nvc0: rework the MP counters infrastructure

This mainly improves how we define the different list of queries.

Signed-off-by: Samuel Pitoiset <[email protected]>
Acked-by: Ilia Mirkin <[email protected]>

1 files changed, 213 insertions, 201 deletions

diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_sm.c b/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_sm.c
index f5f9bb39fd9..7c4ab1c19c8 100644
--- a/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_sm.c
+++ b/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_sm.c
@@ -30,59 +30,85 @@
 #include "nvc0/nve4_compute.xml.h"
 #include "nvc0/nvc0_compute.xml.h"
 
-/* === PERFORMANCE MONITORING COUNTERS for NVE4+ === */
-
 /* NOTE: intentionally using the same names as NV */
-static const char *nve4_hw_sm_query_names[] =
-{
-   /* MP counters */
-   "active_cycles",
-   "active_warps",
-   "atom_cas_count",
-   "atom_count",
-   "branch",
-   "divergent_branch",
-   "gld_request",
-   "global_ld_mem_divergence_replays",
-   "global_store_transaction",
-   "global_st_mem_divergence_replays",
-   "gred_count",
-   "gst_request",
-   "inst_executed",
-   "inst_issued1",
-   "inst_issued2",
-   "l1_global_load_hit",
-   "l1_global_load_miss",
-   "__l1_global_load_transactions",
-   "__l1_global_store_transactions",
-   "l1_local_load_hit",
-   "l1_local_load_miss",
-   "l1_local_store_hit",
-   "l1_local_store_miss",
-   "l1_shared_load_transactions",
-   "l1_shared_store_transactions",
-   "local_load",
-   "local_load_transactions",
-   "local_store",
-   "local_store_transactions",
-   "prof_trigger_00",
-   "prof_trigger_01",
-   "prof_trigger_02",
-   "prof_trigger_03",
-   "prof_trigger_04",
-   "prof_trigger_05",
-   "prof_trigger_06",
-   "prof_trigger_07",
-   "shared_load",
-   "shared_load_replay",
-   "shared_store",
-   "shared_store_replay",
-   "sm_cta_launched",
-   "threads_launched",
-   "uncached_global_load_transaction",
-   "warps_launched",
+#define _Q(t, n) { NVC0_HW_SM_QUERY_##t, n }
+struct {
+   unsigned type;
+   const char *name;
+} nvc0_hw_sm_queries[] = {
+   _Q(ACTIVE_CYCLES,                "active_cycles"                           ),
+   _Q(ACTIVE_WARPS,                 "active_warps"                            ),
+   _Q(ATOM_CAS_COUNT,               "atom_cas_count"                          ),
+   _Q(ATOM_COUNT,                   "atom_count"                              ),
+   _Q(BRANCH,                       "branch"                                  ),
+   _Q(DIVERGENT_BRANCH,             "divergent_branch"                        ),
+   _Q(GLD_REQUEST,                  "gld_request"                             ),
+   _Q(GLD_MEM_DIV_REPLAY,           "global_ld_mem_divergence_replays"        ),
+   _Q(GST_TRANSACTIONS,             "global_store_transaction"                ),
+   _Q(GST_MEM_DIV_REPLAY,           "global_st_mem_divergence_replays"        ),
+   _Q(GRED_COUNT,                   "gred_count"                              ),
+   _Q(GST_REQUEST,                  "gst_request"                             ),
+   _Q(INST_EXECUTED,                "inst_executed"                           ),
+   _Q(INST_ISSUED,                  "inst_issued"                             ),
+   _Q(INST_ISSUED1,                 "inst_issued1"                            ),
+   _Q(INST_ISSUED2,                 "inst_issued2"                            ),
+   _Q(INST_ISSUED1_0,               "inst_issued1_0"                          ),
+   _Q(INST_ISSUED1_1,               "inst_issued1_1"                          ),
+   _Q(INST_ISSUED2_0,               "inst_issued2_0"                          ),
+   _Q(INST_ISSUED2_1,               "inst_issued2_1"                          ),
+   _Q(L1_GLD_HIT,                   "l1_global_load_hit"                      ),
+   _Q(L1_GLD_MISS,                  "l1_global_load_miss"                     ),
+   _Q(L1_GLD_TRANSACTIONS,          "__l1_global_load_transactions"           ),
+   _Q(L1_GST_TRANSACTIONS,          "__l1_global_store_transactions"          ),
+   _Q(L1_LOCAL_LD_HIT,              "l1_local_load_hit"                       ),
+   _Q(L1_LOCAL_LD_MISS,             "l1_local_load_miss"                      ),
+   _Q(L1_LOCAL_ST_HIT,              "l1_local_store_hit"                      ),
+   _Q(L1_LOCAL_ST_MISS,             "l1_local_store_miss"                     ),
+   _Q(L1_SHARED_LD_TRANSACTIONS,    "l1_shared_load_transactions"             ),
+   _Q(L1_SHARED_ST_TRANSACTIONS,    "l1_shared_store_transactions"            ),
+   _Q(LOCAL_LD,                     "local_load"                              ),
+   _Q(LOCAL_LD_TRANSACTIONS,        "local_load_transactions"                 ),
+   _Q(LOCAL_ST,                     "local_store"                             ),
+   _Q(LOCAL_ST_TRANSACTIONS,        "local_store_transactions"                ),
+   _Q(PROF_TRIGGER_0,               "prof_trigger_00"                         ),
+   _Q(PROF_TRIGGER_1,               "prof_trigger_01"                         ),
+   _Q(PROF_TRIGGER_2,               "prof_trigger_02"                         ),
+   _Q(PROF_TRIGGER_3,               "prof_trigger_03"                         ),
+   _Q(PROF_TRIGGER_4,               "prof_trigger_04"                         ),
+   _Q(PROF_TRIGGER_5,               "prof_trigger_05"                         ),
+   _Q(PROF_TRIGGER_6,               "prof_trigger_06"                         ),
+   _Q(PROF_TRIGGER_7,               "prof_trigger_07"                         ),
+   _Q(SHARED_LD,                    "shared_load"                             ),
+   _Q(SHARED_LD_REPLAY,             "shared_load_replay"                      ),
+   _Q(SHARED_ST,                    "shared_store"                            ),
+   _Q(SHARED_ST_REPLAY,             "shared_store_replay"                     ),
+   _Q(SM_CTA_LAUNCHED,              "sm_cta_launched"                         ),
+   _Q(THREADS_LAUNCHED,             "threads_launched"                        ),
+   _Q(TH_INST_EXECUTED_0,           "thread_inst_executed_0"                  ),
+   _Q(TH_INST_EXECUTED_1,           "thread_inst_executed_1"                  ),
+   _Q(TH_INST_EXECUTED_2,           "thread_inst_executed_2"                  ),
+   _Q(TH_INST_EXECUTED_3,           "thread_inst_executed_3"                  ),
+   _Q(UNCACHED_GLD_TRANSACTIONS,    "uncached_global_load_transaction"        ),
+   _Q(WARPS_LAUNCHED,               "warps_launched"                          ),
 };
 
