diff options
| author | Roland Green <[email protected]> | 2019-08-26 15:14:22 -0400 |
|---|---|---|
| committer | GitHub <[email protected]> | 2019-08-26 15:14:22 -0400 |
| commit | b88c9e9a545960b7fcc1c373b60005612296158b (patch) | |
| tree | 1e32c07e932b3d3cc6b977e29a1327274f4f29a8 /src/cuda-sim | |
| parent | 2a6788b59055b5ce694882a282af0cc6311854d4 (diff) | |
| parent | 6be23a1b36a3311c7dafc45a5e692e38cb351337 (diff) | |
Merge branch 'dev' into fix_warnings
Diffstat (limited to 'src/cuda-sim')
| -rw-r--r-- | src/cuda-sim/cuda-sim.cc | 2 | ||||
| -rw-r--r-- | src/cuda-sim/opcodes.h | 6 | ||||
| -rw-r--r-- | src/cuda-sim/ptx.l | 3 | ||||
| -rw-r--r-- | src/cuda-sim/ptx_ir.cc | 6 |
4 files changed, 13 insertions, 4 deletions
diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc index f8d0b3e..fbce75b 100644 --- a/src/cuda-sim/cuda-sim.cc +++ b/src/cuda-sim/cuda-sim.cc @@ -858,7 +858,7 @@ void ptx_instruction::set_opcode_and_latency() op=TENSOR_CORE_OP; break; case SHFL_OP: - latency = 32; + latency = 4; initiation_interval = 4; break; default: diff --git a/src/cuda-sim/opcodes.h b/src/cuda-sim/opcodes.h index b91d92f..86d3b99 100644 --- a/src/cuda-sim/opcodes.h +++ b/src/cuda-sim/opcodes.h @@ -68,6 +68,10 @@ enum wmma_type{ MMA, ROW, COL, - M16N16K16 + M16N16K16, + M32N8K16, + M8N32K16 + + }; #endif diff --git a/src/cuda-sim/ptx.l b/src/cuda-sim/ptx.l index 3a2a839..2dadda4 100644 --- a/src/cuda-sim/ptx.l +++ b/src/cuda-sim/ptx.l @@ -175,6 +175,9 @@ breakaddr TC; yylval->int_value = BREAKADDR_OP; return OPCODE; \.row TC; yylval->int_value = ROW; return LAYOUT; \.col TC; yylval->int_value = COL; return LAYOUT; \.m16n16k16 TC; yylval->int_value = M16N16K16; return CONFIGURATION; +\.m32n8k16 TC; yylval->int_value = M32N8K16; return CONFIGURATION; +\.m8n32k16 TC; yylval->int_value = M8N32K16; return CONFIGURATION; + \.f4e TC; return PRMT_F4E_MODE; \.b4e TC; return PRMT_B4E_MODE; \.rc8 TC; return PRMT_RC8_MODE; diff --git a/src/cuda-sim/ptx_ir.cc b/src/cuda-sim/ptx_ir.cc index 3384d49..6978cc1 100644 --- a/src/cuda-sim/ptx_ir.cc +++ b/src/cuda-sim/ptx_ir.cc @@ -186,8 +186,8 @@ void symbol_table::add_function( function_info *func, const char *filename, unsi //Jin: handle instruction group for cdp symbol_table* symbol_table::start_inst_group() { - char inst_group_name[1024]; - snprintf(inst_group_name, 1024, "%s_inst_group_%u", m_scope_name.c_str(), m_inst_group_id); + char inst_group_name[4096]; + snprintf(inst_group_name, 4096, "%s_inst_group_%u", m_scope_name.c_str(), m_inst_group_id); //previous added assert(m_inst_group_symtab.find(std::string(inst_group_name)) == m_inst_group_symtab.end()); @@ -1154,6 +1154,8 @@ ptx_instruction::ptx_instruction( int opcode, m_wmma_layout[rr++]=last_ptx_inst_option; break; case M16N16K16: + case M32N8K16: + case M8N32K16: break; default: assert(0); |