diff options
| -rw-r--r-- | libcuda/cuda_runtime_api.cc | 1 | ||||
| -rw-r--r-- | src/cuda-sim/cuda-sim.cc | 7 | ||||
| -rw-r--r-- | src/cuda-sim/ptx.l | 5 | ||||
| -rw-r--r-- | src/cuda-sim/ptx_ir.h | 1 |
4 files changed, 11 insertions, 3 deletions
diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index d6ce4e0..ef46f00 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -1923,6 +1923,7 @@ void cuobjdumpParseBinary(unsigned int handle){ if (capability > max_capability) max_capability = capability; } if (max_capability > 20) printf("WARNING: No guarantee that PTX will be parsed for SM version %u\n", max_capability); + if (max_capability == 0) max_capability=context->get_device()->get_gpgpu()->get_config().get_forced_max_capability(); cuobjdumpPTXSection* ptx = NULL; const char* pre_load = getenv("CUOBJDUMP_SIM_FILE"); diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index dce35ca..2c87031 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -1139,8 +1139,13 @@ void function_info::finalize( memory_space *param_mem ) } // copy the parameter over word-by-word so that parameter that crosses a memory page can be copied over //Jin: copy parameter using aligned rules + const type_info *paramtype = param->type(); + int align_amount = paramtype->get_key().get_alignment_spec(); + align_amount = (align_amount == -1) ? size : align_amount; + param_address = (param_address + align_amount - 1) / align_amount * align_amount; //aligned + const size_t word_size = 4; - param_address = (param_address + size - 1) / size * size; //aligned with size + //param_address = (param_address + size - 1) / size * size; //aligned with size for (size_t idx = 0; idx < size; idx += word_size) { const char *pdata = reinterpret_cast<const char*>(param_value.pdata) + idx; // cast to char * for ptr arithmetic param_mem->write(param_address + idx, word_size, pdata,NULL,NULL); diff --git a/src/cuda-sim/ptx.l b/src/cuda-sim/ptx.l index 1b5d7f6..908c5be 100644 --- a/src/cuda-sim/ptx.l +++ b/src/cuda-sim/ptx.l @@ -36,7 +36,8 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #include "ptx.tab.h" #include <string.h> -char linebuf[1024]; +#define LINEBUF_SIZE (64*1024) +char linebuf[LINEBUF_SIZE]; unsigned col = 0; #define TC col+=strlen(ptx_text); #define CHECK_UNSIGNED \ @@ -384,7 +385,7 @@ breakaddr TC; ptx_lval.int_value = BREAKADDR_OP; return OPCODE; "//"[^\n]* TC; // eat single -\n.* col=0; strncpy(linebuf, yytext + 1, 1024); yyless( 1 ); +\n.* col=0; strncpy(linebuf, yytext + 1, LINEBUF_SIZE); yyless( 1 ); " " TC; "\t" TC; diff --git a/src/cuda-sim/ptx_ir.h b/src/cuda-sim/ptx_ir.h index 85b2a3b..6731763 100644 --- a/src/cuda-sim/ptx_ir.h +++ b/src/cuda-sim/ptx_ir.h @@ -91,6 +91,7 @@ public: bool is_tex() const { return m_space_spec == tex_space;} bool is_func_addr() const { return m_is_function?true:false; } int scalar_type() const { return m_scalar_type_spec;} + int get_alignment_spec() const { return m_alignment_spec;} unsigned type_decode( size_t &size, int &t ) const; static unsigned type_decode( int type, size_t &size, int &t ); memory_space_t get_memory_space() const { return m_space_spec; } |