4 files changed, 18 insertions, 5 deletions
diff --git a/src/cuda-sim/cuda-sim.h b/src/cuda-sim/cuda-sim.h
index 9049a84..abd32f9 100644
--- a/src/cuda-sim/cuda-sim.h
+++ b/src/cuda-sim/cuda-sim.h
@@ -98,7 +98,7 @@ public:
         delete[] m_warpAtBarrier;
     }
     //! executes all warps till completion 
-    void execute();
+    void execute(int inst_count, unsigned ctaid_cp);
     virtual void warp_exit( unsigned warp_id );
     virtual bool warp_waiting_at_barrier( unsigned warp_id ) const  
     {
@@ -108,7 +108,7 @@ public:
 private:
     void executeWarp(unsigned, bool &, bool &);
     //initializes threads in the CTA block which we are executing
-    void initializeCTA();
+    void initializeCTA(unsigned ctaid_cp);
     virtual void checkExecutionStatusAndUpdate(warp_inst_t &inst, unsigned t, unsigned tid)
     {
     if(m_thread[tid]==NULL || m_thread[tid]->is_done()){
diff --git a/src/cuda-sim/memory.cc b/src/cuda-sim/memory.cc
index 7bdf4d9..9554f55 100644
--- a/src/cuda-sim/memory.cc
+++ b/src/cuda-sim/memory.cc
@@ -44,9 +44,16 @@ template<unsigned BSIZE> memory_space_impl<BSIZE>::memory_space_impl( std::strin
    assert( m_log2_block_size != (unsigned)-1 );
 }
 
+template<unsigned BSIZE> void memory_space_impl<BSIZE>::write_only( mem_addr_t offset, mem_addr_t index, size_t length, const void *data)
+{
+   m_data[index].write(offset,length,(const unsigned char*)data);
+}
+
 template<unsigned BSIZE> void memory_space_impl<BSIZE>::write( mem_addr_t addr, size_t length, const void *data, class ptx_thread_info *thd, const ptx_instruction *pI)
 {
+
    mem_addr_t index = addr >> m_log2_block_size;
+
    if ( (addr+length) <= (index+1)*BSIZE ) {
       // fast route for intra-block access 
       unsigned offset = addr & (BSIZE-1);
@@ -142,8 +149,9 @@ template<unsigned BSIZE> void memory_space_impl<BSIZE>::read( mem_addr_t addr, s
 template<unsigned BSIZE> void memory_space_impl<BSIZE>::print( const char *format, FILE *fout ) const
 {
    typename map_t::const_iterator i_page;
-   for (i_page = m_data.begin(); i_page != m_data.end(); ++i_page) {
-      fprintf(fout, "%s - %#x:", m_name.c_str(), i_page->first);
+
+   for ( i_page = m_data.begin(); i_page != m_data.end(); ++i_page) {
+      fprintf(fout, "%s %08x:", m_name.c_str(), i_page->first);
       i_page->second.print(format, fout);
    }
 }
diff --git a/src/cuda-sim/memory.h b/src/cuda-sim/memory.h
index f785b8b..ab588bc 100644
--- a/src/cuda-sim/memory.h
+++ b/src/cuda-sim/memory.h
@@ -81,7 +81,7 @@ public:
    {
       unsigned int *i_data = (unsigned int*)m_data;
       for (int d = 0; d < (BSIZE / sizeof(unsigned int)); d++) {
-         if (d % 8 == 0) {
+         if (d % 1 == 0) {
             fprintf(fout, "\n");
          }
          fprintf(fout, format, i_data[d]);
@@ -104,6 +104,7 @@ class memory_space
 public:
    virtual ~memory_space() {}
    virtual void write( mem_addr_t addr, size_t length, const void *data, ptx_thread_info *thd, const ptx_instruction *pI ) = 0;
+   virtual void write_only( mem_addr_t index, mem_addr_t offset,  size_t length, const void *data ) = 0;
    virtual void read( mem_addr_t addr, size_t length, void *data ) const = 0;
    virtual void print( const char *format, FILE *fout ) const = 0;
    virtual void set_watch( addr_t addr, unsigned watchpoint ) = 0;
@@ -114,8 +115,10 @@ public:
    memory_space_impl( std::string name, unsigned hash_size );
 
    virtual void write( mem_addr_t addr, size_t length, const void *data, ptx_thread_info *thd, const ptx_instruction *pI );
+   virtual void write_only( mem_addr_t index, mem_addr_t offset, size_t length, const void *data);
    virtual void read( mem_addr_t addr, size_t length, void *data ) const;
    virtual void print( const char *format, FILE *fout ) const;
+   
    virtual void set_watch( addr_t addr, unsigned watchpoint ); 
 
 private:
diff --git a/src/cuda-sim/ptx_sim.h b/src/cuda-sim/ptx_sim.h
index 0a6e8eb..d226fbe 100644
--- a/src/cuda-sim/ptx_sim.h
+++ b/src/cuda-sim/ptx_sim.h
@@ -306,6 +306,8 @@ public:
 
    const ptx_version &get_ptx_version() const;
    void set_reg( const symbol *reg, const ptx_reg_t &value );
+   void print_reg_thread (char * fname);
+   void resume_reg_thread(char * fname,  symbol_table * symtab);
    ptx_reg_t get_reg( const symbol *reg );
    ptx_reg_t get_operand_value( const operand_info &op, operand_info dstInfo, unsigned opType, ptx_thread_info *thread, int derefFlag );
    void set_operand_value( const operand_info &dst, const ptx_reg_t &data, unsigned type, ptx_thread_info *thread, const ptx_instruction *pI );