_Q1A(PROF_TRIGGER_5, 0x0001, B6, USER, 0x00000014, 1, 1),
_Q1A(PROF_TRIGGER_6, 0x0001, B6, USER, 0x00000018, 1, 1),
_Q1A(PROF_TRIGGER_7, 0x0001, B6, USER, 0x0000001c, 1, 1),
- _Q1A(LAUNCHED_WARPS, 0x0001, B6, LAUNCH, 0x00000004, 1, 1),
- _Q1A(LAUNCHED_THREADS, 0x003f, B6, LAUNCH, 0x398a4188, 1, 1),
- _Q1B(LAUNCHED_CTA, 0x0001, B6, WARP, 0x0000001c, 1, 1),
+ _Q1A(WARPS_LAUNCHED, 0x0001, B6, LAUNCH, 0x00000004, 1, 1),
+ _Q1A(THREADS_LAUNCHED, 0x003f, B6, LAUNCH, 0x398a4188, 1, 1),
+ _Q1B(SM_CTA_LAUNCHED, 0x0001, B6, WARP, 0x0000001c, 1, 1),
_Q1A(INST_ISSUED1, 0x0001, B6, ISSUE, 0x00000004, 1, 1),
_Q1A(INST_ISSUED2, 0x0001, B6, ISSUE, 0x00000008, 1, 1),
_Q1A(INST_ISSUED, 0x0003, B6, ISSUE, 0x00000104, 1, 1),
_Q1A(INST_EXECUTED, 0x0003, B6, EXEC, 0x00000398, 1, 1),
- _Q1A(LD_SHARED, 0x0001, B6, LDST, 0x00000000, 1, 1),
- _Q1A(ST_SHARED, 0x0001, B6, LDST, 0x00000004, 1, 1),
- _Q1A(LD_LOCAL, 0x0001, B6, LDST, 0x00000008, 1, 1),
- _Q1A(ST_LOCAL, 0x0001, B6, LDST, 0x0000000c, 1, 1),
+ _Q1A(SHARED_LD, 0x0001, B6, LDST, 0x00000000, 1, 1),
+ _Q1A(SHARED_ST, 0x0001, B6, LDST, 0x00000004, 1, 1),
+ _Q1A(LOCAL_LD, 0x0001, B6, LDST, 0x00000008, 1, 1),
+ _Q1A(LOCAL_ST, 0x0001, B6, LDST, 0x0000000c, 1, 1),
_Q1A(GLD_REQUEST, 0x0001, B6, LDST, 0x00000010, 1, 1),
_Q1A(GST_REQUEST, 0x0001, B6, LDST, 0x00000014, 1, 1),
- _Q1B(L1_LOCAL_LOAD_HIT, 0x0001, B6, L1, 0x00000000, 1, 1),
- _Q1B(L1_LOCAL_LOAD_MISS, 0x0001, B6, L1, 0x00000004, 1, 1),
- _Q1B(L1_LOCAL_STORE_HIT, 0x0001, B6, L1, 0x00000008, 1, 1),
- _Q1B(L1_LOCAL_STORE_MISS, 0x0001, B6, L1, 0x0000000c, 1, 1),
- _Q1B(L1_GLOBAL_LOAD_HIT, 0x0001, B6, L1, 0x00000010, 1, 1),
- _Q1B(L1_GLOBAL_LOAD_MISS, 0x0001, B6, L1, 0x00000014, 1, 1),
- _Q1B(GLD_TRANSACTIONS_UNCACHED, 0x0001, B6, MEM, 0x00000000, 1, 1),
+ _Q1B(L1_LOCAL_LD_HIT, 0x0001, B6, L1, 0x00000000, 1, 1),
+ _Q1B(L1_LOCAL_LD_MISS, 0x0001, B6, L1, 0x00000004, 1, 1),
+ _Q1B(L1_LOCAL_ST_HIT, 0x0001, B6, L1, 0x00000008, 1, 1),
+ _Q1B(L1_LOCAL_ST_MISS, 0x0001, B6, L1, 0x0000000c, 1, 1),
+ _Q1B(L1_GLD_HIT, 0x0001, B6, L1, 0x00000010, 1, 1),
