a/src/cuda-sim/ptx_sim.h +++ b/src/cuda-sim/ptx_sim.h @@ -424,6 +424,10 @@ public: void enable_debug_trace() { m_enable_debug_trace = true; } + unsigned get_local_mem_stack_pointer() const { return m_local_mem_stack_pointer; } + void push_local_mem_stack(); + void pop_local_mem_stack(); + public: addr_t m_last_effective_address; bool m_branch_taken; @@ -464,8 +468,8 @@ private: function_info *m_func_info; std::list<stack_entry> m_callstack; + unsigned m_local_mem_stack_pointer; - // typedef std::unordered_map<std::string,ptx_reg_t> reg_map_t; typedef std::unordered_map<const symbol*,ptx_reg_t> reg_map_t; std::list<reg_map_t> m_regs; |