+#undef _Q
+
+static inline const char *
+nvc0_hw_sm_query_get_name(unsigned query_type)
+{
+   unsigned i;
+
+   for (i = 0; i < ARRAY_SIZE(nvc0_hw_sm_queries); i++) {
+      if (nvc0_hw_sm_queries[i].type == query_type)
+         return nvc0_hw_sm_queries[i].name;
+   }
+   assert(0);
+   return NULL;
+}
+
+/* === PERFORMANCE MONITORING COUNTERS for NVE4+ === */
+
 /* Code to read out MP counters: They are accessible via mmio, too, but let's
  * just avoid mapping registers in userspace. We'd have to know which MPs are
  * enabled/present, too, and that information is not presently exposed.
@@ -187,19 +213,20 @@ struct nvc0_hw_sm_counter_cfg
 
 struct nvc0_hw_sm_query_cfg
 {
+   unsigned type;
    struct nvc0_hw_sm_counter_cfg ctr[8];
    uint8_t num_counters;
    uint8_t norm[2]; /* normalization num,denom */
 };
 
-#define _Q1A(n, f, m, g, s, nu, dn) [NVE4_HW_SM_QUERY_##n] = { { { f, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m, 0, NVE4_COMPUTE_MP_PM_A_SIGSEL_##g, 0, s }, {}, {}, {} }, 1, { nu, dn } }
-#define _Q1B(n, f, m, g, s, nu, dn) [NVE4_HW_SM_QUERY_##n] = { { { f, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m, 1, NVE4_COMPUTE_MP_PM_B_SIGSEL_##g, 0, s }, {}, {}, {} }, 1, { nu, dn } }
+#define _Q1A(n, f, m, g, s, nu, dn) { NVC0_HW_SM_QUERY_##n, { { f, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m, 0, NVE4_COMPUTE_MP_PM_A_SIGSEL_##g, 0, s }, {}, {}, {} }, 1, { nu, dn } }
+#define _Q1B(n, f, m, g, s, nu, dn) { NVC0_HW_SM_QUERY_##n, { { f, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m, 1, NVE4_COMPUTE_MP_PM_B_SIGSEL_##g, 0, s }, {}, {}, {} }, 1, { nu, dn } }
 
