1 files changed, 54 insertions, 5 deletions

diff --git a/src/cuda-sim/ptx_parser.cc b/src/cuda-sim/ptx_parser.cc
index 0ba9a5b..25758dd 100644
--- a/src/cuda-sim/ptx_parser.cc
+++ b/src/cuda-sim/ptx_parser.cc
@@ -75,6 +75,7 @@ symbol *g_label;
 int g_opcode = -1;
 std::list<operand_info> g_operands;
 std::list<int> g_options;
+std::list<int> g_wmma_options;
 std::list<int> g_scalar_type;
 
 #define PTX_PARSE_DPRINTF(...) \
@@ -133,6 +134,7 @@ void init_instruction_state()
    g_label = NULL;
    g_opcode = -1;
    g_options.clear();
+   g_wmma_options.clear();
    g_return_var = operand_info();
    init_directive_state();
 }
@@ -144,7 +146,9 @@ symbol_table *init_parser( const char *ptx_filename )
        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;
@@ -273,7 +277,7 @@ void parse_assert_impl( int test_value, const char *file, unsigned line, const c
       parse_error_impl(file,line, msg);
 }
 
-extern char linebuf[1024];
+extern char linebuf[4096];
 
 
 void set_return()
@@ -308,6 +312,7 @@ void add_instruction()
                                              g_operands,
                                              g_return_var,
                                              g_options, 
+                                             g_wmma_options, 
                                              g_scalar_type,
                                              g_space_spec,
                                              g_filename,
@@ -451,6 +456,20 @@ void add_identifier( const char *identifier, int array_dim, unsigned array_ident
       g_last_symbol->set_address( addr+addr_pad );
       g_current_symbol_table->alloc_shared( num_bits/8 + addr_pad );
       break;
+   case sstarr_space:
+         printf("GPGPU-Sim PTX: allocating sstarr region for \"%s\" ",
+                identifier);
+         fflush(stdout);
+         assert( (num_bits%8) == 0  );
+         addr = g_current_symbol_table->get_sstarr_next();
+         addr_pad = pad_address(addr, num_bits/8, 128);
+         printf("from 0x%x to 0x%lx (sstarr memory space)\n",
+                 addr+addr_pad,
+                 addr+addr_pad + num_bits/8);
+            fflush(stdout);
+         g_last_symbol->set_address( addr+addr_pad );
+         g_current_symbol_table->alloc_sstarr( num_bits/8 + addr_pad );
+         break;
    case const_space:
       if( array_ident == ARRAY_IDENTIFIER_NO_DIM ) {
          printf("GPGPU-Sim PTX: deferring allocation of constant region for \"%s\" (need size information)\n", identifier );
@@ -657,8 +676,8 @@ void add_scalar_type_spec( int type_spec )
    g_scalar_type.push_back( type_spec );
    if ( g_scalar_type.size() > 1 ) {
       parse_assert( (g_opcode == -1) || (g_opcode == CVT_OP) || (g_opcode == SET_OP) || (g_opcode == SLCT_OP)
-         || (g_opcode == TEX_OP)|| (g_opcode == DP4A_OP), 
-                    "only cvt, set, slct, tex, and dp4a can have more than one type specifier.");
+                    || (g_opcode == TEX_OP)|| (g_opcode==MMA_OP)|| (g_opcode == DP4A_OP),
+                    "only cvt, set, slct, tex and dp4a can have more than one type specifier.");
    }
    g_scalar_type_spec = type_spec;
 }
@@ -697,7 +716,11 @@ void add_option( int option )
    PTX_PARSE_DPRINTF("add_option");
    g_options.push_back( option );
 }
-
+void add_wmma_option( int option ) 
+{
+   PTX_PARSE_DPRINTF("add_option");
+   g_wmma_options.push_back( option );
+}
 void add_double_operand( const char *d1, const char *d2 )
 {
    //operands that access two variables.
@@ -753,6 +776,28 @@ void add_4vector_operand( const char *d1, const char *d2, const char *d3, const
    if ( s4 == null_op ) s4 = NULL;
    g_operands.push_back( operand_info(s1,s2,s3,s4) );
 }
+void add_8vector_operand( const char *d1, const char *d2, const char *d3, const char *d4,const char *d5,const char *d6,const char *d7,const char *d8 ) 
+{
+   PTX_PARSE_DPRINTF("add_8vector_operand");
+   const symbol *s1 = g_current_symbol_table->lookup(d1);
+   const symbol *s2 = g_current_symbol_table->lookup(d2);
+   const symbol *s3 = g_current_symbol_table->lookup(d3);
+   const symbol *s4 = g_current_symbol_table->lookup(d4);
+   const symbol *s5 = g_current_symbol_table->lookup(d5);
+   const symbol *s6 = g_current_symbol_table->lookup(d6);
+   const symbol *s7 = g_current_symbol_table->lookup(d7);
+   const symbol *s8 = g_current_symbol_table->lookup(d8);
+   parse_assert( s1 != NULL && s2 != NULL && s3 != NULL && s4 != NULL && s5 !=NULL && s6 !=NULL && s7 !=NULL && s8 !=NULL, "v4 component(s) missing declarations.");
+   const symbol *null_op = g_current_symbol_table->lookup("_");
+   if ( s2 == null_op ) s2 = NULL;
+   if ( s3 == null_op ) s3 = NULL;
+   if ( s4 == null_op ) s4 = NULL;
+   if ( s5 == null_op ) s5 = NULL;
+   if ( s6 == null_op ) s6 = NULL;
+   if ( s7 == null_op ) s7 = NULL;
+   if ( s8 == null_op ) s8 = NULL;
+   g_operands.push_back( operand_info(s1,s2,s3,s4,s5,s6,s7,s8) );
+}
 
 void add_builtin_operand( int builtin, int dim_modifier ) 
 {
@@ -1011,6 +1056,10 @@ void target_header3(char* a, char* b, char* c)
    g_global_symbol_table->set_sm_target(a,b,c);
 }
 
+void maxnt_id(int x, int y, int z) {
+  g_func_info->set_maxnt_id(x * y * z);
+}
+
 void func_header(const char* a) {} //intentional dummy function
 void func_header_info(const char* a) {} //intentional dummy function
 void func_header_info_int(const char* a, int b) {} //intentional dummy function