diff options
| author | aamir <[email protected]> | 2018-10-24 21:18:27 -0700 |
|---|---|---|
| committer | aamir <[email protected]> | 2018-10-24 21:18:27 -0700 |
| commit | 7c441c450e40bf07bdf1acfe1eb2258952e1f7b7 (patch) | |
| tree | ff4f3ab1e5d42c8284d419209afb2ccc5f25603b /src/cuda-sim/ptx_ir.cc | |
| parent | 68134d5eb326552fc1ef4b02b2eb21103266283b (diff) | |
| parent | 09e6092ace5213a5d5a49bf80b052802c06a4268 (diff) | |
merged tensor-cores code
Diffstat (limited to 'src/cuda-sim/ptx_ir.cc')
| -rw-r--r-- | src/cuda-sim/ptx_ir.cc | 55 |
1 files changed, 45 insertions, 10 deletions
diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index 8ebdcf8..d12c741 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -995,7 +995,7 @@ static std::list<operand_info> check_operands( int opcode, const std::list<operand_info> &operands ) { static int g_warn_literal_operands_two_type_inst; - if( (opcode == CVT_OP) || (opcode == SET_OP) || (opcode == SLCT_OP) || (opcode == TEX_OP) ) { + if( (opcode == CVT_OP) || (opcode == SET_OP) || (opcode == SLCT_OP) || (opcode == TEX_OP) || (opcode==MMA_OP)) { // just make sure these do not have have const operands... if( !g_warn_literal_operands_two_type_inst ) { std::list<operand_info>::const_iterator o; @@ -1043,6 +1043,7 @@ ptx_instruction::ptx_instruction( int opcode, const std::list<operand_info> &operands, const operand_info &return_var, const std::list<int> &options, + const std::list<int> &wmma_options, const std::list<int> &scalar_type, memory_space_t space_spec, const char *file, @@ -1061,6 +1062,7 @@ ptx_instruction::ptx_instruction( int opcode, m_operands.insert(m_operands.begin(), checked_operands.begin(), checked_operands.end() ); m_return_var = return_var; m_options = options; + m_wmma_options = wmma_options; m_wide = false; m_hi = false; m_lo = false; @@ -1078,9 +1080,33 @@ ptx_instruction::ptx_instruction( int opcode, m_atomic_spec = 0; m_membar_level = 0; m_inst_size = 8; // bytes - + int rr=0; std::list<int>::const_iterator i; unsigned n=1; + for ( i=wmma_options.begin(); i!= wmma_options.end(); i++, n++ ) { + int last_ptx_inst_option = *i; + switch ( last_ptx_inst_option ) { + case SYNC_OPTION: + case LOAD_A: + case LOAD_B: + case LOAD_C: + case STORE_D: + case MMA: + m_wmma_type=last_ptx_inst_option; + break; + case ROW: + case COL: + m_wmma_layout[rr++]=last_ptx_inst_option; + break; + case M16N16K16: + break; + default: + assert(0); + break; + } + } + rr=0; + n=1; for ( i=options.begin(); i!= options.end(); i++, n++ ) { int last_ptx_inst_option = *i; switch ( last_ptx_inst_option ) { @@ -1207,16 +1233,25 @@ ptx_instruction::ptx_instruction( int opcode, case HALF_OPTION: m_inst_size = 4; // bytes break; - case EXTP_OPTION: - break; - case NC_OPTION: - break; - case UP_OPTION: - case DOWN_OPTION: - case BFLY_OPTION: - case IDX_OPTION: + case EXTP_OPTION: + break; + case NC_OPTION: + break; + case UP_OPTION: + case DOWN_OPTION: + case BFLY_OPTION: + case IDX_OPTION: m_shfl_op = last_ptx_inst_option; break; + + case PRMT_F4E_MODE: + case PRMT_B4E_MODE: + case PRMT_RC8_MODE: + case PRMT_ECL_MODE: + case PRMT_ECR_MODE: + case PRMT_RC16_MODE: + m_prmt_op = last_ptx_inst_option; + break; default: assert(0); break; |