This changelist implements the following:

1. Adds support for using cuobjdump for both ptx and ptxplus execution. This has been tested with CUDA 4.0
  . Ptxplus is no longer supported through decuda/decuda_to_ptxplus
2. Adds support for converting the SASS output by cuobjdump to ptxplus. This has been tested with CUDA 4.0
  . The old path that extracts ptx from cubin files is still preserved
3. Adds a bank group model. (WARNING: memory config has changed, please adapt yours). To disable the bank groups model, set nbkgrp to 1 and tCCDL and tRTPL to 0

Diff the configuration files to learn about how to use those new options.

Merging

//depot/gpgpu_sim_research/fermi-test/distribution/...

to //depot/gpgpu_sim_research/fermi/distribution/...

[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 12023]

8 files changed, 233 insertions, 81 deletions

diff --git a/src/cuda-sim/Makefile b/src/cuda-sim/Makefile
index a475ca2..a70ee5f 100644
--- a/src/cuda-sim/Makefile
+++ b/src/cuda-sim/Makefile
@@ -159,13 +159,15 @@ $(OUTPUT_DIR)/lex.ptx_.o: ptx.tab.c
 ../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/gnu/stubs-64.h
 ../../build/release/cuda-sim/cuda_device_printf.o: ptx.tab.h ptx_sim.h
 ../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/stdlib.h
+../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/bits/waitflags.h
+../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/bits/waitstatus.h
+../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/endian.h
+../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/bits/endian.h
+../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/bits/byteswap.h
 ../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/sys/types.h
 ../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/bits/types.h
 ../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/bits/typesizes.h
 ../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/time.h
-../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/endian.h
-../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/bits/endian.h
-../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/bits/byteswap.h
 ../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/sys/select.h
 ../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/bits/select.h
 ../../build/release/cuda-sim/cuda_device_printf.o: /usr/include/bits/sigset.h
@@ -193,13 +195,15 @@ $(OUTPUT_DIR)/lex.ptx_.o: ptx.tab.c
 ../../build/release/cuda-sim/cuda-sim.o: /usr/include/bits/wordsize.h
 ../../build/release/cuda-sim/cuda-sim.o: /usr/include/gnu/stubs.h
 ../../build/release/cuda-sim/cuda-sim.o: /usr/include/gnu/stubs-64.h
+../../build/release/cuda-sim/cuda-sim.o: /usr/include/bits/waitflags.h
+../../build/release/cuda-sim/cuda-sim.o: /usr/include/bits/waitstatus.h
+../../build/release/cuda-sim/cuda-sim.o: /usr/include/endian.h
+../../build/release/cuda-sim/cuda-sim.o: /usr/include/bits/endian.h
+../../build/release/cuda-sim/cuda-sim.o: /usr/include/bits/byteswap.h
 ../../build/release/cuda-sim/cuda-sim.o: /usr/include/sys/types.h
 ../../build/release/cuda-sim/cuda-sim.o: /usr/include/bits/types.h
 ../../build/release/cuda-sim/cuda-sim.o: /usr/include/bits/typesizes.h
 ../../build/release/cuda-sim/cuda-sim.o: /usr/include/time.h
-../../build/release/cuda-sim/cuda-sim.o: /usr/include/endian.h
-../../build/release/cuda-sim/cuda-sim.o: /usr/include/bits/endian.h
-../../build/release/cuda-sim/cuda-sim.o: /usr/include/bits/byteswap.h
 ../../build/release/cuda-sim/cuda-sim.o: /usr/include/sys/select.h
 ../../build/release/cuda-sim/cuda-sim.o: /usr/include/bits/select.h
 ../../build/release/cuda-sim/cuda-sim.o: /usr/include/bits/sigset.h
@@ -242,6 +246,7 @@ $(OUTPUT_DIR)/lex.ptx_.o: ptx.tab.c
 ../../build/release/cuda-sim/cuda-sim.o: /usr/include/zconf.h
 ../../build/release/cuda-sim/cuda-sim.o: /usr/include/unistd.h
 ../../build/release/cuda-sim/cuda-sim.o: /usr/include/bits/posix_opt.h