 /* NOTES:
  * active_warps: bit 0 alternates btw 0 and 1 for odd nr of warps
  * inst_executed etc.: we only count a single warp scheduler
  */
-static const struct nvc0_hw_sm_query_cfg nve4_hw_sm_queries[] =
+static const struct nvc0_hw_sm_query_cfg sm30_hw_sm_queries[] =
 {
    _Q1B(ACTIVE_CYCLES, 0x0001, B6, WARP, 0x00000000, 1, 1),
    _Q1B(ACTIVE_WARPS,  0x003f, B6, WARP, 0x31483104, 2, 1),
@@ -257,43 +284,6 @@ static const struct nvc0_hw_sm_query_cfg nve4_hw_sm_queries[] =
  *   because there is a context-switch problem that we need to fix.
  *   Results might be wrong sometimes, be careful!
  */
-static const char *nvc0_hw_sm_query_names[] =
-{
-   /* MP counters */
-   "active_cycles",
-   "active_warps",
-   "atom_count",
-   "branch",
-   "divergent_branch",
-   "gld_request",
-   "gred_count",
-   "gst_request",
-   "inst_executed",
-   "inst_issued",
-   "inst_issued1_0",
-   "inst_issued1_1",
-   "inst_issued2_0",
-   "inst_issued2_1",
-   "local_load",
-   "local_store",
-   "prof_trigger_00",
-   "prof_trigger_01",
-   "prof_trigger_02",
-   "prof_trigger_03",
-   "prof_trigger_04",
-   "prof_trigger_05",
-   "prof_trigger_06",
-   "prof_trigger_07",
-   "shared_load",
-   "shared_store",
-   "threads_launched",
-   "thread_inst_executed_0",
-   "thread_inst_executed_1",
-   "thread_inst_executed_2",
-   "thread_inst_executed_3",
-   "warps_launched",
-};
-
 static const uint64_t nvc0_read_hw_sm_counters_code[] =
 {
    /* mov b32 $r8 $tidx
@@ -345,12 +335,12 @@ static const uint64_t nvc0_read_hw_sm_counters_code[] =
 };
 
 #define _C(f, o, g, m, s) { f, NVC0_COMPUTE_MP_PM_OP_MODE_##o, 0, g, m, s }
-#define _Q(n, c) [NVC0_HW_SM_QUERY_##n] = c
 