+ _Q1B(L1_GLD_MISS, 0x0001, B6, L1, 0x00000014, 1, 1),
+ _Q1B(UNCACHED_GLD_TRANSACTIONS, 0x0001, B6, MEM, 0x00000000, 1, 1),
_Q1B(GST_TRANSACTIONS, 0x0001, B6, MEM, 0x00000004, 1, 1),
_Q1A(BRANCH, 0x0001, B6, BRANCH, 0x0000000c, 1, 1),
- _Q1A(BRANCH_DIVERGENT, 0x0001, B6, BRANCH, 0x00000010, 1, 1),
+ _Q1A(DIVERGENT_BRANCH, 0x0001, B6, BRANCH, 0x00000010, 1, 1),
_Q1B(ACTIVE_WARPS, 0x003f, B6, WARP, 0x31483104, 2, 1),
_Q1B(ACTIVE_CYCLES, 0x0001, B6, WARP, 0x00000000, 1, 1),
_Q1A(ATOM_COUNT, 0x0001, B6, BRANCH, 0x00000000, 1, 1),
_Q1A(GRED_COUNT, 0x0001, B6, BRANCH, 0x00000008, 1, 1),
- _Q1B(LD_SHARED_REPLAY, 0x0001, B6, REPLAY, 0x00000008, 1, 1),
- _Q1B(ST_SHARED_REPLAY, 0x0001, B6, REPLAY, 0x0000000c, 1, 1),
- _Q1B(LD_LOCAL_TRANSACTIONS, 0x0001, B6, TRANSACTION, 0x00000000, 1, 1),
- _Q1B(ST_LOCAL_TRANSACTIONS, 0x0001, B6, TRANSACTION, 0x00000004, 1, 1),
- _Q1B(L1_LD_SHARED_TRANSACTIONS, 0x0001, B6, TRANSACTION, 0x00000008, 1, 1),
- _Q1B(L1_ST_SHARED_TRANSACTIONS, 0x0001, B6, TRANSACTION, 0x0000000c, 1, 1),
+ _Q1B(SHARED_LD_REPLAY, 0x0001, B6, REPLAY, 0x00000008, 1, 1),
+ _Q1B(SHARED_ST_REPLAY, 0x0001, B6, REPLAY, 0x0000000c, 1, 1),
+ _Q1B(LOCAL_LD_TRANSACTIONS, 0x0001, B6, TRANSACTION, 0x00000000, 1, 1),
+ _Q1B(LOCAL_ST_TRANSACTIONS, 0x0001, B6, TRANSACTION, 0x00000004, 1, 1),
+ _Q1B(L1_SHARED_LD_TRANSACTIONS, 0x0001, B6, TRANSACTION, 0x00000008, 1, 1),
+ _Q1B(L1_SHARED_ST_TRANSACTIONS, 0x0001, B6, TRANSACTION, 0x0000000c, 1, 1),
_Q1B(GLD_MEM_DIV_REPLAY, 0x0001, B6, REPLAY, 0x00000010, 1, 1),
_Q1B(GST_MEM_DIV_REPLAY, 0x0001, B6, REPLAY, 0x00000014, 1, 1),
_M2AB(IPC, 0x3, B6, EXEC, 0x398, 0xffff, LOGOP, WARP, 0x0, DIV_SUM_M0, 10, 1),
{
_Q(INST_EXECUTED, 0xaaaa, LOGOP, 0x2d, 3, 0x00, 0x11, 0x22, 0x00, 0x00, 0x00),
_Q(BRANCH, 0xaaaa, LOGOP, 0x1a, 2, 0x00, 0x11, 0x00, 0x00, 0x00, 0x00),
- _Q(BRANCH_DIVERGENT, 0xaaaa, LOGOP, 0x19, 2, 0x20, 0x31, 0x00, 0x00, 0x00, 0x00),
+ _Q(DIVERGENT_BRANCH, 0xaaaa, LOGOP, 0x19, 2, 0x20, 0x31, 0x00, 0x00, 0x00, 0x00),