+../../build/release/cuda-sim/cuda-sim.o: /usr/include/bits/environments.h
 ../../build/release/cuda-sim/cuda-sim.o: /usr/include/bits/confname.h
 ../../build/release/cuda-sim/cuda-sim.o: /usr/include/getopt.h
 ../../build/release/cuda-sim/cuda-sim.o: ../gpgpu-sim/scoreboard.h
@@ -261,13 +266,15 @@ $(OUTPUT_DIR)/lex.ptx_.o: ptx.tab.c
 ../../build/release/cuda-sim/instructions.o: /usr/include/gnu/stubs-64.h
 ../../build/release/cuda-sim/instructions.o: ptx.tab.h ptx_sim.h
 ../../build/release/cuda-sim/instructions.o: /usr/include/stdlib.h
+../../build/release/cuda-sim/instructions.o: /usr/include/bits/waitflags.h
+../../build/release/cuda-sim/instructions.o: /usr/include/bits/waitstatus.h
+../../build/release/cuda-sim/instructions.o: /usr/include/endian.h
+../../build/release/cuda-sim/instructions.o: /usr/include/bits/endian.h
+../../build/release/cuda-sim/instructions.o: /usr/include/bits/byteswap.h
 ../../build/release/cuda-sim/instructions.o: /usr/include/sys/types.h
 ../../build/release/cuda-sim/instructions.o: /usr/include/bits/types.h
 ../../build/release/cuda-sim/instructions.o: /usr/include/bits/typesizes.h
 ../../build/release/cuda-sim/instructions.o: /usr/include/time.h
-../../build/release/cuda-sim/instructions.o: /usr/include/endian.h
-../../build/release/cuda-sim/instructions.o: /usr/include/bits/endian.h
-../../build/release/cuda-sim/instructions.o: /usr/include/bits/byteswap.h
 ../../build/release/cuda-sim/instructions.o: /usr/include/sys/select.h
 ../../build/release/cuda-sim/instructions.o: /usr/include/bits/select.h
 ../../build/release/cuda-sim/instructions.o: /usr/include/bits/sigset.h
@@ -311,6 +318,7 @@ $(OUTPUT_DIR)/lex.ptx_.o: ptx.tab.c
 ../../build/release/cuda-sim/instructions.o: /usr/include/zconf.h
 ../../build/release/cuda-sim/instructions.o: /usr/include/unistd.h
 ../../build/release/cuda-sim/instructions.o: /usr/include/bits/posix_opt.h
+../../build/release/cuda-sim/instructions.o: /usr/include/bits/environments.h
 ../../build/release/cuda-sim/instructions.o: /usr/include/bits/confname.h
 ../../build/release/cuda-sim/instructions.o: /usr/include/getopt.h
 ../../build/release/cuda-sim/instructions.o: ../gpgpu-sim/scoreboard.h
@@ -336,11 +344,13 @@ $(OUTPUT_DIR)/lex.ptx_.o: ptx.tab.c
 ../../build/release/cuda-sim/memory.o: /usr/include/bits/stdio_lim.h
 ../../build/release/cuda-sim/memory.o: /usr/include/bits/sys_errlist.h
 ../../build/release/cuda-sim/memory.o: /usr/include/stdlib.h
-../../build/release/cuda-sim/memory.o: /usr/include/sys/types.h
-../../build/release/cuda-sim/memory.o: /usr/include/time.h
+../../build/release/cuda-sim/memory.o: /usr/include/bits/waitflags.h
+../../build/release/cuda-sim/memory.o: /usr/include/bits/waitstatus.h
 ../../build/release/cuda-sim/memory.o: /usr/include/endian.h
 ../../build/release/cuda-sim/memory.o: /usr/include/bits/endian.h
 ../../build/release/cuda-sim/memory.o: /usr/include/bits/byteswap.h
+../../build/release/cuda-sim/memory.o: /usr/include/sys/types.h
+../../build/release/cuda-sim/memory.o: /usr/include/time.h
 ../../build/release/cuda-sim/memory.o: /usr/include/sys/select.h
 ../../build/release/cuda-sim/memory.o: /usr/include/bits/select.h
 ../../build/release/cuda-sim/memory.o: /usr/include/bits/sigset.h
@@ -358,13 +368,15 @@ $(OUTPUT_DIR)/lex.ptx_.o: ptx.tab.c
 ../../build/release/cuda-sim/ptx_ir.o: /usr/include/gnu/stubs.h
 ../../build/release/cuda-sim/ptx_ir.o: /usr/include/gnu/stubs-64.h ptx.tab.h
 ../../build/release/cuda-sim/ptx_ir.o: ptx_sim.h /usr/include/stdlib.h
+../../build/release/cuda-sim/ptx_ir.o: /usr/include/bits/waitflags.h
+../../build/release/cuda-sim/ptx_ir.o: /usr/include/bits/waitstatus.h
+../../build/release/cuda-sim/ptx_ir.o: /usr/include/endian.h
+../../build/release/cuda-sim/ptx_ir.o: /usr/include/bits/endian.h
+../../build/release/cuda-sim/ptx_ir.o: /usr/include/bits/byteswap.h
 ../../build/release/cuda-sim/ptx_ir.o: /usr/include/sys/types.h
 ../../build/release/cuda-sim/ptx_ir.o: /usr/include/bits/types.h