 /* ==== Compute capability 2.0 (GF100/GF110) ==== */
 static const struct nvc0_hw_sm_query_cfg
 sm20_active_cycles =
 {
+   .type         = NVC0_HW_SM_QUERY_ACTIVE_CYCLES,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x11, 0x000000ff, 0x00000000),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -359,6 +349,7 @@ sm20_active_cycles =
 static const struct nvc0_hw_sm_query_cfg
 sm20_active_warps =
 {
+   .type         = NVC0_HW_SM_QUERY_ACTIVE_WARPS,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x24, 0x000000ff, 0x00000010),
    .ctr[1]       = _C(0xaaaa, LOGOP, 0x24, 0x000000ff, 0x00000020),
    .ctr[2]       = _C(0xaaaa, LOGOP, 0x24, 0x000000ff, 0x00000030),
@@ -372,6 +363,7 @@ sm20_active_warps =
 static const struct nvc0_hw_sm_query_cfg
 sm20_atom_count =
 {
+   .type         = NVC0_HW_SM_QUERY_ATOM_COUNT,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x63, 0x000000ff, 0x00000030),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -380,6 +372,7 @@ sm20_atom_count =
 static const struct nvc0_hw_sm_query_cfg
 sm20_branch =
 {
+   .type         = NVC0_HW_SM_QUERY_BRANCH,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x1a, 0x000000ff, 0x00000000),
    .ctr[1]       = _C(0xaaaa, LOGOP, 0x1a, 0x000000ff, 0x00000010),
    .num_counters = 2,
@@ -389,6 +382,7 @@ sm20_branch =
 static const struct nvc0_hw_sm_query_cfg
 sm20_divergent_branch =
 {
+   .type         = NVC0_HW_SM_QUERY_DIVERGENT_BRANCH,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x19, 0x000000ff, 0x00000020),
    .ctr[1]       = _C(0xaaaa, LOGOP, 0x19, 0x000000ff, 0x00000030),
    .num_counters = 2,
@@ -398,6 +392,7 @@ sm20_divergent_branch =
 static const struct nvc0_hw_sm_query_cfg
 sm20_gld_request =
 {
+   .type         = NVC0_HW_SM_QUERY_GLD_REQUEST,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x64, 0x000000ff, 0x00000030),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -406,6 +401,7 @@ sm20_gld_request =
 static const struct nvc0_hw_sm_query_cfg
 sm20_gred_count =
 {
+   .type         = NVC0_HW_SM_QUERY_GRED_COUNT,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x63, 0x000000ff, 0x00000040),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -414,6 +410,7 @@ sm20_gred_count =
 static const struct nvc0_hw_sm_query_cfg
 sm20_gst_request =
 {
+   .type         = NVC0_HW_SM_QUERY_GST_REQUEST,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x64, 0x000000ff, 0x00000060),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -422,6 +419,7 @@ sm20_gst_request =
 static const struct nvc0_hw_sm_query_cfg
 sm20_inst_executed =
 {
+   .type         = NVC0_HW_SM_QUERY_INST_EXECUTED,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x2d, 0x0000ffff, 0x00001000),
    .ctr[1]       = _C(0xaaaa, LOGOP, 0x2d, 0x0000ffff, 0x00001010),
    .num_counters = 2,
@@ -431,6 +429,7 @@ sm20_inst_executed =
 static const struct nvc0_hw_sm_query_cfg
 sm20_inst_issued =
 {
+   .type         = NVC0_HW_SM_QUERY_INST_ISSUED,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x27, 0x0000ffff, 0x00007060),
    .ctr[1]       = _C(0xaaaa, LOGOP, 0x27, 0x0000ffff, 0x00007070),
    .num_counters = 2,
@@ -440,6 +439,7 @@ sm20_inst_issued =
 static const struct nvc0_hw_sm_query_cfg
 sm20_local_ld =
 {
+   .type         = NVC0_HW_SM_QUERY_LOCAL_LD,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x64, 0x000000ff, 0x00000020),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -448,6 +448,7 @@ sm20_local_ld =
 static const struct nvc0_hw_sm_query_cfg
 sm20_local_st =
 {
+   .type         = NVC0_HW_SM_QUERY_LOCAL_ST,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x64, 0x000000ff, 0x00000050),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -456,6 +457,7 @@ sm20_local_st =
 static const struct nvc0_hw_sm_query_cfg
 sm20_prof_trigger_0 =
 {
+   .type         = NVC0_HW_SM_QUERY_PROF_TRIGGER_0,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x01, 0x000000ff, 0x00000000),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -464,6 +466,7 @@ sm20_prof_trigger_0 =
 static const struct nvc0_hw_sm_query_cfg
 sm20_prof_trigger_1 =
 {
+   .type         = NVC0_HW_SM_QUERY_PROF_TRIGGER_1,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x01, 0x000000ff, 0x00000010),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -472,6 +475,7 @@ sm20_prof_trigger_1 =
 static const struct nvc0_hw_sm_query_cfg
 sm20_prof_trigger_2 =
 {
+   .type         = NVC0_HW_SM_QUERY_PROF_TRIGGER_2,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x01, 0x000000ff, 0x00000020),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -480,6 +484,7 @@ sm20_prof_trigger_2 =
 static const struct nvc0_hw_sm_query_cfg
 sm20_prof_trigger_3 =
 {
+   .type         = NVC0_HW_SM_QUERY_PROF_TRIGGER_3,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x01, 0x000000ff, 0x00000030),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -488,6 +493,7 @@ sm20_prof_trigger_3 =
 static const struct nvc0_hw_sm_query_cfg
 sm20_prof_trigger_4 =
 {
+   .type         = NVC0_HW_SM_QUERY_PROF_TRIGGER_4,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x01, 0x000000ff, 0x00000040),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -496,6 +502,7 @@ sm20_prof_trigger_4 =
 static const struct nvc0_hw_sm_query_cfg
 sm20_prof_trigger_5 =
 {
+   .type         = NVC0_HW_SM_QUERY_PROF_TRIGGER_5,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x01, 0x000000ff, 0x00000050),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -504,6 +511,7 @@ sm20_prof_trigger_5 =
 static const struct nvc0_hw_sm_query_cfg
 sm20_prof_trigger_6 =
 {
+   .type         = NVC0_HW_SM_QUERY_PROF_TRIGGER_6,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x01, 0x000000ff, 0x00000060),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -512,6 +520,7 @@ sm20_prof_trigger_6 =
 static const struct nvc0_hw_sm_query_cfg
 sm20_prof_trigger_7 =
 {
+   .type         = NVC0_HW_SM_QUERY_PROF_TRIGGER_7,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x01, 0x000000ff, 0x00000070),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -520,6 +529,7 @@ sm20_prof_trigger_7 =
 static const struct nvc0_hw_sm_query_cfg
 sm20_shared_ld =
 {
+   .type         = NVC0_HW_SM_QUERY_SHARED_LD,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x64, 0x000000ff, 0x00000010),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -528,6 +538,7 @@ sm20_shared_ld =
 static const struct nvc0_hw_sm_query_cfg
 sm20_shared_st =
 {
+   .type         = NVC0_HW_SM_QUERY_SHARED_ST,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x64, 0x000000ff, 0x00000040),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -536,6 +547,7 @@ sm20_shared_st =
 static const struct nvc0_hw_sm_query_cfg
 sm20_threads_launched =
 {
+   .type         = NVC0_HW_SM_QUERY_THREADS_LAUNCHED,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x26, 0x000000ff, 0x00000010),
    .ctr[1]       = _C(0xaaaa, LOGOP, 0x26, 0x000000ff, 0x00000020),
    .ctr[2]       = _C(0xaaaa, LOGOP, 0x26, 0x000000ff, 0x00000030),
@@ -549,6 +561,7 @@ sm20_threads_launched =
 static const struct nvc0_hw_sm_query_cfg
 sm20_th_inst_executed_0 =
 {
+   .type         = NVC0_HW_SM_QUERY_TH_INST_EXECUTED_0,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x2f, 0x000000ff, 0x00000000),
    .ctr[1]       = _C(0xaaaa, LOGOP, 0x2f, 0x000000ff, 0x00000010),
    .ctr[2]       = _C(0xaaaa, LOGOP, 0x2f, 0x000000ff, 0x00000020),
@@ -562,6 +575,7 @@ sm20_th_inst_executed_0 =
 static const struct nvc0_hw_sm_query_cfg
 sm20_th_inst_executed_1 =
 {
+   .type         = NVC0_HW_SM_QUERY_TH_INST_EXECUTED_1,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x30, 0x000000ff, 0x00000000),
    .ctr[1]       = _C(0xaaaa, LOGOP, 0x30, 0x000000ff, 0x00000010),
    .ctr[2]       = _C(0xaaaa, LOGOP, 0x30, 0x000000ff, 0x00000020),
@@ -575,6 +589,7 @@ sm20_th_inst_executed_1 =
 static const struct nvc0_hw_sm_query_cfg
 sm20_warps_launched =
 {
+   .type         = NVC0_HW_SM_QUERY_WARPS_LAUNCHED,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x26, 0x000000ff, 0x00000000),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -582,44 +597,39 @@ sm20_warps_launched =
 
