3 files changed, 54 insertions, 9 deletions
diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc
index 737b94a..a55fb79 100644
--- a/src/cuda-sim/cuda-sim.cc
+++ b/src/cuda-sim/cuda-sim.cc
@@ -79,6 +79,7 @@
 #include "ptx-stats.h"
 #include "ptx_loader.h"
 #include "ptx_parser.h"
+#include "../gpgpu-sim/gpu-sim.h"
 
 extern bool g_interactive_debugger_enabled;
 
@@ -759,7 +760,6 @@ unsigned datatype2size( unsigned data_type )
    return data_size; 
 }
 
-extern unsigned long long  gpu_sim_cycle;
 unsigned g_warp_active_mask;
 
 void function_info::ptx_exec_inst( ptx_thread_info *thread, 
diff --git a/src/cuda-sim/cuda-sim.h b/src/cuda-sim/cuda-sim.h
index 2c5e0a7..7e05f7e 100644
--- a/src/cuda-sim/cuda-sim.h
+++ b/src/cuda-sim/cuda-sim.h
@@ -2,6 +2,7 @@
 #define CUDASIM_H_INCLUDED
 
 #include "../abstract_hardware_model.h"
+#include "dram_callback.h"
 #include <stdlib.h>
 #include <map>
 #include <string>
@@ -17,10 +18,16 @@ extern int g_debug_execution;
 extern int g_debug_thread_uid;
 extern std::map<std::string,function_info*> *g_kernel_name_to_function_lookup;
 
-extern void   gpgpu_cuda_ptx_sim_init_grid(const char *kernel_key,struct gpgpu_ptx_sim_arg *args, struct dim3 gridDim, struct dim3 blockDim );
-extern void   gpgpu_opencl_ptx_sim_init_grid(class function_info *entry,struct gpgpu_ptx_sim_arg *args, struct dim3 gridDim, struct dim3 blockDim );
+extern void   gpgpu_cuda_ptx_sim_init_grid( const char *kernel_key,
+                                            struct gpgpu_ptx_sim_arg *args, 
+                                            struct dim3 gridDim, 
+                                            struct dim3 blockDim );
+extern void   gpgpu_opencl_ptx_sim_init_grid(class function_info *entry,
+                                            struct gpgpu_ptx_sim_arg *args, 
+                                            struct dim3 gridDim, 
+                                            struct dim3 blockDim );
 extern void   gpgpu_cuda_ptx_sim_main_func( const char *kernel_key, dim3 gridDim, dim3 blockDim, struct gpgpu_ptx_sim_arg *);
-extern void print_splash();
+extern void   print_splash();
 extern void*  gpgpu_ptx_sim_malloc( size_t count );
 extern void*  gpgpu_ptx_sim_mallocarray( size_t count );
 extern void   gpgpu_ptx_sim_memcpy_to_gpu( size_t dst_start_addr, const void *src, size_t count );
@@ -43,5 +50,47 @@ extern void read_sim_environment_variables();
 extern void register_function_implementation( const char *name, function_info *impl );
 extern void ptxinfo_cuda_addinfo();
 extern void ptxinfo_opencl_addinfo( std::map<std::string,function_info*> &kernels );
+extern void ptx_sim_free_sm( class ptx_thread_info** thread_info );
+unsigned ptx_sim_init_thread( class ptx_thread_info** thread_info,
+                              int sid,
+                              unsigned tid,
+                              unsigned threads_left,
+                              unsigned num_threads, 
+                              class core_t *core, 
+                              unsigned hw_cta_id, 
+                              unsigned hw_warp_id );
+unsigned ptx_sim_cta_size();
+const struct gpgpu_ptx_sim_kernel_info* ptx_sim_kernel_info();
+void set_option_gpgpu_spread_blocks_across_cores(int option);
+int ptx_thread_done( void *thd );
+unsigned ptx_thread_donecycle( void *thr );
+int ptx_thread_get_next_pc( void *thd );
+void* ptx_thread_get_next_finfo( void *thd );
+int ptx_thread_at_barrier( void *thd );
+int ptx_thread_all_at_barrier( void *thd );
+unsigned long long ptx_thread_get_cta_uid( void *thd );
+void ptx_thread_reset_barrier( void *thd );
+void ptx_thread_release_barrier( void *thd );
+void ptx_print_insn( address_type pc, FILE *fp );
+unsigned int ptx_set_tex_cache_linesize( unsigned linesize);
+void ptx_decode_inst( void *thd, 
+                      unsigned *op, 
+                      int *i1, 
+                      int *i2, 
+                      int *i3, 
+                      int *i4, 
+                      int *o1, 
+                      int *o2, 
+                      int *o3, 
+                      int *o4, 
+                      int *vectorin, 
+                      int *vectorout, 
+                      int *arch_reg );
+void ptx_exec_inst( void *thd, 
+                    address_type *addr, 
+                    memory_space_t *space, 
+                    unsigned *data_size, 
+                    dram_callback_t* callback, 
+                    unsigned warp_active_mask );
 
 #endif
diff --git a/src/cuda-sim/ptx_sim.cc b/src/cuda-sim/ptx_sim.cc
index 31c8c81..2156bd6 100644
--- a/src/cuda-sim/ptx_sim.cc
+++ b/src/cuda-sim/ptx_sim.cc
@@ -64,6 +64,7 @@
 #include "ptx_sim.h"
 #include <string>
 #include "ptx_ir.h"
+#include "../gpgpu-sim/gpu-sim.h"
 
 void feature_not_implemented( const char *f );
 
@@ -249,8 +250,6 @@ const ptx_version &ptx_thread_info::get_ptx_version() const
    return m_func_info->get_ptx_version(); 
 }
 
-extern unsigned long long  gpu_sim_cycle;
-
 void ptx_thread_info::set_done() 
 {
    assert( !m_at_barrier );
@@ -258,9 +257,6 @@ void ptx_thread_info::set_done()
    m_cycle_done = gpu_sim_cycle; 
 }
 
-extern unsigned long long  gpu_sim_cycle;
-extern signed long long gpu_tot_sim_cycle;
-
 unsigned ptx_thread_info::get_builtin( int builtin_id, unsigned dim_mod ) 
 {
    assert( m_valid );