 ../../build/release/cuda-sim/ptx_ir.o: /usr/include/bits/typesizes.h
 ../../build/release/cuda-sim/ptx_ir.o: /usr/include/time.h
-../../build/release/cuda-sim/ptx_ir.o: /usr/include/endian.h
-../../build/release/cuda-sim/ptx_ir.o: /usr/include/bits/endian.h
-../../build/release/cuda-sim/ptx_ir.o: /usr/include/bits/byteswap.h
 ../../build/release/cuda-sim/ptx_ir.o: /usr/include/sys/select.h
 ../../build/release/cuda-sim/ptx_ir.o: /usr/include/bits/select.h
 ../../build/release/cuda-sim/ptx_ir.o: /usr/include/bits/sigset.h
@@ -393,13 +405,15 @@ $(OUTPUT_DIR)/lex.ptx_.o: ptx.tab.c
 ../../build/release/cuda-sim/ptx_loader.o: /usr/include/gnu/stubs-64.h
 ../../build/release/cuda-sim/ptx_loader.o: ptx.tab.h ptx_sim.h
 ../../build/release/cuda-sim/ptx_loader.o: /usr/include/stdlib.h
+../../build/release/cuda-sim/ptx_loader.o: /usr/include/bits/waitflags.h
+../../build/release/cuda-sim/ptx_loader.o: /usr/include/bits/waitstatus.h
+../../build/release/cuda-sim/ptx_loader.o: /usr/include/endian.h
+../../build/release/cuda-sim/ptx_loader.o: /usr/include/bits/endian.h
+../../build/release/cuda-sim/ptx_loader.o: /usr/include/bits/byteswap.h
 ../../build/release/cuda-sim/ptx_loader.o: /usr/include/sys/types.h
 ../../build/release/cuda-sim/ptx_loader.o: /usr/include/bits/types.h
 ../../build/release/cuda-sim/ptx_loader.o: /usr/include/bits/typesizes.h
 ../../build/release/cuda-sim/ptx_loader.o: /usr/include/time.h
-../../build/release/cuda-sim/ptx_loader.o: /usr/include/endian.h
-../../build/release/cuda-sim/ptx_loader.o: /usr/include/bits/endian.h
-../../build/release/cuda-sim/ptx_loader.o: /usr/include/bits/byteswap.h
 ../../build/release/cuda-sim/ptx_loader.o: /usr/include/sys/select.h
 ../../build/release/cuda-sim/ptx_loader.o: /usr/include/bits/select.h
 ../../build/release/cuda-sim/ptx_loader.o: /usr/include/bits/sigset.h
@@ -433,13 +447,15 @@ $(OUTPUT_DIR)/lex.ptx_.o: ptx.tab.c
 ../../build/release/cuda-sim/ptx_parser.o: /usr/include/gnu/stubs-64.h
 ../../build/release/cuda-sim/ptx_parser.o: ptx.tab.h ptx_sim.h
 ../../build/release/cuda-sim/ptx_parser.o: /usr/include/stdlib.h
+../../build/release/cuda-sim/ptx_parser.o: /usr/include/bits/waitflags.h
+../../build/release/cuda-sim/ptx_parser.o: /usr/include/bits/waitstatus.h
+../../build/release/cuda-sim/ptx_parser.o: /usr/include/endian.h
+../../build/release/cuda-sim/ptx_parser.o: /usr/include/bits/endian.h
+../../build/release/cuda-sim/ptx_parser.o: /usr/include/bits/byteswap.h
 ../../build/release/cuda-sim/ptx_parser.o: /usr/include/sys/types.h
 ../../build/release/cuda-sim/ptx_parser.o: /usr/include/bits/types.h
 ../../build/release/cuda-sim/ptx_parser.o: /usr/include/bits/typesizes.h
 ../../build/release/cuda-sim/ptx_parser.o: /usr/include/time.h
-../../build/release/cuda-sim/ptx_parser.o: /usr/include/endian.h
-../../build/release/cuda-sim/ptx_parser.o: /usr/include/bits/endian.h
-../../build/release/cuda-sim/ptx_parser.o: /usr/include/bits/byteswap.h
 ../../build/release/cuda-sim/ptx_parser.o: /usr/include/sys/select.h
 ../../build/release/cuda-sim/ptx_parser.o: /usr/include/bits/select.h
 ../../build/release/cuda-sim/ptx_parser.o: /usr/include/bits/sigset.h
@@ -465,13 +481,15 @@ $(OUTPUT_DIR)/lex.ptx_.o: ptx.tab.c
 ../../build/release/cuda-sim/ptx_sim.o: /usr/include/bits/wordsize.h
 ../../build/release/cuda-sim/ptx_sim.o: /usr/include/gnu/stubs.h
 ../../build/release/cuda-sim/ptx_sim.o: /usr/include/gnu/stubs-64.h
+../../build/release/cuda-sim/ptx_sim.o: /usr/include/bits/waitflags.h
+../../build/release/cuda-sim/ptx_sim.o: /usr/include/bits/waitstatus.h
+../../build/release/cuda-sim/ptx_sim.o: /usr/include/endian.h
+../../build/release/cuda-sim/ptx_sim.o: /usr/include/bits/endian.h
+../../build/release/cuda-sim/ptx_sim.o: /usr/include/bits/byteswap.h
 ../../build/release/cuda-sim/ptx_sim.o: /usr/include/sys/types.h
 ../../build/release/cuda-sim/ptx_sim.o: /usr/include/bits/types.h
