freedreno/a6xx: Use counter allocation helper

If the kernel supports PERFCNTR_CONFIG for counter reservation, we can
expose perfcntrs by default.

Signed-off-by: Rob Clark <rob.clark@oss.qualcomm.com>

src/gallium/drivers/freedreno/a6xx/fd6_query.cc

@@ -824,6 +824,7 @@ static const struct fd_acc_sample_provider so_overflow_predicate = {
 struct fd_batch_query_entry {
    uint8_t gid; /* group-id */
    uint8_t cid; /* countable-id within the group */
+   const struct fd_perfcntr_counter *counter;
 };
 
 struct fd_batch_query_data {
@@ -839,33 +840,23 @@ perfcntr_resume(struct fd_acc_query *aq, struct fd_batch *batch) assert_dt
    struct fd_screen *screen = data->screen;
    fd_cs cs(batch->draw);
 
-   unsigned counters_per_group[screen->num_perfcntr_groups];
-   memset(counters_per_group, 0, sizeof(counters_per_group));
-
    fd_pkt7(cs, CP_WAIT_FOR_IDLE, 0);
 
    /* configure performance counters for the requested queries: */
    for (unsigned i = 0; i < data->num_query_entries; i++) {
       struct fd_batch_query_entry *entry = &data->query_entries[i];
       const struct fd_perfcntr_group *g = &screen->perfcntr_groups[entry->gid];
-      unsigned counter_idx = counters_per_group[entry->gid]++;
-
-      assert(counter_idx < g->num_counters);
 
       fd_pkt4(cs, 1).add((fd_reg_pair){
-         .reg = g->counters[counter_idx].select_reg,
+         .reg = entry->counter->select_reg,
          .value = g->countables[entry->cid].selector,
       });
    }
 
-   memset(counters_per_group, 0, sizeof(counters_per_group));
-
    /* and snapshot the start values */
    for (unsigned i = 0; i < data->num_query_entries; i++) {
       struct fd_batch_query_entry *entry = &data->query_entries[i];
-      const struct fd_perfcntr_group *g = &screen->perfcntr_groups[entry->gid];
-      unsigned counter_idx = counters_per_group[entry->gid]++;
-      const struct fd_perfcntr_counter *counter = &g->counters[counter_idx];
+      const struct fd_perfcntr_counter *counter = entry->counter;
 
       fd_pkt7(cs, CP_REG_TO_MEM, 3)
          .add(CP_REG_TO_MEM_0(.reg = counter->counter_reg_lo, ._64b = true))
@@ -877,12 +868,8 @@ static void
 perfcntr_pause(struct fd_acc_query *aq, struct fd_batch *batch) assert_dt
 {
    struct fd_batch_query_data *data = (struct fd_batch_query_data *)aq->query_data;
-   struct fd_screen *screen = data->screen;
    fd_cs cs(batch->draw);
 
-   unsigned counters_per_group[screen->num_perfcntr_groups];
-   memset(counters_per_group, 0, sizeof(counters_per_group));
-
    fd_pkt7(cs, CP_WAIT_FOR_IDLE, 0);
 
    /* TODO do we need to bother to turn anything off? */
@@ -890,9 +877,7 @@ perfcntr_pause(struct fd_acc_query *aq, struct fd_batch *batch) assert_dt
    /* snapshot the end values: */
    for (unsigned i = 0; i < data->num_query_entries; i++) {
       struct fd_batch_query_entry *entry = &data->query_entries[i];
-      const struct fd_perfcntr_group *g = &screen->perfcntr_groups[entry->gid];
-      unsigned counter_idx = counters_per_group[entry->gid]++;
-      const struct fd_perfcntr_counter *counter = &g->counters[counter_idx];
+      const struct fd_perfcntr_counter *counter = entry->counter;
 
       fd_pkt7(cs, CP_REG_TO_MEM, 3)
          .add(CP_REG_TO_MEM_0(.reg = counter->counter_reg_lo, ._64b = true))
@@ -925,12 +910,24 @@ perfcntr_accumulate_result(struct fd_acc_query *aq,
    }
 }
 