 static const struct nvc0_hw_sm_query_cfg *sm20_hw_sm_queries[] =
 {
-   _Q(ACTIVE_CYCLES,       &sm20_active_cycles),
-   _Q(ACTIVE_WARPS,        &sm20_active_warps),
-   _Q(ATOM_COUNT,          &sm20_atom_count),
-   _Q(BRANCH,              &sm20_branch),
-   _Q(DIVERGENT_BRANCH,    &sm20_divergent_branch),
-   _Q(GLD_REQUEST,         &sm20_gld_request),
-   _Q(GRED_COUNT,          &sm20_gred_count),
-   _Q(GST_REQUEST,         &sm20_gst_request),
-   _Q(INST_EXECUTED,       &sm20_inst_executed),
-   _Q(INST_ISSUED,         &sm20_inst_issued),
-   _Q(INST_ISSUED1_0,      NULL),
-   _Q(INST_ISSUED1_1,      NULL),
-   _Q(INST_ISSUED2_0,      NULL),
-   _Q(INST_ISSUED2_1,      NULL),
-   _Q(LOCAL_LD,            &sm20_local_ld),
-   _Q(LOCAL_ST,            &sm20_local_st),
-   _Q(PROF_TRIGGER_0,      &sm20_prof_trigger_0),
-   _Q(PROF_TRIGGER_1,      &sm20_prof_trigger_1),
-   _Q(PROF_TRIGGER_2,      &sm20_prof_trigger_2),
-   _Q(PROF_TRIGGER_3,      &sm20_prof_trigger_3),
-   _Q(PROF_TRIGGER_4,      &sm20_prof_trigger_4),
-   _Q(PROF_TRIGGER_5,      &sm20_prof_trigger_5),
-   _Q(PROF_TRIGGER_6,      &sm20_prof_trigger_6),
-   _Q(PROF_TRIGGER_7,      &sm20_prof_trigger_7),
-   _Q(SHARED_LD,           &sm20_shared_ld),
-   _Q(SHARED_ST,           &sm20_shared_st),
-   _Q(THREADS_LAUNCHED,    &sm20_threads_launched),
-   _Q(TH_INST_EXECUTED_0,  &sm20_th_inst_executed_0),
-   _Q(TH_INST_EXECUTED_1,  &sm20_th_inst_executed_1),
-   _Q(TH_INST_EXECUTED_2,  NULL),
-   _Q(TH_INST_EXECUTED_3,  NULL),
-   _Q(WARPS_LAUNCHED,      &sm20_warps_launched),
+   &sm20_active_cycles,
+   &sm20_active_warps,
+   &sm20_atom_count,
+   &sm20_branch,
+   &sm20_divergent_branch,
+   &sm20_gld_request,
+   &sm20_gred_count,
+   &sm20_gst_request,
+   &sm20_inst_executed,
+   &sm20_inst_issued,
+   &sm20_local_ld,
+   &sm20_local_st,
+   &sm20_prof_trigger_0,
+   &sm20_prof_trigger_1,
+   &sm20_prof_trigger_2,
+   &sm20_prof_trigger_3,
+   &sm20_prof_trigger_4,
+   &sm20_prof_trigger_5,
+   &sm20_prof_trigger_6,
+   &sm20_prof_trigger_7,
+   &sm20_shared_ld,
+   &sm20_shared_st,
+   &sm20_threads_launched,
+   &sm20_th_inst_executed_0,
+   &sm20_th_inst_executed_1,
+   &sm20_warps_launched,
 };
 
