- reduce 3.1 sdk regression test runtimes (as per 2.3 sdk versions)

- add ldu support (just use ld for now)
- get rid of asserts for .const[n] 
- status:
	cuda sdk 3.1 dct8x8, eigenvalues, MersenneTwister, reduction, scan still running (too slow)
	ocl sdk 3.1 oclQuasirandomGenerator still running (too slow)
	ocl sdk 3.1 oclReduction fail comparison (cuda 3.1)
	ocl sdk 3.1 oclMatrixMul exits mid computation
	ocl sdk 3.1 oclTranspose hits "clCreateBuffer - buffer already created for this host variable"
	ocl sdk 3.1 oclSortingNetworks fails w/ message about min workgroup size
	remaining 30 tests pass

[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 6927]

6 files changed, 33 insertions, 7 deletions

diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc
index 2a41e4b..bf90938 100644
--- a/src/cuda-sim/cuda-sim.cc
+++ b/src/cuda-sim/cuda-sim.cc
@@ -555,7 +555,7 @@ void gpgpu_ptx_sim_init_memory()
    static bool initialized = false;
    if ( !initialized ) {
       g_global_mem = new memory_space_impl<8192>("global",64*1024);
-      g_param_mem = new memory_space_impl<64>("param",64*1024);
+      g_param_mem = new memory_space_impl<8192>("param",64*1024);
       g_tex_mem = new memory_space_impl<8192>("tex",64*1024);
       g_surf_mem = new memory_space_impl<8192>("surf",64*1024);
       initialized = true;
@@ -992,12 +992,14 @@ void function_info::add_param_name_type_size( unsigned index, std::string name,
 void function_info::add_param_data( unsigned argn, struct gpgpu_ptx_sim_arg *args )
 {
    const void *data = args->m_start;
-   param_t tmp;
 
    if( data ) {
+      param_t tmp;
+
       tmp.pdata = args->m_start;
       tmp.size = args->m_nbytes;
       tmp.offset = args->m_offset;
+      tmp.type = 0;
       std::map<unsigned,param_info>::iterator i=m_ptx_kernel_param_info.find(argn);
       if( i != m_ptx_kernel_param_info.end()) {
          i->second.add_data(tmp);
@@ -1139,6 +1141,8 @@ void function_info::ptx_exec_inst( ptx_thread_info *thread,
    unsigned index = pc - m_start_PC;
    assert( index < m_instr_mem_size );
    ptx_instruction *pI = m_instr_mem[index];
+   try {
+
    g_current_symbol_table = thread->get_finfo()->get_symtab();
    thread->clearRPC();
    thread->m_last_set_operand_value.u64 = 0;
@@ -1248,6 +1252,12 @@ void function_info::ptx_exec_inst( ptx_thread_info *thread,
    *space = insn_space;
    *addr = insn_memaddr;
    *data_size = insn_data_size;
+
+   } catch ( int x  ) {
+      printf("GPGPU-Sim PTX: ERROR (%d) executing intruction (%s:%u)\n", x, pI->source_file(), pI->source_line() );
+      printf("GPGPU-Sim PTX:       '%s'\n", pI->get_source() );
+      abort();
+   }
 }
 
 unsigned g_gx, g_gy, g_gz;
diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc
index a19f1a8..3ce0a2f 100644
--- a/src/cuda-sim/instructions.cc
+++ b/src/cuda-sim/instructions.cc
@@ -1452,7 +1452,7 @@ void decode_space( memory_space_t &space, const ptx_thread_info *thread, const o
    case surf_space:   mem = g_surf_mem; break; 
    case param_space_kernel:  mem = g_param_mem; break;
    case shared_space:  mem = thread->m_shared_mem; break; 
-   case const_space:  assert(space.get_bank()==0); mem = g_global_mem; break;
+   case const_space:  mem = g_global_mem; break;
    case generic_space:
       if( thread->get_ptx_version().ver() >= 2.0 ) {
          // convert generic address to memory space address
@@ -1474,7 +1474,7 @@ void decode_space( memory_space_t &space, const ptx_thread_info *thread, const o
    }
 }
 
-void ld_impl( const ptx_instruction *pI, ptx_thread_info *thread ) 
+void ld_exec( const ptx_instruction *pI, ptx_thread_info *thread ) 
 { 
    const operand_info &dst = pI->dst();
    const operand_info &src1 = pI->src1();
@@ -1515,7 +1515,14 @@ void ld_impl( const ptx_instruction *pI, ptx_thread_info *thread )
    thread->m_last_memory_space = space; 
 }
 
