diff options
author | Samuel Pitoiset <[email protected]> | 2015-12-02 16:56:23 +0100 |
---|---|---|
committer | Samuel Pitoiset <[email protected]> | 2015-12-05 19:23:34 +0100 |
commit | 0afd8f7bd793eb111757a2921a392b2d129061c4 (patch) | |
tree | 560fd923df9d48fd8fa63b001f304958d5c6d70f /src | |
parent | af275b8839be4ecaee4777cafeeb940bbbfcccab (diff) |
nvc0: re-introduce performance metrics for SM30 (Kepler)
This implements more performance metrics than the previous support,
but some other metrics still need to be figured out.
Signed-off-by: Samuel Pitoiset <[email protected]>
Diffstat (limited to 'src')
-rw-r--r-- | src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_metric.c | 176 | ||||
-rw-r--r-- | src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_metric.h | 17 |
2 files changed, 188 insertions, 5 deletions
diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_metric.c b/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_metric.c index fb2806a805e..756ba4e8185 100644 --- a/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_metric.c +++ b/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_metric.c @@ -157,6 +157,109 @@ static const struct nvc0_hw_metric_query_cfg *sm21_hw_metric_queries[] = #undef _SM #undef _M +/* === PERFORMANCE MONITORING METRICS for NVE4+ === */ +static const char *nve4_hw_metric_names[] = +{ + "metric-achieved_occupancy", + "metric-branch_efficiency", + "metric-inst_issued", + "metric-inst_per_wrap", + "metric-inst_replay_overhead", + "metric-issued_ipc", + "metric-issue_slots", + "metric-issue_slot_utilization", + "metric-ipc", + "metric-shared_replay_overhead", +}; + +#define _SM(n) NVE4_HW_SM_QUERY(NVE4_HW_SM_QUERY_ ##n) +#define _M(n, c) [NVE4_HW_METRIC_QUERY_##n] = c + +/* ==== Compute capability 3.0 (GK104/GK106/GK107) ==== */ +static const struct nvc0_hw_metric_query_cfg +sm30_achieved_occupancy = +{ + .queries[0] = _SM(ACTIVE_WARPS), + .queries[1] = _SM(ACTIVE_CYCLES), + .num_queries = 2, +}; + +static const struct nvc0_hw_metric_query_cfg +sm30_branch_efficiency = +{ + .queries[0] = _SM(BRANCH), + .queries[1] = _SM(DIVERGENT_BRANCH), + .num_queries = 2, +}; + +static const struct nvc0_hw_metric_query_cfg +sm30_inst_issued = +{ + .queries[0] = _SM(INST_ISSUED1), + .queries[1] = _SM(INST_ISSUED2), + .num_queries = 2, +}; + +static const struct nvc0_hw_metric_query_cfg +sm30_inst_per_wrap = +{ + .queries[0] = _SM(INST_EXECUTED), + .queries[1] = _SM(WARPS_LAUNCHED), + .num_queries = 2, +}; + +static const struct nvc0_hw_metric_query_cfg +sm30_inst_replay_overhead = +{ + .queries[0] = _SM(INST_ISSUED1), + .queries[1] = _SM(INST_ISSUED2), + .queries[2] = _SM(INST_EXECUTED), + .num_queries = 3, +}; + +static const struct nvc0_hw_metric_query_cfg +sm30_issued_ipc = +{ + .queries[0] = _SM(INST_ISSUED1), + .queries[1] = _SM(INST_ISSUED2), + .queries[2] = _SM(ACTIVE_CYCLES), + .num_queries = 3, +}; + +static const struct nvc0_hw_metric_query_cfg +sm30_ipc = +{ + .queries[0] = _SM(INST_EXECUTED), + .queries[1] = _SM(ACTIVE_CYCLES), + .num_queries = 2, +}; + +static const struct nvc0_hw_metric_query_cfg +sm30_shared_replay_overhead = +{ + .queries[0] = _SM(SHARED_LD_REPLAY), + .queries[1] = _SM(SHARED_ST_REPLAY), + .queries[2] = _SM(INST_EXECUTED), + .num_queries = 3, +}; + +static const struct nvc0_hw_metric_query_cfg *sm30_hw_metric_queries[] = +{ + _M(ACHIEVED_OCCUPANCY, &sm30_achieved_occupancy), + _M(BRANCH_EFFICIENCY, &sm30_branch_efficiency), + _M(INST_ISSUED, &sm30_inst_issued), + _M(INST_PER_WRAP, &sm30_inst_per_wrap), + _M(INST_REPLAY_OVERHEAD, &sm30_inst_replay_overhead), + _M(ISSUED_IPC, &sm30_issued_ipc), + _M(ISSUE_SLOTS, &sm30_inst_issued), + _M(ISSUE_SLOT_UTILIZATION, &sm30_issued_ipc), + _M(IPC, &sm30_ipc), + _M(SHARED_REPLAY_OVERHEAD, &sm30_shared_replay_overhead), +}; + +#undef _SM +#undef _M + static inline const struct nvc0_hw_metric_query_cfg ** nvc0_hw_metric_get_queries(struct nvc0_screen *screen) { @@ -175,6 +278,9 @@ nvc0_hw_metric_query_get_cfg(struct nvc0_context *nvc0, struct nvc0_screen *screen = nvc0->screen; struct nvc0_query *q = &hq->base; + if (screen->base.class_3d >= NVE4_3D_CLASS) + return sm30_hw_metric_queries[q->type - NVE4_HW_METRIC_QUERY(0)]; + queries = nvc0_hw_metric_get_queries(screen); return queries[q->type - NVC0_HW_METRIC_QUERY(0)]; } @@ -309,6 +415,52 @@ sm21_hw_metric_calc_result(struct nvc0_hw_query *hq, uint64_t res64[8]) return 0; } +static uint64_t +sm30_hw_metric_calc_result(struct nvc0_hw_query *hq, uint64_t res64[8]) +{ + switch (hq->base.type - NVE4_HW_METRIC_QUERY(0)) { + case NVE4_HW_METRIC_QUERY_ACHIEVED_OCCUPANCY: + return sm20_hw_metric_calc_result(hq, res64); + case NVE4_HW_METRIC_QUERY_BRANCH_EFFICIENCY: + return sm20_hw_metric_calc_result(hq, res64); + case NVE4_HW_METRIC_QUERY_INST_ISSUED: + /* inst_issued1 + inst_issued2 * 2 */ + return res64[0] + res64[1] * 2; + case NVE4_HW_METRIC_QUERY_INST_PER_WRAP: + return sm20_hw_metric_calc_result(hq, res64); + case NVE4_HW_METRIC_QUERY_INST_REPLAY_OVERHEAD: + /* (metric-inst_issued - inst_executed) / inst_executed */ + if (res64[2]) + return (((res64[0] + res64[1] * 2) - res64[2]) / (double)res64[2]); + break; + case NVE4_HW_METRIC_QUERY_ISSUED_IPC: + /* metric-inst_issued / active_cycles */ + if (res64[2]) + return (res64[0] + res64[1] * 2) / (double)res64[2]; + break; + case NVE4_HW_METRIC_QUERY_ISSUE_SLOTS: + /* inst_issued1 + inst_issued2 */ + return res64[0] + res64[1]; + case NVE4_HW_METRIC_QUERY_ISSUE_SLOT_UTILIZATION: + /* ((metric-issue_slots / 2) / active_cycles) * 100 */ + if (res64[2]) + return (((res64[0] + res64[1]) / 2) / (double)res64[2]) * 100; + break; + case NVE4_HW_METRIC_QUERY_IPC: + return sm20_hw_metric_calc_result(hq, res64); + case NVE4_HW_METRIC_QUERY_SHARED_REPLAY_OVERHEAD: + /* (shared_load_replay + shared_store_replay) / inst_executed */ + if (res64[2]) + return (res64[0] + res64[1]) / (double)res64[2]; + break; + default: + debug_printf("invalid metric type: %d\n", + hq->base.type - NVE4_HW_METRIC_QUERY(0)); + break; + } + return 0; +} + static boolean