 /* ==== Compute capability 2.1 (GF108+ except GF110) ==== */
 static const struct nvc0_hw_sm_query_cfg
 sm21_inst_executed =
 {
+   .type         = NVC0_HW_SM_QUERY_INST_EXECUTED,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x2d, 0x000000ff, 0x00000000),
    .ctr[1]       = _C(0xaaaa, LOGOP, 0x2d, 0x000000ff, 0x00000010),
    .ctr[2]       = _C(0xaaaa, LOGOP, 0x2d, 0x000000ff, 0x00000020),
@@ -630,6 +640,7 @@ sm21_inst_executed =
 static const struct nvc0_hw_sm_query_cfg
 sm21_inst_issued1_0 =
 {
+   .type         = NVC0_HW_SM_QUERY_INST_ISSUED1_0,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x7e, 0x000000ff, 0x00000010),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -638,6 +649,7 @@ sm21_inst_issued1_0 =
 static const struct nvc0_hw_sm_query_cfg
 sm21_inst_issued1_1 =
 {
+   .type         = NVC0_HW_SM_QUERY_INST_ISSUED1_1,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x7e, 0x000000ff, 0x00000040),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -646,6 +658,7 @@ sm21_inst_issued1_1 =
 static const struct nvc0_hw_sm_query_cfg
 sm21_inst_issued2_0 =
 {
+   .type         = NVC0_HW_SM_QUERY_INST_ISSUED2_0,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x7e, 0x000000ff, 0x00000020),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -654,6 +667,7 @@ sm21_inst_issued2_0 =
 static const struct nvc0_hw_sm_query_cfg
 sm21_inst_issued2_1 =
 {
+   .type         = NVC0_HW_SM_QUERY_INST_ISSUED2_1,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0x7e, 0x000000ff, 0x00000050),
    .num_counters = 1,
    .norm         = { 1, 1 },
@@ -662,6 +676,7 @@ sm21_inst_issued2_1 =
 static const struct nvc0_hw_sm_query_cfg
 sm21_th_inst_executed_0 =
 {
+   .type         = NVC0_HW_SM_QUERY_TH_INST_EXECUTED_0,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0xa3, 0x000000ff, 0x00000000),
    .ctr[1]       = _C(0xaaaa, LOGOP, 0xa3, 0x000000ff, 0x00000010),
    .ctr[2]       = _C(0xaaaa, LOGOP, 0xa3, 0x000000ff, 0x00000020),
@@ -675,6 +690,7 @@ sm21_th_inst_executed_0 =
 static const struct nvc0_hw_sm_query_cfg
 sm21_th_inst_executed_1 =
 {
+   .type         = NVC0_HW_SM_QUERY_TH_INST_EXECUTED_1,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0xa5, 0x000000ff, 0x00000000),
    .ctr[1]       = _C(0xaaaa, LOGOP, 0xa5, 0x000000ff, 0x00000010),
    .ctr[2]       = _C(0xaaaa, LOGOP, 0xa5, 0x000000ff, 0x00000020),
@@ -688,6 +704,7 @@ sm21_th_inst_executed_1 =
 static const struct nvc0_hw_sm_query_cfg
 sm21_th_inst_executed_2 =
 {
+   .type         = NVC0_HW_SM_QUERY_TH_INST_EXECUTED_2,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0xa4, 0x000000ff, 0x00000000),
    .ctr[1]       = _C(0xaaaa, LOGOP, 0xa4, 0x000000ff, 0x00000010),
    .ctr[2]       = _C(0xaaaa, LOGOP, 0xa4, 0x000000ff, 0x00000020),
@@ -701,6 +718,7 @@ sm21_th_inst_executed_2 =
 static const struct nvc0_hw_sm_query_cfg
 sm21_th_inst_executed_3 =
 {
+   .type         = NVC0_HW_SM_QUERY_TH_INST_EXECUTED_3,
    .ctr[0]       = _C(0xaaaa, LOGOP, 0xa6, 0x000000ff, 0x00000000),
    .ctr[1]       = _C(0xaaaa, LOGOP, 0xa6, 0x000000ff, 0x00000010),
    .ctr[2]       = _C(0xaaaa, LOGOP, 0xa6, 0x000000ff, 0x00000020),
@@ -713,41 +731,39 @@ sm21_th_inst_executed_3 =
 