_Q(ACTIVE_WARPS, 0xaaaa, LOGOP, 0x24, 6, 0x10, 0x21, 0x32, 0x43, 0x54, 0x65),
_Q(ACTIVE_CYCLES, 0xaaaa, LOGOP, 0x11, 1, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00),
- _Q(LAUNCHED_WARPS, 0xaaaa, LOGOP, 0x26, 1, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00),
- _Q(LAUNCHED_THREADS, 0xaaaa, LOGOP, 0x26, 6, 0x10, 0x21, 0x32, 0x43, 0x54, 0x65),
- _Q(LD_SHARED, 0xaaaa, LOGOP, 0x64, 1, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00),
- _Q(ST_SHARED, 0xaaaa, LOGOP, 0x64, 1, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00),
- _Q(LD_LOCAL, 0xaaaa, LOGOP, 0x64, 1, 0x20, 0x00, 0x00, 0x00, 0x00, 0x00),
- _Q(ST_LOCAL, 0xaaaa, LOGOP, 0x64, 1, 0x50, 0x00, 0x00, 0x00, 0x00, 0x00),
+ _Q(WARPS_LAUNCHED, 0xaaaa, LOGOP, 0x26, 1, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00),
+ _Q(THREADS_LAUNCHED, 0xaaaa, LOGOP, 0x26, 6, 0x10, 0x21, 0x32, 0x43, 0x54, 0x65),
+ _Q(SHARED_LD, 0xaaaa, LOGOP, 0x64, 1, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00),
+ _Q(SHARED_ST, 0xaaaa, LOGOP, 0x64, 1, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00),
+ _Q(LOCAL_LD, 0xaaaa, LOGOP, 0x64, 1, 0x20, 0x00, 0x00, 0x00, 0x00, 0x00),
+ _Q(LOCAL_ST, 0xaaaa, LOGOP, 0x64, 1, 0x50, 0x00, 0x00, 0x00, 0x00, 0x00),
_Q(GRED_COUNT, 0xaaaa, LOGOP, 0x63, 1, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00),
_Q(ATOM_COUNT, 0xaaaa, LOGOP, 0x63, 1, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00),
_Q(GLD_REQUEST, 0xaaaa, LOGOP, 0x64, 1, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00),
NVE4_PM_QUERY_PROF_TRIGGER_5,
NVE4_PM_QUERY_PROF_TRIGGER_6,
NVE4_PM_QUERY_PROF_TRIGGER_7,
- NVE4_PM_QUERY_LAUNCHED_WARPS,
- NVE4_PM_QUERY_LAUNCHED_THREADS,
- NVE4_PM_QUERY_LAUNCHED_CTA,
+ NVE4_PM_QUERY_WARPS_LAUNCHED,
+ NVE4_PM_QUERY_THREADS_LAUNCHED,
+ NVE4_PM_QUERY_SM_CTA_LAUNCHED,
NVE4_PM_QUERY_INST_ISSUED1,
NVE4_PM_QUERY_INST_ISSUED2,
NVE4_PM_QUERY_INST_EXECUTED,
- NVE4_PM_QUERY_LD_LOCAL,
- NVE4_PM_QUERY_ST_LOCAL,
- NVE4_PM_QUERY_LD_SHARED,
- NVE4_PM_QUERY_ST_SHARED,
- NVE4_PM_QUERY_L1_LOCAL_LOAD_HIT,