 ../../build/release/cuda-sim/ptx_sim.o: /usr/include/bits/typesizes.h
 ../../build/release/cuda-sim/ptx_sim.o: /usr/include/time.h
-../../build/release/cuda-sim/ptx_sim.o: /usr/include/endian.h
-../../build/release/cuda-sim/ptx_sim.o: /usr/include/bits/endian.h
-../../build/release/cuda-sim/ptx_sim.o: /usr/include/bits/byteswap.h
 ../../build/release/cuda-sim/ptx_sim.o: /usr/include/sys/select.h
 ../../build/release/cuda-sim/ptx_sim.o: /usr/include/bits/select.h
 ../../build/release/cuda-sim/ptx_sim.o: /usr/include/bits/sigset.h
@@ -513,6 +531,7 @@ $(OUTPUT_DIR)/lex.ptx_.o: ptx.tab.c
 ../../build/release/cuda-sim/ptx_sim.o: /usr/include/zconf.h
 ../../build/release/cuda-sim/ptx_sim.o: /usr/include/unistd.h
 ../../build/release/cuda-sim/ptx_sim.o: /usr/include/bits/posix_opt.h
+../../build/release/cuda-sim/ptx_sim.o: /usr/include/bits/environments.h
 ../../build/release/cuda-sim/ptx_sim.o: /usr/include/bits/confname.h
 ../../build/release/cuda-sim/ptx_sim.o: /usr/include/getopt.h
 ../../build/release/cuda-sim/ptx_sim.o: ../gpgpu-sim/scoreboard.h
@@ -530,13 +549,15 @@ $(OUTPUT_DIR)/lex.ptx_.o: ptx.tab.c
 ../../build/release/cuda-sim/ptx-stats.o: /usr/include/gnu/stubs-64.h
 ../../build/release/cuda-sim/ptx-stats.o: ptx.tab.h ptx_sim.h
 ../../build/release/cuda-sim/ptx-stats.o: /usr/include/stdlib.h
+../../build/release/cuda-sim/ptx-stats.o: /usr/include/bits/waitflags.h
+../../build/release/cuda-sim/ptx-stats.o: /usr/include/bits/waitstatus.h
+../../build/release/cuda-sim/ptx-stats.o: /usr/include/endian.h
+../../build/release/cuda-sim/ptx-stats.o: /usr/include/bits/endian.h
+../../build/release/cuda-sim/ptx-stats.o: /usr/include/bits/byteswap.h
 ../../build/release/cuda-sim/ptx-stats.o: /usr/include/sys/types.h
 ../../build/release/cuda-sim/ptx-stats.o: /usr/include/bits/types.h
 ../../build/release/cuda-sim/ptx-stats.o: /usr/include/bits/typesizes.h
 ../../build/release/cuda-sim/ptx-stats.o: /usr/include/time.h
-../../build/release/cuda-sim/ptx-stats.o: /usr/include/endian.h
-../../build/release/cuda-sim/ptx-stats.o: /usr/include/bits/endian.h
-../../build/release/cuda-sim/ptx-stats.o: /usr/include/bits/byteswap.h
 ../../build/release/cuda-sim/ptx-stats.o: /usr/include/sys/select.h
 ../../build/release/cuda-sim/ptx-stats.o: /usr/include/bits/select.h
 ../../build/release/cuda-sim/ptx-stats.o: /usr/include/bits/sigset.h
diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc
index d8c2c00..7d8d254 100644
--- a/src/cuda-sim/cuda-sim.cc
+++ b/src/cuda-sim/cuda-sim.cc
@@ -43,6 +43,7 @@
 #include "ptx_loader.h"
 #include "ptx_parser.h"
 #include "../gpgpu-sim/gpu-sim.h"
+#include "ptx_sim.h"
 #include "../gpgpusim_entrypoint.h"
 #include "decuda_pred_table/decuda_pred_table.h"
 #include "../stream_manager.h"
diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc
index 550a076..4537bd0 100644
--- a/src/cuda-sim/instructions.cc
+++ b/src/cuda-sim/instructions.cc
@@ -80,11 +80,12 @@ ptx_reg_t ptx_thread_info::get_reg( const symbol *reg )
       const std::string &name = reg->name();
       unsigned call_uid = m_callstack.back().m_call_uid;
       ptx_reg_t uninit_reg;
-      uninit_reg.u32 = 0xDEADBEEF;
+      uninit_reg.u32 = 0x0;
       set_reg(reg, uninit_reg); // give it a value since we are going to warn the user anyway
       std::string file_loc = get_location();
       if( !unfound_register_warned ) {
-          printf("GPGPU-Sim PTX: WARNING (%s) ** reading undefined register \'%s\' (cuid:%u). Setting to 0XDEADBEEF.\n",
+          printf("GPGPU-Sim PTX: WARNING (%s) ** reading undefined register \'%s\' (cuid:%u). Setting to 0X00000000. This is okay if you are simulating the native ISA"
+        		  "\n",
                  file_loc.c_str(), name.c_str(), call_uid );
           unfound_register_warned = true;
       }
@@ -818,13 +819,21 @@ void atom_callback( const inst_t* inst, ptx_thread_info* thread )
    const operand_info &src2 = pI->src2();    // b
 