 static const struct nvc0_hw_sm_query_cfg *sm21_hw_sm_queries[] =
 {
-   _Q(ACTIVE_CYCLES,       &sm20_active_cycles),
-   _Q(ACTIVE_WARPS,        &sm20_active_warps),
-   _Q(ATOM_COUNT,          &sm20_atom_count),
-   _Q(BRANCH,              &sm20_branch),
-   _Q(DIVERGENT_BRANCH,    &sm20_divergent_branch),
-   _Q(GLD_REQUEST,         &sm20_gld_request),
-   _Q(GRED_COUNT,          &sm20_gred_count),
-   _Q(GST_REQUEST,         &sm20_gst_request),
-   _Q(INST_EXECUTED,       &sm21_inst_executed),
-   _Q(INST_ISSUED,         NULL),
-   _Q(INST_ISSUED1_0,      &sm21_inst_issued1_0),
-   _Q(INST_ISSUED1_1,      &sm21_inst_issued1_1),
-   _Q(INST_ISSUED2_0,      &sm21_inst_issued2_0),
-   _Q(INST_ISSUED2_1,      &sm21_inst_issued2_1),
-   _Q(LOCAL_LD,            &sm20_local_ld),
-   _Q(LOCAL_ST,            &sm20_local_st),
-   _Q(PROF_TRIGGER_0,      &sm20_prof_trigger_0),
-   _Q(PROF_TRIGGER_1,      &sm20_prof_trigger_1),
-   _Q(PROF_TRIGGER_2,      &sm20_prof_trigger_2),
-   _Q(PROF_TRIGGER_3,      &sm20_prof_trigger_3),
-   _Q(PROF_TRIGGER_4,      &sm20_prof_trigger_4),
-   _Q(PROF_TRIGGER_5,      &sm20_prof_trigger_5),
-   _Q(PROF_TRIGGER_6,      &sm20_prof_trigger_6),
-   _Q(PROF_TRIGGER_7,      &sm20_prof_trigger_7),
-   _Q(SHARED_LD,           &sm20_shared_ld),
-   _Q(SHARED_ST,           &sm20_shared_st),
-   _Q(THREADS_LAUNCHED,    &sm20_threads_launched),
-   _Q(TH_INST_EXECUTED_0,  &sm21_th_inst_executed_0),
-   _Q(TH_INST_EXECUTED_1,  &sm21_th_inst_executed_1),
-   _Q(TH_INST_EXECUTED_2,  &sm21_th_inst_executed_2),
-   _Q(TH_INST_EXECUTED_3,  &sm21_th_inst_executed_3),
-   _Q(WARPS_LAUNCHED,      &sm20_warps_launched),
+   &sm20_active_cycles,
+   &sm20_active_warps,
+   &sm20_atom_count,
+   &sm20_branch,
+   &sm20_divergent_branch,
+   &sm20_gld_request,
+   &sm20_gred_count,
+   &sm20_gst_request,
+   &sm21_inst_executed,
+   &sm21_inst_issued1_0,
+   &sm21_inst_issued1_1,
+   &sm21_inst_issued2_0,
+   &sm21_inst_issued2_1,
+   &sm20_local_ld,
+   &sm20_local_st,
+   &sm20_prof_trigger_0,
+   &sm20_prof_trigger_1,
+   &sm20_prof_trigger_2,
+   &sm20_prof_trigger_3,
+   &sm20_prof_trigger_4,
+   &sm20_prof_trigger_5,
+   &sm20_prof_trigger_6,
+   &sm20_prof_trigger_7,
+   &sm20_shared_ld,
+   &sm20_shared_st,
+   &sm20_threads_launched,
+   &sm21_th_inst_executed_0,
+   &sm21_th_inst_executed_1,
+   &sm21_th_inst_executed_2,
+   &sm21_th_inst_executed_3,
+   &sm20_warps_launched,
 };
 
-#undef _Q
 #undef _C
 
 static inline const struct nvc0_hw_sm_query_cfg **
@@ -760,21 +776,47 @@ nvc0_hw_sm_get_queries(struct nvc0_screen *screen)
    return sm21_hw_sm_queries;
 }
 
+unsigned
+nvc0_hw_sm_get_num_queries(struct nvc0_screen *screen)
+{
+   struct nouveau_device *dev = screen->base.device;
+
+   switch (screen->base.class_3d) {
+   case NVE4_3D_CLASS:
+      return ARRAY_SIZE(sm30_hw_sm_queries);
+   default:
+      if (dev->chipset == 0xc0 || dev->chipset == 0xc8)
+         return ARRAY_SIZE(sm20_hw_sm_queries);
+      return ARRAY_SIZE(sm21_hw_sm_queries);
+   }
+   return 0;
+}
+
 static const struct nvc0_hw_sm_query_cfg *
 nvc0_hw_sm_query_get_cfg(struct nvc0_context *nvc0, struct nvc0_hw_query *hq)
 {
    struct nvc0_screen *screen = nvc0->screen;
    struct nvc0_query *q = &hq->base;
+   unsigned num_queries;
+   unsigned i;
 
-   if (screen->base.class_3d >= NVE4_3D_CLASS)
-      return &nve4_hw_sm_queries[q->type - PIPE_QUERY_DRIVER_SPECIFIC];
+   num_queries = nvc0_hw_sm_get_num_queries(screen);
 