- NVE4_PM_QUERY_L1_LOCAL_LOAD_MISS,
- NVE4_PM_QUERY_L1_LOCAL_STORE_HIT,
- NVE4_PM_QUERY_L1_LOCAL_STORE_MISS,
+ NVE4_PM_QUERY_LOCAL_LD,
+ NVE4_PM_QUERY_LOCAL_ST,
+ NVE4_PM_QUERY_SHARED_LD,
+ NVE4_PM_QUERY_SHARED_ST,
+ NVE4_PM_QUERY_L1_LOCAL_LD_HIT,
+ NVE4_PM_QUERY_L1_LOCAL_LD_MISS,
+ NVE4_PM_QUERY_L1_LOCAL_ST_HIT,
+ NVE4_PM_QUERY_L1_LOCAL_ST_MISS,
NVE4_PM_QUERY_GLD_REQUEST,
NVE4_PM_QUERY_GST_REQUEST,
- NVE4_PM_QUERY_L1_GLOBAL_LOAD_HIT,
- NVE4_PM_QUERY_L1_GLOBAL_LOAD_MISS,
- NVE4_PM_QUERY_GLD_TRANSACTIONS_UNCACHED,
+ NVE4_PM_QUERY_L1_GLD_HIT,
+ NVE4_PM_QUERY_L1_GLD_MISS,
+ NVE4_PM_QUERY_UNCACHED_GLD_TRANSACTIONS,
NVE4_PM_QUERY_GST_TRANSACTIONS,
NVE4_PM_QUERY_BRANCH,
- NVE4_PM_QUERY_BRANCH_DIVERGENT,
+ NVE4_PM_QUERY_DIVERGENT_BRANCH,
NVE4_PM_QUERY_ACTIVE_WARPS,
NVE4_PM_QUERY_ACTIVE_CYCLES,
NVE4_PM_QUERY_INST_ISSUED,
NVE4_PM_QUERY_ATOM_COUNT,
NVE4_PM_QUERY_GRED_COUNT,
- NVE4_PM_QUERY_LD_SHARED_REPLAY,
- NVE4_PM_QUERY_ST_SHARED_REPLAY,
- NVE4_PM_QUERY_LD_LOCAL_TRANSACTIONS,
- NVE4_PM_QUERY_ST_LOCAL_TRANSACTIONS,
- NVE4_PM_QUERY_L1_LD_SHARED_TRANSACTIONS,
- NVE4_PM_QUERY_L1_ST_SHARED_TRANSACTIONS,
+ NVE4_PM_QUERY_SHARED_LD_REPLAY,
+ NVE4_PM_QUERY_SHARED_ST_REPLAY,
+ NVE4_PM_QUERY_LOCAL_LD_TRANSACTIONS,
+ NVE4_PM_QUERY_LOCAL_ST_TRANSACTIONS,
+ NVE4_PM_QUERY_L1_SHARED_LD_TRANSACTIONS,
+ NVE4_PM_QUERY_L1_SHARED_ST_TRANSACTIONS,
NVE4_PM_QUERY_GLD_MEM_DIV_REPLAY,
NVE4_PM_QUERY_GST_MEM_DIV_REPLAY,
NVE4_PM_QUERY_METRIC_IPC,
{
NVC0_PM_QUERY_INST_EXECUTED = 0,
NVC0_PM_QUERY_BRANCH,
- NVC0_PM_QUERY_BRANCH_DIVERGENT,
+ NVC0_PM_QUERY_DIVERGENT_BRANCH,
NVC0_PM_QUERY_ACTIVE_WARPS,
NVC0_PM_QUERY_ACTIVE_CYCLES,
- NVC0_PM_QUERY_LAUNCHED_WARPS,
- NVC0_PM_QUERY_LAUNCHED_THREADS,
- NVC0_PM_QUERY_LD_SHARED,
- NVC0_PM_QUERY_ST_SHARED,
- NVC0_PM_QUERY_LD_LOCAL,
- NVC0_PM_QUERY_ST_LOCAL,
+ NVC0_PM_QUERY_WARPS_LAUNCHED,
+ NVC0_PM_QUERY_THREADS_LAUNCHED,
+ NVC0_PM_QUERY_SHARED_LD,
+ NVC0_PM_QUERY_SHARED_ST,
+ NVC0_PM_QUERY_LOCAL_LD,
+ NVC0_PM_QUERY_LOCAL_ST,
NVC0_PM_QUERY_GRED_COUNT,
NVC0_PM_QUERY_ATOM_COUNT,
NVC0_PM_QUERY_GLD_REQUEST,