    // Get operand values
-   src1_data = thread->get_operand_value(src1, dst, to_type, thread, 1);      // a
-   src2_data = thread->get_operand_value(src2, dst, to_type, thread, 1);      // b
+   if (dst.get_symbol()->type()){
+	   src1_data = thread->get_operand_value(src1, dst, to_type, thread, 1);      // a
+	   src2_data = thread->get_operand_value(src2, dst, to_type, thread, 1);      // b
+   } else {
+	   //This is the case whent he first argument (dest) is '_'
+	   src1_data = thread->get_operand_value(src1, src1, to_type, thread, 1);      // a
+	   src2_data = thread->get_operand_value(src2, src1, to_type, thread, 1);      // b
+   }
 
    // Copy value pointed to in operand 'a' into register 'd'
    // (i.e. copy src1_data to dst)
    thread->get_global_memory()->read(src1_data.u32,size/8,&data.s64);
-   thread->set_operand_value(dst, data, to_type, thread, pI);                         // Write value into register 'd'
+   if (dst.get_symbol()->type()){
+	   thread->set_operand_value(dst, data, to_type, thread, pI);                         // Write value into register 'd'
+   }
 
    // Get the atomic operation to be performed
    unsigned m_atomic_spec = pI->get_atomic();
@@ -1080,7 +1089,12 @@ void atom_impl( const ptx_instruction *pI, ptx_thread_info *thread )
    const operand_info &src1 = pI->src1();
    const operand_info &dst  = pI->dst();
    unsigned i_type = pI->get_type();
-   ptx_reg_t src1_data = thread->get_operand_value(src1, dst, i_type, thread, 1);
+   ptx_reg_t src1_data;
+   if (dst.get_symbol()->type()){
+	   src1_data = thread->get_operand_value(src1, dst, i_type, thread, 1);
+   } else {
+	   src1_data = thread->get_operand_value(src1, src1, i_type, thread, 1);
+   }
 
    memory_space_t space = pI->get_space();
 
@@ -1944,10 +1958,10 @@ void ex2_impl( const ptx_instruction *pI, ptx_thread_info *thread )
 }
 
 void exit_impl( const ptx_instruction *pI, ptx_thread_info *thread ) 
-{    
+{
+   thread->set_done();
    thread->exitCore();
    thread->registerExit();
-   thread->set_done();
 }
 
 void mad_def( const ptx_instruction *pI, ptx_thread_info *thread );
@@ -2755,9 +2769,9 @@ void ret_impl( const ptx_instruction *pI, ptx_thread_info *thread )
 { 
    bool empty = thread->callstack_pop();
    if( empty ) {
+   thread->set_done();
    thread->exitCore();
    thread->registerExit();
-   thread->set_done();
    }
 }
 
@@ -2766,9 +2780,9 @@ void retp_impl( const ptx_instruction *pI, ptx_thread_info *thread )
 {
    bool empty = thread->callstack_pop_plus();
    if( empty ) {
+   thread->set_done();
    thread->exitCore();
    thread->registerExit();
-   thread->set_done();
    }
 }
 
diff --git a/src/cuda-sim/ptx.l b/src/cuda-sim/ptx.l
index 70d4c23..8c330fc 100644
--- a/src/cuda-sim/ptx.l
+++ b/src/cuda-sim/ptx.l
@@ -381,8 +381,9 @@ breakaddr  TC; ptx_lval.int_value = BREAKADDR_OP; return OPCODE;
 [^\"]*	TC; ptx_lval.string_value = strdup(yytext); 
 }
 
-<INITIAL,NOT_OPCODE>.  TC; ptx_error((const char*)NULL);
+<*>\t@@DWARF.*\n
 
+<INITIAL,NOT_OPCODE>.  TC; ptx_error((const char*)NULL);
 %%
 
 extern int g_error_detected;
diff --git a/src/cuda-sim/ptx.y b/src/cuda-sim/ptx.y
index f0b67fb..430fc00 100644
--- a/src/cuda-sim/ptx.y
+++ b/src/cuda-sim/ptx.y
@@ -476,7 +476,9 @@ operand: IDENTIFIER  { add_scalar_operand( $1 ); }
 	| tex_operand
 	| IDENTIFIER PLUS INT_OPERAND { add_address_operand($1,$3); }
 	| IDENTIFIER LO_OPTION { add_scalar_operand( $1 ); change_operand_lohi(1);}
+	| MINUS IDENTIFIER LO_OPTION { add_scalar_operand( $2 ); change_operand_lohi(1); change_operand_neg();}
 	| IDENTIFIER HI_OPTION { add_scalar_operand( $1 ); change_operand_lohi(2);}
+	| MINUS IDENTIFIER HI_OPTION { add_scalar_operand( $2 ); change_operand_lohi(2); change_operand_neg();}
 	| IDENTIFIER PIPE IDENTIFIER { add_2vector_operand($1,$3); change_double_operand_type(-1);}
 	| IDENTIFIER PIPE IDENTIFIER LO_OPTION { add_2vector_operand($1,$3); change_double_operand_type(-1); change_operand_lohi(1);}
 	| IDENTIFIER PIPE IDENTIFIER HI_OPTION { add_2vector_operand($1,$3); change_double_operand_type(-1); change_operand_lohi(2);}
diff --git a/src/cuda-sim/ptx_loader.cc b/src/cuda-sim/ptx_loader.cc
index 61a357e..d3d87f7 100644
--- a/src/cuda-sim/ptx_loader.cc
+++ b/src/cuda-sim/ptx_loader.cc
@@ -29,6 +29,7 @@
 #include "ptx_ir.h"
 #include "cuda-sim.h"
 #include "ptx_parser.h"
+#include <unistd.h>
 #include <dirent.h>
 #include <fstream>
 
@@ -45,7 +46,7 @@ bool g_override_embedded_ptx = false;
 extern "C" int ptx_parse();
 extern "C" int ptx__scan_string(const char*);
 
-const char *g_ptxinfo_filename = NULL;
+extern "C" const char *g_ptxinfo_filename = NULL;
 extern "C" int ptxinfo_parse();
 extern "C" int ptxinfo_debug;
 extern "C" FILE *ptxinfo_in;
@@ -89,8 +90,9 @@ void print_ptx_file( const char *p, unsigned source_num, const char *filename )
    fflush(stdout);
 }
 