-   if (q->type >= NVC0_HW_SM_QUERY(0) && q->type <= NVC0_HW_SM_QUERY_LAST) {
+   if (screen->base.class_3d >= NVE4_3D_CLASS) {
+      for (i = 0; i < num_queries; i++) {
+         if (NVC0_HW_SM_QUERY(sm30_hw_sm_queries[i].type) == q->type)
+            return &sm30_hw_sm_queries[i];
+      }
+   } else {
       const struct nvc0_hw_sm_query_cfg **queries =
-         nvc0_hw_sm_get_queries(screen);
-      return queries[q->type - NVC0_HW_SM_QUERY(0)];
+          nvc0_hw_sm_get_queries(screen);
+
+      for (i = 0; i < num_queries; i++) {
+         if (NVC0_HW_SM_QUERY(queries[i]->type) == q->type)
+            return queries[i];
+      }
    }
-   debug_printf("invalid query type: %d\n", q->type);
+   assert(0);
    return NULL;
 }
 
@@ -1132,8 +1174,7 @@ nvc0_hw_sm_create_query(struct nvc0_context *nvc0, unsigned type)
    if (nvc0->screen->base.drm->version < 0x01000101)
       return NULL;
 
-   if ((type < NVE4_HW_SM_QUERY(0) || type > NVE4_HW_SM_QUERY_LAST) &&
-       (type < NVC0_HW_SM_QUERY(0) || type > NVC0_HW_SM_QUERY_LAST))
+   if (type < NVC0_HW_SM_QUERY(0) || type > NVC0_HW_SM_QUERY_LAST)
       return NULL;
 
    hsq = CALLOC_STRUCT(nvc0_hw_sm_query);
@@ -1201,23 +1242,6 @@ nvc0_hw_sm_create_query(struct nvc0_context *nvc0, unsigned type)
    return hq;
 }
 
-static int
-nvc0_hw_sm_get_next_query_id(const struct nvc0_hw_sm_query_cfg **queries,
-                             unsigned id)
-{
-   unsigned i, next = 0;
-
-   for (i = 0; i < NVC0_HW_SM_QUERY_COUNT; i++) {
-      if (!queries[i]) {
-         next++;
-      } else
-      if (i >= id && queries[id + next]) {
-         break;
-      }
-   }
-   return id + next;
-}
-
 int
 nvc0_hw_sm_get_driver_query_info(struct nvc0_screen *screen, unsigned id,
                                  struct pipe_driver_query_info *info)
@@ -1225,21 +1249,8 @@ nvc0_hw_sm_get_driver_query_info(struct nvc0_screen *screen, unsigned id,
    int count = 0;
 
    if (screen->base.drm->version >= 0x01000101) {
-      if (screen->compute) {
-         if (screen->base.class_3d == NVE4_3D_CLASS) {
-            count += NVE4_HW_SM_QUERY_COUNT;
-         } else
-         if (screen->base.class_3d < NVE4_3D_CLASS) {
-            const struct nvc0_hw_sm_query_cfg **queries =
-               nvc0_hw_sm_get_queries(screen);
-            unsigned i;
-
-            for (i = 0; i < NVC0_HW_SM_QUERY_COUNT; i++) {
-               if (queries[i])
-                  count++;
-            }
-         }
-      }
+      if (screen->compute)
+         count = nvc0_hw_sm_get_num_queries(screen);
    }
 
    if (!info)
@@ -1248,8 +1259,10 @@ nvc0_hw_sm_get_driver_query_info(struct nvc0_screen *screen, unsigned id,
    if (id < count) {
       if (screen->compute) {
          if (screen->base.class_3d == NVE4_3D_CLASS) {
-            info->name = nve4_hw_sm_query_names[id];
-            info->query_type = NVE4_HW_SM_QUERY(id);
+            const struct nvc0_hw_sm_query_cfg *q = &sm30_hw_sm_queries[id];
+
+            info->name = nvc0_hw_sm_query_get_name(q->type);
+            info->query_type = NVC0_HW_SM_QUERY(q->type);
             info->group_id = NVC0_HW_SM_QUERY_GROUP;
             return 1;
          } else
@@ -1257,9 +1270,8 @@ nvc0_hw_sm_get_driver_query_info(struct nvc0_screen *screen, unsigned id,
             const struct nvc0_hw_sm_query_cfg **queries =
                nvc0_hw_sm_get_queries(screen);
 
-            id = nvc0_hw_sm_get_next_query_id(queries, id);
-            info->name = nvc0_hw_sm_query_names[id];
-            info->query_type = NVC0_HW_SM_QUERY(id);
+            info->name = nvc0_hw_sm_query_get_name(queries[id]->type);
+            info->query_type = NVC0_HW_SM_QUERY(queries[id]->type);
             info->group_id = NVC0_HW_SM_QUERY_GROUP;
             return 1;
          }