nvc0_hw_metric_get_query_result(struct nvc0_context *nvc0, struct nvc0_hw_query *hq, boolean wait, @@ -331,10 +483,14 @@ nvc0_hw_metric_get_query_result(struct nvc0_context *nvc0, res64[i] = *(uint64_t *)&results[i]; } - if (dev->chipset == 0xc0 || dev->chipset == 0xc8) - value = sm20_hw_metric_calc_result(hq, res64); - else - value = sm21_hw_metric_calc_result(hq, res64); + if (screen->base.class_3d >= NVE4_3D_CLASS) { + value = sm30_hw_metric_calc_result(hq, res64); + } else { + if (dev->chipset == 0xc0 || dev->chipset == 0xc8) + value = sm20_hw_metric_calc_result(hq, res64); + else + value = sm21_hw_metric_calc_result(hq, res64); + } *(uint64_t *)result = value; return ret; @@ -355,7 +511,8 @@ nvc0_hw_metric_create_query(struct nvc0_context *nvc0, unsigned type) struct nvc0_hw_query *hq; unsigned i; - if (type < NVC0_HW_METRIC_QUERY(0) || type > NVC0_HW_METRIC_QUERY_LAST) + if ((type < NVE4_HW_METRIC_QUERY(0) || type > NVE4_HW_METRIC_QUERY_LAST) && + (type < NVC0_HW_METRIC_QUERY(0) || type > NVC0_HW_METRIC_QUERY_LAST)) return NULL; hmq = CALLOC_STRUCT(nvc0_hw_metric_query); @@ -406,6 +563,9 @@ nvc0_hw_metric_get_driver_query_info(struct nvc0_screen *screen, unsigned id, if (screen->base.device->drm_version >= 0x01000101) { if (screen->compute) { + if (screen->base.class_3d == NVE4_3D_CLASS) { + count += NVE4_HW_METRIC_QUERY_COUNT; + } else if (class_3d < NVE4_3D_CLASS) { const struct nvc0_hw_metric_query_cfg **queries = nvc0_hw_metric_get_queries(screen); @@ -424,6 +584,12 @@ nvc0_hw_metric_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_metric_names[id]; + info->query_type = NVE4_HW_METRIC_QUERY(id); + info->group_id = -1; + return 1; + } else if (class_3d < NVE4_3D_CLASS) { const struct nvc0_hw_metric_query_cfg **queries = nvc0_hw_metric_get_queries(screen); diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_metric.h b/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_metric.h index 95675fd19b7..06cb355db40 100644 --- a/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_metric.h +++ b/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw_metric.h @@ -18,6 +18,23 @@ nvc0_hw_metric_query(struct nvc0_hw_query *hq) /* * Driver metrics queries: */ +#define NVE4_HW_METRIC_QUERY(i) (PIPE_QUERY_DRIVER_SPECIFIC + 3072 + (i)) +#define NVE4_HW_METRIC_QUERY_LAST NVE4_HW_METRIC_QUERY(NVE4_HW_METRIC_QUERY_COUNT - 1) +enum nve4_hw_metric_queries +{ + NVE4_HW_METRIC_QUERY_ACHIEVED_OCCUPANCY = 0, + NVE4_HW_METRIC_QUERY_BRANCH_EFFICIENCY, + NVE4_HW_METRIC_QUERY_INST_ISSUED, + NVE4_HW_METRIC_QUERY_INST_PER_WRAP, + NVE4_HW_METRIC_QUERY_INST_REPLAY_OVERHEAD, + NVE4_HW_METRIC_QUERY_ISSUED_IPC, + NVE4_HW_METRIC_QUERY_ISSUE_SLOTS, + NVE4_HW_METRIC_QUERY_ISSUE_SLOT_UTILIZATION, + NVE4_HW_METRIC_QUERY_IPC, + NVE4_HW_METRIC_QUERY_SHARED_REPLAY_OVERHEAD, + NVE4_HW_METRIC_QUERY_COUNT +}; + #define NVC0_HW_METRIC_QUERY(i) (PIPE_QUERY_DRIVER_SPECIFIC + 3072 + (i)) #define NVC0_HW_METRIC_QUERY_LAST NVC0_HW_METRIC_QUERY(NVC0_HW_METRIC_QUERY_COUNT - 1) enum nvc0_hw_metric_queries |