-char* gpgpu_ptx_sim_convert_ptx_to_ptxplus(const char *ptx_str, const char *cubin_str, unsigned source_num, bool save_converted )
+char* gpgpu_ptx_sim_convert_ptx_to_ptxplus(const char *ptx_str, const char *cubin_str, unsigned source_num, bool save_converted, bool cuobjdump)
 {
+
 	printf("GPGPU-Sim PTX: converting EMBEDDED .ptx file to ptxplus \n");
 
 	// Extract ptx to a file
@@ -104,49 +106,60 @@ char* gpgpu_ptx_sim_convert_ptx_to_ptxplus(const char *ptx_str, const char *cubi
     fprintf(ptxfile,"%s",ptx_str);
     fclose(ptxfile);
 
-    // Extract cubin to a file
-    char fname_cubin[1024];
-    snprintf(fname_cubin,1024,"_cubin_XXXXXX");
-    int fd2=mkstemp(fname_cubin);
-    close(fd2);
-
-    printf("GPGPU-Sim PTX: extracting embedded cubin to temporary file \"%s\"\n", fname_cubin);
-    FILE *cubinfile = fopen(fname_cubin,"w");
-    fprintf(cubinfile,"%s",cubin_str);
-    fclose(cubinfile);
+    char fname_ptxplus[1024];
+    snprintf(fname_ptxplus,1024,"_ptxplus_XXXXXX");
+    int fd4=mkstemp(fname_ptxplus);
+    close(fd4);
 
-    // Run decuda
+    // Extract cubin to a file
     char fname_decuda[1024];
-    snprintf(fname_decuda,1024,"_decuda_XXXXXX");
-    int fd3=mkstemp(fname_decuda);
-    close(fd3);
+    char fname_cubin[1024];
+    if( !cuobjdump){
 
-    char decuda_commandline[1024];
-    snprintf(decuda_commandline,1024,"$DECUDA_INSTALL_PATH/decuda.py -o %s %s", fname_decuda, fname_cubin);
+		snprintf(fname_cubin,1024,"_cubin_XXXXXX");
+		int fd2=mkstemp(fname_cubin);
+		close(fd2);
 
-    printf("GPGPU-Sim PTX: calling decuda on cubin file, decuda output file = \"%s\"\n", fname_decuda);
-    int decuda_result = system(decuda_commandline);
-	if( decuda_result != 0 ) {
-	   printf("GPGPU-Sim PTX: ERROR ** while calling decuda (b) %d\n", decuda_result);
-	   printf("               Ensure env variable DECUDA_INSTALL_PATH is set and points to decuda base directory.\n");
-	   exit(1);
-	}
+		printf("GPGPU-Sim PTX: extracting embedded cubin to temporary file \"%s\"\n", fname_cubin);
+		FILE *cubinfile = fopen(fname_cubin,"w");
+		fprintf(cubinfile,"%s",cubin_str);
+		fclose(cubinfile);
 
-	// Run decuda_to_ptxplus
-    char fname_ptxplus[1024];
-    snprintf(fname_ptxplus,1024,"_ptxplus_XXXXXX");
-    int fd4=mkstemp(fname_ptxplus);
-    close(fd4);
 
-    char d2pp_commandline[1024];
-    snprintf(d2pp_commandline,1024,"$D2PP_INSTALL_PATH/decuda_to_ptxplus %s %s %s %s > /dev/null", fname_decuda, fname_ptx, fname_cubin, fname_ptxplus);
+		snprintf(fname_decuda,1024,"_decuda_XXXXXX");
+		int fd3=mkstemp(fname_decuda);
+		close(fd3);
 
-    printf("GPGPU-Sim PTX: calling decuda_to_ptxplus, ptxplus output file = \"%s\"\n", fname_ptxplus);
-    int d2pp_result = system(d2pp_commandline);
-	if( d2pp_result != 0 ) {
-	   printf("GPGPU-Sim PTX: ERROR ** while calling decuda_to_ptxplus %d\n", d2pp_result);
-	   printf("               Ensure env variable D2PP_INSTALL_PATH is set and points to decuda_to_ptxplus base directory.\n");
-	   exit(1);
+		// Run decuda
+		char decuda_commandline[1024];
+		snprintf(decuda_commandline,1024,"$DECUDA_INSTALL_PATH/decuda.py -o %s %s", fname_decuda, fname_cubin);
+		printf("GPGPU-Sim PTX: calling decuda on cubin file, decuda output file = \"%s\"\n", fname_decuda);
+		int decuda_result = system(decuda_commandline);
+		if( decuda_result != 0 ) {
+			printf("GPGPU-Sim PTX: ERROR ** while calling decuda (b) %d\n", decuda_result);
+			printf("               Ensure env variable DECUDA_INSTALL_PATH is set and points to decuda base directory.\n");
+			exit(1);
+		}
+		// Run decuda_to_ptxplus
+		char d2pp_commandline[1024];
+		snprintf(d2pp_commandline,1024,"$D2PP_INSTALL_PATH/decuda_to_ptxplus %s %s %s %s > /dev/null", fname_decuda, fname_ptx, fname_cubin, fname_ptxplus);
+		printf("GPGPU-Sim PTX: calling decuda_to_ptxplus, ptxplus output file = \"%s\"\n", fname_ptxplus);
+		int d2pp_result = system(d2pp_commandline);
+		if( d2pp_result != 0 ) {
+			printf("GPGPU-Sim PTX: ERROR ** while calling decuda_to_ptxplus %d\n", d2pp_result);
+			printf("               Ensure env variable D2PP_INSTALL_PATH is set and points to decuda_to_ptxplus base directory.\n");
+			exit(1);
+		}
+    }
+    else {
+		// Run cuobjdump_to_ptxplus
+		char commandline[1024];
+		int result;
+		snprintf(commandline, 1024, "$GPGPUSIM_ROOT/cuobjdump_to_ptxplus/cuobjdump_to_ptxplus /proc/%d/exe %s", getpid(), fname_ptxplus);
+		fflush(stdout);
+		printf("GPGPU-Sim PTX: calling cuobjdump_to_ptxplus on exe file, output file = \"%s\"\n", fname_ptxplus);
+		result = system(commandline);
+		if(result){printf("GPGPU-Sim PTX: ERROR ** could not execute %s\n", commandline); exit(1);}
 	}
 