-void ldu_impl( const ptx_instruction *pI, ptx_thread_info *thread ) { inst_not_implemented(pI); }
+void ld_impl( const ptx_instruction *pI, ptx_thread_info *thread ) 
+{
+   ld_exec(pI,thread);
+}
+void ldu_impl( const ptx_instruction *pI, ptx_thread_info *thread ) 
+{ 
+   ld_exec(pI,thread);
+}
 
 void lg2_impl( const ptx_instruction *pI, ptx_thread_info *thread ) 
 { 
diff --git a/src/cuda-sim/memory.cc b/src/cuda-sim/memory.cc
index 754fe7f..09c9e81 100644
--- a/src/cuda-sim/memory.cc
+++ b/src/cuda-sim/memory.cc
@@ -94,7 +94,13 @@ template<unsigned BSIZE> void memory_space_impl<BSIZE>::read( mem_addr_t addr, s
    unsigned offset = addr & (BSIZE-1);
    unsigned nbytes = length;
    typename map_t::const_iterator i = m_data.find(index);
-   assert( (addr+length) <= (index+1)*BSIZE );
+   if( (addr+length) > (index+1)*BSIZE ) {
+      printf("GPGPU-Sim PTX: ERROR * access to memory \'%s\' is unaligned : addr=0x%x, length=%zu\n",
+             m_name.c_str(), addr, length);
+      printf("GPGPU-Sim PTX: (addr+length)=0x%lx > 0x%x=(index+1)*BSIZE, index=0x%x, BSIZE=0x%x\n",
+             (addr+length),(index+1)*BSIZE, index, BSIZE);
+      throw 1;
+   }
    if( i == m_data.end() ) {
       for( size_t n=0; n < length; n++ ) 
          ((unsigned char*)data)[n] = (unsigned char) 0;
diff --git a/src/cuda-sim/opcodes.def b/src/cuda-sim/opcodes.def
index 143bb93..4566468 100644
--- a/src/cuda-sim/opcodes.def
+++ b/src/cuda-sim/opcodes.def
@@ -95,6 +95,7 @@ OP_DEF(EXIT_OP,exit_impl,"exit",1,3)
 OP_DEF(FMA_OP,fma_impl,"fma",1,2)
 OP_DEF(ISSPACEP_OP,isspacep_impl,"isspacep",1,1)
 OP_DEF(LD_OP,ld_impl,"ld",1,5)
+OP_DEF(LDU_OP,ldu_impl,"ldu",1,5)
 OP_DEF(LG2_OP,lg2_impl,"lg2",1,4)
 OP_DEF(MAD24_OP,mad24_impl,"mad24",1,2)
 OP_DEF(MAD_OP,mad_impl,"mad",1,2)
diff --git a/src/cuda-sim/ptx.l b/src/cuda-sim/ptx.l
index 3e98a75..3b83e44 100644
--- a/src/cuda-sim/ptx.l
+++ b/src/cuda-sim/ptx.l
@@ -109,6 +109,7 @@ fma     TC; ptx_lval.int_value = FMA_OP; return OPCODE;
 isspacep TC; ptx_lval.int_value = ISSPACEP_OP; return OPCODE;
 ld      TC; ptx_lval.int_value = LD_OP; return OPCODE;
 ld.volatile TC; ptx_lval.int_value = LD_OP; return OPCODE;
+ldu     TC; ptx_lval.int_value = LDU_OP; return OPCODE;
 lg2	TC; ptx_lval.int_value = LG2_OP; return OPCODE;
 mad24   TC; ptx_lval.int_value = MAD24_OP; return OPCODE;
 mad     TC; ptx_lval.int_value = MAD_OP; return OPCODE;
diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc
index fa87ad9..cfa6650 100644
--- a/src/cuda-sim/ptx_ir.cc
+++ b/src/cuda-sim/ptx_ir.cc
@@ -419,7 +419,6 @@ void add_identifier( const char *identifier, int array_dim, unsigned array_ident
       g_current_symbol_table->alloc_shared( num_bits/8 + addr_pad );
       break;
    case const_space:
-      assert(ti.get_memory_space().get_bank()==0);
       if( array_ident == ARRAY_IDENTIFIER_NO_DIM ) {
          printf("GPGPU-Sim PTX: deferring allocation of constant region for \"%s\" (need size information)\n", identifier );
       } else {
@@ -1363,6 +1362,8 @@ unsigned type_info_key::type_decode( int type, size_t &size, int &basic_type )
    case B16_TYPE: size=16; basic_type=0; return 13;
    case B32_TYPE: size=32; basic_type=0; return 14;
    case B64_TYPE: size=64; basic_type=0; return 15;
+   case TEXREF_TYPE: case SAMPLERREF_TYPE: case SURFREF_TYPE:
+      size=32; basic_type=3; return 16;
    default: 
       printf("ERROR ** type_decode() does not know about \"%s\"\n", g_ptx_token_decode[type].c_str() ); 
       assert(0);