+static void
+perfcntr_cleanup(void *query_data)
+{
+   struct fd_batch_query_data *data = (struct fd_batch_query_data *)query_data;
+
+   for (unsigned i = 0; i < data->num_query_entries; i++) {
+      struct fd_batch_query_entry *entry = &data->query_entries[i];
+      fd_perfcntr_release(data->screen->perfcntrs, entry->counter);
+   }
+}
+
 static const struct fd_acc_sample_provider perfcntr = {
    .query_type = FD_QUERY_FIRST_PERFCNTR,
    .always = true,
    .resume = perfcntr_resume,
    .pause = perfcntr_pause,
    .result = perfcntr_accumulate_result,
+   .cleanup = perfcntr_cleanup,
 };
 
 static struct pipe_query *
@@ -949,13 +946,6 @@ fd6_create_batch_query(struct pipe_context *pctx, unsigned num_queries,
    data->screen = screen;
    data->num_query_entries = num_queries;
 
-   /* validate the requested query_types and ensure we don't try
-    * to request more query_types of a given group than we have
-    * counters:
-    */
-   unsigned counters_per_group[screen->num_perfcntr_groups];
-   memset(counters_per_group, 0, sizeof(counters_per_group));
-
    for (unsigned i = 0; i < num_queries; i++) {
       unsigned idx = query_types[i] - FD_QUERY_FIRST_PERFCNTR;
 
@@ -985,13 +975,15 @@ fd6_create_batch_query(struct pipe_context *pctx, unsigned num_queries,
             entry->cid++;
       }
 
-      if (counters_per_group[entry->gid] >=
-          screen->perfcntr_groups[entry->gid].num_counters) {
-         mesa_loge("too many counters for group %u", entry->gid);
+      const struct fd_perfcntr_group *g = &screen->perfcntr_groups[entry->gid];
+      const struct fd_perfcntr_countable *c = &g->countables[entry->cid];
+
+      entry->counter = fd_perfcntr_reserve(screen->perfcntrs, g, c);
+
+      if (!entry->counter) {
+         mesa_loge("Could not reserve counter for %s.%s", g->name, c->name);
          goto error;
       }
-
-      counters_per_group[entry->gid]++;
    }
 
    q = fd_acc_create_query2(ctx, 0, 0, &perfcntr);
@@ -1004,6 +996,7 @@ fd6_create_batch_query(struct pipe_context *pctx, unsigned num_queries,
    return (struct pipe_query *)q;
 
 error:
+   perfcntr_cleanup(data);
    free(data);
    return NULL;
 }

src/gallium/drivers/freedreno/freedreno_query_acc.c

@@ -21,6 +21,9 @@ fd_acc_destroy_query(struct fd_context *ctx, struct fd_query *q) assert_dt
 
    DBG("%p", q);
 
+   if (aq->provider->cleanup)
+      aq->provider->cleanup(aq->query_data);
+
    pipe_resource_reference(&aq->prsc, NULL);
    list_del(&aq->node);
 

src/gallium/drivers/freedreno/freedreno_query_acc.h

@@ -72,6 +72,7 @@ struct fd_acc_sample_provider {
    void (*result_resource)(struct fd_acc_query *aq, struct fd_ringbuffer *ring,
                            enum pipe_query_value_type result_type, int index,
                            struct fd_resource *dst, unsigned offset);
+   void (*cleanup)(void *query_data);  /* optional cleanup */
 };
 
 struct fd_acc_query {

src/gallium/drivers/freedreno/freedreno_screen.c

@@ -165,6 +165,8 @@ fd_screen_destroy(struct pipe_screen *pscreen)
    if (screen->ro)
       screen->ro->destroy(screen->ro);
 
+   fd_perfcntr_state_free(screen->perfcntrs);
+
    fd_bc_fini(&screen->batch_cache);
    fd_gmem_screen_fini(pscreen);
 
@@ -1057,7 +1059,10 @@ fd_screen_create(int fd,
       if (screen->primtypes[i])
          screen->primtypes_mask |= (1 << i);
 
-   if (FD_DBG(PERFC)) {
+   screen->perfcntrs = fd_perfcntr_state_alloc(screen->dev_id, fd);
+
+   if (FD_DBG(PERFC) ||
+       (screen->perfcntrs && fd_perfcntr_has_reservation(screen->perfcntrs))) {
       screen->perfcntr_groups =
          fd_perfcntrs(screen->dev_id, &screen->num_perfcntr_groups);
    }

src/gallium/drivers/freedreno/freedreno_screen.h

@@ -106,6 +106,7 @@ struct fd_screen {
 
    unsigned num_perfcntr_groups;
    const struct fd_perfcntr_group *perfcntr_groups;
+   struct fd_perfcntr_state *perfcntrs;
 
    /* generated at startup from the perfcntr groups: */
    unsigned num_perfcntr_queries;