 	// Get ptxplus from file
@@ -173,7 +186,12 @@ char* gpgpu_ptx_sim_convert_ptx_to_ptxplus(const char *ptx_str, const char *cubi
 
 	// Remove temporary files
 	char rm_commandline[1024];
-	snprintf(rm_commandline,1024,"rm -f %s %s %s %s", fname_ptx, fname_cubin, fname_decuda, fname_ptxplus);
+	if (!cuobjdump) {
+		snprintf(rm_commandline,1024,"rm -f %s %s %s %s", fname_ptx, fname_cubin, fname_decuda, fname_ptxplus);
+	}
+	else {
+		snprintf(rm_commandline,1024,"rm -f %s %s", fname_ptx, fname_ptxplus);
+	}
 	printf("GPGPU-Sim PTX: removing temporary files using \"%s\"\n", rm_commandline);
 	int rm_result = system(rm_commandline);
 	if( rm_result != 0 ) {
@@ -186,6 +204,92 @@ char* gpgpu_ptx_sim_convert_ptx_to_ptxplus(const char *ptx_str, const char *cubi
 	return ptxplus_str;
 }
 
+char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const char *ptxfilename, const char *elffilename, const char *sassfilename)
+{
+
+	printf("GPGPU-Sim PTX: converting EMBEDDED .ptx file to ptxplus \n");
+
+/*
+	// Extract ptx to a file
+    char fname_ptx[1024];
+    snprintf(fname_ptx,1024,"_ptx_XXXXXX");
+    int fd1=mkstemp(fname_ptx);
+    close(fd1);
+
+    printf("GPGPU-Sim PTX: extracting embedded .ptx to temporary file \"%s\"\n", fname_ptx);
+    FILE *ptxfile = fopen(fname_ptx,"w");
+    fprintf(ptxfile,"%s",ptx_str);
+    fclose(ptxfile);
+
+    char fname_sass[1024];
+    snprintf(fname_sass,1024,"_sass_XXXXXX");
+    int fd2=mkstemp(fname_sass);
+    close(fd2);
+
+    printf("GPGPU-Sim PTX: extracting .sass to temporary file \"%s\"\n", fname_sass);
+    FILE *sassfile = fopen(fname_sass,"w");
+    fprintf(sassfile,"%s",sass_str);
+    fclose(sassfile);
+
+    char fname_elf[1024];
+    snprintf(fname_elf,1024,"_elf_XXXXXX");
+    int fd3=mkstemp(fname_elf);
+    close(fd3);
+
+    printf("GPGPU-Sim PTX: extracting .elf to temporary file \"%s\"\n", fname_elf);
+    FILE *elffile = fopen(fname_elf,"w");
+    fprintf(elffile,"%s",elf_str);
+    fclose(elffile);
+*/
+    char fname_ptxplus[1024];
+    snprintf(fname_ptxplus,1024,"_ptxplus_XXXXXX");
+    int fd4=mkstemp(fname_ptxplus);
+    close(fd4);
+
+    // Extract cubin to a file
+    char fname_decuda[1024];
+    char fname_cubin[1024];
+
+
+	// Run cuobjdump_to_ptxplus
+	char commandline[1024];
+	int result;
+	snprintf(commandline, 1024, "$GPGPUSIM_ROOT/cuobjdump_to_ptxplus/cuobjdump_to_ptxplus %s %s %s %s", ptxfilename, sassfilename, elffilename, fname_ptxplus);
+	fflush(stdout);
+	printf("GPGPU-Sim PTX: calling cuobjdump_to_ptxplus\ncommandline: %s\n", commandline);
+	result = system(commandline);
+	if(result){printf("GPGPU-Sim PTX: ERROR ** could not execute %s\n", commandline); exit(1);}
+
+
+	// Get ptxplus from file
+	std::ifstream fileStream(fname_ptxplus, std::ios::in);
+	std::string text, line;
+	while(getline(fileStream,line)) {
+		text += (line + "\n");
+	}
+	fileStream.close();
+
+	char* ptxplus_str = new char [strlen(text.c_str())+1];
+	strcpy(ptxplus_str, text.c_str());
+
+	// Remove temporary files
+	/*
+	char rm_commandline[1024];
+
+	snprintf(rm_commandline,1024,"rm -f %s", fname_ptxplus);
+
+	printf("GPGPU-Sim PTX: removing temporary files using \"%s\"\n", rm_commandline);
+	int rm_result = system(rm_commandline);
+	if( rm_result != 0 ) {
+	   printf("GPGPU-Sim PTX: ERROR ** while removing temporary files %d\n", rm_result);
+	   exit(1);
+	}
+	*/
+	printf("GPGPU-Sim PTX: DONE converting EMBEDDED .ptx file to ptxplus \n");
+
+	return ptxplus_str;
+}
+
 
 symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num )
 {
@@ -237,6 +341,7 @@ void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num
     close(fd);
     char commandline2[4096];
     snprintf(commandline2,4096,"cat %s | sed 's/.version 1.5/.version 1.4/' | sed 's/, texmode_independent//' | sed 's/\\(\\.extern \\.const\\[1\\] .b8 \\w\\+\\)\\[\\]/\\1\\[1\\]/' | sed 's/const\\[.\\]/const\\[0\\]/g' > %s", fname, fname2);
+    printf("Running: %s\n", commandline2);
     int result = system(commandline2);
     if( result != 0 ) {
        printf("GPGPU-Sim PTX: ERROR ** while loading PTX (a) %d\n", result);
@@ -252,10 +357,10 @@ void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num
     extra_flags[0]=0;
 
 #if CUDART_VERSION >= 3000
-    snprintf(extra_flags,1024,"--gpu-name=sm_20");
+    snprintf(extra_flags,1024,"--gpu-name=sm_13");
 #endif
 
-    snprintf(commandline,1024,"ptxas %s -v %s --output-file /dev/null 2> %s", 
+    snprintf(commandline,1024,"ptxas %s -v %s --output-file  /dev/null 2> %s",
              extra_flags, fname2, tempfile_ptxinfo);
     printf("GPGPU-Sim PTX: generating ptxinfo using \"%s\"\n", commandline);
     result = system(commandline);
diff --git a/src/cuda-sim/ptx_loader.h b/src/cuda-sim/ptx_loader.h
index 0266755..aa98a91 100644
--- a/src/cuda-sim/ptx_loader.h
+++ b/src/cuda-sim/ptx_loader.h
@@ -32,6 +32,7 @@ extern bool g_override_embedded_ptx;
  
 class symbol_table *gpgpu_ptx_sim_load_ptx_from_string( const char *p, unsigned source_num );
 void gpgpu_ptxinfo_load_from_string( const char *p_for_info, unsigned source_num );
-char* gpgpu_ptx_sim_convert_ptx_to_ptxplus(const char *ptx_str, const char *cubin_str, unsigned source_num, bool save_converted );
+char* gpgpu_ptx_sim_convert_ptx_to_ptxplus(const char *ptx_str, const char *cubin_str, unsigned source_num, bool save_converted, bool cuobjdump );
+char* gpgpu_ptx_sim_convert_ptx_and_sass_to_ptxplus(const char *ptx_str, const char *sass_str, const char *elf_str);
 
 #endif
diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc
index 3db984d..308405a 100644
--- a/src/cuda-sim/ptx_parser.cc
+++ b/src/cuda-sim/ptx_parser.cc
@@ -43,6 +43,7 @@ const char *g_filename;
 unsigned g_max_regs_per_thread = 0;
 
 // the program intermediate representation...
+static symbol_table *g_global_allfiles_symbol_table = NULL;
 static symbol_table *g_global_symbol_table = NULL;
 std::map<std::string,symbol_table*> g_sym_name_to_symbol_table;
 static symbol_table *g_current_symbol_table = NULL;
@@ -105,7 +106,13 @@ void read_parser_environment_variables()
 symbol_table *init_parser( const char *ptx_filename )
 {
    g_filename = strdup(ptx_filename);
-   g_global_symbol_table = g_current_symbol_table = new symbol_table("global",0,NULL);
+   if  (g_global_allfiles_symbol_table == NULL) {
+	   g_global_allfiles_symbol_table = new symbol_table("global_allfiles", 0, NULL);
+	   g_global_symbol_table = g_current_symbol_table = g_global_allfiles_symbol_table;
+   }
+   else {
+	   g_global_symbol_table = g_current_symbol_table = new symbol_table("global",0,g_global_allfiles_symbol_table);
+   }
    ptx_lineno = 1;
 
 #define DEF(X,Y) g_ptx_token_decode[X] = Y;
@@ -767,7 +774,7 @@ void add_scalar_operand( const char *identifier )
    DPRINTF("add_scalar_operand");
    const symbol *s = g_current_symbol_table->lookup(identifier);
    if ( s == NULL ) {
-      if ( g_opcode == BRA_OP || g_opcode == CALLP_OP || g_opcode == BREAKADDR_OP) {
+      if ( g_opcode == BRA_OP || g_opcode == CALLP_OP) {
          // forward branch target...
          s = g_current_symbol_table->add_variable(identifier,NULL,0,g_filename,ptx_lineno);
       } else {