Merge remote-tracking branch 'upstream/dev' into dev

4 files changed, 58 insertions, 11 deletions

diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h
index ef4cf48..e726ab9 100644
--- a/src/cuda-sim/ptx_ir.h
+++ b/src/cuda-sim/ptx_ir.h
@@ -1293,6 +1293,17 @@ public:
    bool is_pdom_set() const { return pdom_done; } //return pdom flag
    void set_pdom() { pdom_done = true; } //set pdom flag
 
+   void add_config_param( size_t size, unsigned alignment ){
+      unsigned offset = 0;
+      if (m_param_configs.size()>0){
+          unsigned offset_nom = m_param_configs.back().first + m_param_configs.back().second;
+          offset = offset_nom%alignment ? (offset_nom/alignment + 1) * alignment : offset_nom;
+      }
+      m_param_configs.push_back(std::pair<size_t,unsigned>(size, offset));
+   }
+
+   std::pair<size_t, unsigned> get_param_config(unsigned param_num) const { return m_param_configs[param_num]; }
+
 private:
    unsigned m_uid;
    unsigned m_local_mem_framesize;
@@ -1306,6 +1317,7 @@ 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::vector< std::pair<size_t, unsigned> > m_param_configs;
    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_parser.cc b/src/cuda-sim/ptx_parser.cc
index c418fac..e6d6325 100644
--- a/src/cuda-sim/ptx_parser.cc
+++ b/src/cuda-sim/ptx_parser.cc
@@ -61,6 +61,7 @@ memory_space_t g_ptr_spec = undefined_space;
 int g_scalar_type_spec = -1;
 int g_vector_spec = -1;
 int g_alignment_spec = -1;
+int g_size = -1;
 int g_extern_spec = 0;
 
 // variable declaration stuff:
@@ -116,6 +117,7 @@ void init_directive_state()
    g_vector_spec=-1;
    g_opcode=-1;
    g_alignment_spec = -1;
+   g_size = -1;
    g_extern_spec = 0;
    g_scalar_type.clear();
    g_operands.clear();
@@ -373,6 +375,9 @@ int pad_address (new_addr_type address, unsigned size, unsigned maxalign) {
 
 void add_identifier( const char *identifier, int array_dim, unsigned array_ident ) 
 {
+   if(array_ident==ARRAY_IDENTIFIER){
+       g_size *= array_dim;
+   }
    if( g_func_decl && (g_func_info == NULL) ) {
       // return variable decl...
       assert( g_add_identifier_cached__identifier == NULL );
@@ -562,10 +567,13 @@ void add_constptr(const char* identifier1, const char* identifier2, int offset)
 
 void add_function_arg()
 {
+   assert(g_size>0);
    if( g_func_info ) {
       PTX_PARSE_DPRINTF("add_function_arg \"%s\"", g_last_symbol->name().c_str() );
       g_func_info->add_arg(g_last_symbol);
+      g_func_info->add_config_param( g_size, g_alignment_spec );
    }
+
 }
 
 void add_extern_spec() 
@@ -617,6 +625,32 @@ void add_vector_spec(int spec )
 
 void add_scalar_type_spec( int type_spec ) 
 {
+   //save size of parameter
+   switch ( type_spec ) {
+      case B8_TYPE:
+      case S8_TYPE:
+      case U8_TYPE: 
+         g_size = 1; break;
+      case B16_TYPE:
+      case S16_TYPE:
+      case U16_TYPE:
+      case F16_TYPE: 
+         g_size = 2; break;
+      case B32_TYPE:
+      case S32_TYPE:
+      case U32_TYPE:
+      case F32_TYPE: 
+         g_size = 4; break;
+      case B64_TYPE:
+      case BB64_TYPE:
+      case S64_TYPE:
+      case U64_TYPE:
+      case F64_TYPE: 
+      case FF64_TYPE:
+         g_size = 8; break;
+      case BB128_TYPE: 
+         g_size = 16; break;
+   }
    PTX_PARSE_DPRINTF("add_scalar_type_spec \"%s\"", g_ptx_token_decode[type_spec].c_str());
    g_scalar_type.push_back( type_spec );
    if ( g_scalar_type.size() > 1 ) {
diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc
index 4640d65..0a5e11f 100644
--- a/src/gpgpu-sim/shader.cc
+++ b/src/gpgpu-sim/shader.cc
@@ -3025,19 +3025,20 @@ int register_bank(int regnum, int wid, unsigned num_banks, unsigned bank_warp_sh
    return bank % num_banks;
 }
 
-bool opndcoll_rfu_t::writeback( const warp_inst_t &inst )
+bool opndcoll_rfu_t::writeback( warp_inst_t &inst )
 {
    assert( !inst.empty() );
    std::list<unsigned> regs = m_shader->get_regs_written(inst);
-   std::list<unsigned>::iterator r;
-   unsigned n=0;
-   for( r=regs.begin(); r!=regs.end();r++,n++ ) {
-      unsigned reg = *r;
-      unsigned bank = register_bank(reg,inst.warp_id(),m_num_banks,m_bank_warp_shift);
-      if( m_arbiter.bank_idle(bank) ) {
-          m_arbiter.allocate_bank_for_write(bank,op_t(&inst,reg,m_num_banks,m_bank_warp_shift));
-      } else {
-          return false;
+   for( unsigned op=0; op < MAX_REG_OPERANDS; op++ ) {
+      int reg_num = inst.arch_reg.dst[op]; // this math needs to match that used in function_info::ptx_decode_inst
+      if( reg_num >= 0 ){ // valid register
+         unsigned bank = register_bank(reg_num,inst.warp_id(),m_num_banks,m_bank_warp_shift);
+         if( m_arbiter.bank_idle(bank) ) {
+             m_arbiter.allocate_bank_for_write(bank,op_t(&inst,reg_num,m_num_banks,m_bank_warp_shift));
+             inst.arch_reg.dst[op] = -1;
+         } else {
+             return false;
+         }
       }
    }
    for(unsigned i=0;i<(unsigned)regs.size();i++){
diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h
index 6201840..958abc4 100644
--- a/src/gpgpu-sim/shader.h
+++ b/src/gpgpu-sim/shader.h
@@ -525,7 +525,7 @@ public:
    void init( unsigned num_banks, shader_core_ctx *shader );
 
    // modifiers
-   bool writeback( const warp_inst_t &warp ); // might cause stall 
+   bool writeback( warp_inst_t &warp ); 
 
    void step()
    {