diff options
| author | Aamir Raihan <[email protected]> | 2018-06-08 13:56:54 -0700 |
|---|---|---|
| committer | Aamir Raihan <[email protected]> | 2018-06-08 13:56:54 -0700 |
| commit | 774e28201d239433357db4cbd00c0a62273773cc (patch) | |
| tree | 2b5a8f5cc1108009c4f8edcb1fc09232773c31ae /src | |
| parent | 3e167c76f73822f6ff044d4ecd769aab6e199be1 (diff) | |
regression passed
Diffstat (limited to 'src')
| -rw-r--r-- | src/Makefile | 2 | ||||
| -rw-r--r-- | src/cuda-sim/Makefile | 2 | ||||
| -rw-r--r-- | src/cuda-sim/instructions.cc | 11 | ||||
| -rw-r--r-- | src/cuda-sim/ptx.l | 4 | ||||
| -rw-r--r-- | src/gpgpu-sim/Makefile | 2 | ||||
| -rw-r--r-- | src/intersim2/Makefile | 2 |
6 files changed, 6 insertions, 17 deletions
diff --git a/src/Makefile b/src/Makefile index 09194f3..6001669 100644 --- a/src/Makefile +++ b/src/Makefile @@ -46,7 +46,7 @@ ifeq ($(TRACE),1) endif ifneq ($(DEBUG),1) - OPTFLAGS += -O0 + OPTFLAGS += -O3 else CXXFLAGS += endif diff --git a/src/cuda-sim/Makefile b/src/cuda-sim/Makefile index d08a696..e8977a6 100644 --- a/src/cuda-sim/Makefile +++ b/src/cuda-sim/Makefile @@ -42,7 +42,7 @@ include ../../version_detection.mk OUTPUT_DIR=$(SIM_OBJ_FILES_DIR)/cuda-sim -OPT := -O0 -g3 -Wall -Wno-unused-function -Wno-sign-compare +OPT := -O3 -g3 -Wall -Wno-unused-function -Wno-sign-compare ifeq ($(DEBUG),1) OPT := -g3 -Wall -Wno-unused-function -Wno-sign-compare endif diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc index 59415f4..92c8529 100644 --- a/src/cuda-sim/instructions.cc +++ b/src/cuda-sim/instructions.cc @@ -835,17 +835,6 @@ void add_impl( const ptx_instruction *pI, ptx_thread_info *thread ) unsigned i_type = pI->get_type(); src1_data = thread->get_operand_value(src1, dst, i_type, thread, 1); src2_data = thread->get_operand_value(src2, dst, i_type, thread, 1); - //unsigned warpId_aa,warp_size_aa; - //warpId_aa = pI->warp_id(); - //warp_size_aa=32; - //dim3 t=thread->get_tid(); - //unsigned tid_aa=warp_size_aa*warpId_aa; - - ptx_thread_info *thread2; - thread2=thread; - src1_data = thread2->get_operand_value(src1, dst, i_type, thread2, 1); - src2_data = thread2->get_operand_value(src2, dst, i_type, thread2, 1); - unsigned rounding_mode = pI->rounding_mode(); int orig_rm = fegetround(); diff --git a/src/cuda-sim/ptx.l b/src/cuda-sim/ptx.l index f67dddd..03d6838 100644 --- a/src/cuda-sim/ptx.l +++ b/src/cuda-sim/ptx.l @@ -68,7 +68,7 @@ bra TC; ptx_lval.int_value = BRA_OP; return OPCODE; brx TC; ptx_lval.int_value = BRX_OP; return OPCODE; brev TC; ptx_lval.int_value = BREV_OP; return OPCODE; brkpt TC; ptx_lval.int_value = BRKPT_OP; return OPCODE; -wmma\.mma TC; ptx_lval.int_value = MMA_OP; return OPCODE; +wmma TC; ptx_lval.int_value = MMA_OP; return OPCODE; wmma\.load TC; ptx_lval.int_value = MMA_LD_OP; return OPCODE; wmma\.store TC; ptx_lval.int_value = MMA_ST_OP; return OPCODE; @@ -160,7 +160,7 @@ breakaddr TC; ptx_lval.int_value = BREAKADDR_OP; return OPCODE; \.b\.sync TC; ptx_lval.int_value = LOAD_B; return WMMA_DIRECTIVE; \.c\.sync TC; ptx_lval.int_value = LOAD_C; return WMMA_DIRECTIVE; \.d\.sync TC; ptx_lval.int_value = STORE_D; return WMMA_DIRECTIVE; -\.sync TC;ptx_lval.int_value=MMA; return WMMA_DIRECTIVE; +\.mma\.sync TC;ptx_lval.int_value=MMA; return WMMA_DIRECTIVE; \.row TC; ptx_lval.int_value = ROW; return LAYOUT; \.col TC; ptx_lval.int_value = COL; return LAYOUT; \.m16n16k16 TC; ptx_lval.int_value = M16N16K16; return CONFIGURATION; diff --git a/src/gpgpu-sim/Makefile b/src/gpgpu-sim/Makefile index 4f77699..f10a8a4 100644 --- a/src/gpgpu-sim/Makefile +++ b/src/gpgpu-sim/Makefile @@ -48,7 +48,7 @@ ifeq ($(GNUC_CPP0X), 1) endif ifneq ($(DEBUG),1) - OPTFLAGS += -O0 + OPTFLAGS += -O3 else CXXFLAGS += endif diff --git a/src/intersim2/Makefile b/src/intersim2/Makefile index 4ef21ac..bd42000 100644 --- a/src/intersim2/Makefile +++ b/src/intersim2/Makefile @@ -44,7 +44,7 @@ endif CPPFLAGS += -Wall $(INCPATH) $(DEFINE) ifneq ($(DEBUG),1) -CPPFLAGS += -O0 +CPPFLAGS += -O3 endif CPPFLAGS += -g CPPFLAGS += -fPIC |