summaryrefslogtreecommitdiffstats
path: root/src/intel/perf
diff options
context:
space:
mode:
authorMark Janes <[email protected]>2019-07-10 16:57:16 -0700
committerMark Janes <[email protected]>2019-08-07 21:33:56 -0700
commitdeea3798b6e7bc54e62cd6a7b1c08c4e51801144 (patch)
tree76fdf3ae9c8e798ecbcb2efcdb3fd5dedb99b3aa /src/intel/perf
parent1f4f421ce07e16d7bbef16c016fbc1d5b2055516 (diff)
intel/perf: make perf context private
Encapsulate the details of this data structure. Reviewed-by: Kenneth Graunke <[email protected]>
Diffstat (limited to 'src/intel/perf')
-rw-r--r--src/intel/perf/gen_perf.c102
-rw-r--r--src/intel/perf/gen_perf.h71
2 files changed, 109 insertions, 64 deletions
diff --git a/src/intel/perf/gen_perf.c b/src/intel/perf/gen_perf.c
index 0bac1fd2888..03f29d5222f 100644
--- a/src/intel/perf/gen_perf.c
+++ b/src/intel/perf/gen_perf.c
@@ -189,6 +189,86 @@ struct oa_sample_buf {
uint32_t last_timestamp;
};
+struct gen_perf_context {
+ struct gen_perf_config *perf;
+
+ void * ctx; /* driver context (eg, brw_context) */
+ void * bufmgr;
+ const struct gen_device_info *devinfo;
+
+ uint32_t hw_ctx;
+ int drm_fd;
+
+ /* The i915 perf stream we open to setup + enable the OA counters */
+ int oa_stream_fd;
+
+ /* An i915 perf stream fd gives exclusive access to the OA unit that will
+ * report counter snapshots for a specific counter set/profile in a
+ * specific layout/format so we can only start OA queries that are
+ * compatible with the currently open fd...
+ */
+ int current_oa_metrics_set_id;
+ int current_oa_format;
+
+ /* List of buffers containing OA reports */
+ struct exec_list sample_buffers;
+
+ /* Cached list of empty sample buffers */
+ struct exec_list free_sample_buffers;
+
+ int n_active_oa_queries;
+ int n_active_pipeline_stats_queries;
+
+ /* The number of queries depending on running OA counters which
+ * extends beyond brw_end_perf_query() since we need to wait until
+ * the last MI_RPC command has parsed by the GPU.
+ *
+ * Accurate accounting is important here as emitting an
+ * MI_REPORT_PERF_COUNT command while the OA unit is disabled will
+ * effectively hang the gpu.
+ */
+ int n_oa_users;
+
+ /* To help catch an spurious problem with the hardware or perf
+ * forwarding samples, we emit each MI_REPORT_PERF_COUNT command
+ * with a unique ID that we can explicitly check for...
+ */
+ int next_query_start_report_id;
+
+ /**
+ * An array of queries whose results haven't yet been assembled
+ * based on the data in buffer objects.
+ *
+ * These may be active, or have already ended. However, the
+ * results have not been requested.
+ */
+ struct gen_perf_query_object **unaccumulated;
+ int unaccumulated_elements;
+ int unaccumulated_array_size;
+
+ /* The total number of query objects so we can relinquish
+ * our exclusive access to perf if the application deletes
+ * all of its objects. (NB: We only disable perf while
+ * there are no active queries)
+ */
+ int n_query_instances;
+};
+
+struct gen_perf_context *
+gen_perf_new_context(void *parent)
+{
+ struct gen_perf_context *ctx = rzalloc(parent, struct gen_perf_context);
+ if (! ctx)
+ fprintf(stderr, "%s: failed to alloc context\n", __func__);
+ return ctx;
+}
+
+struct gen_perf_config *
+gen_perf_config(struct gen_perf_context *ctx)
+{
+ return ctx->perf;
+}
+
struct gen_perf_query_object *
gen_perf_new_query(struct gen_perf_context *perf_ctx, unsigned query_index)
{
@@ -206,6 +286,28 @@ gen_perf_new_query(struct gen_perf_context *perf_ctx, unsigned query_index)
return obj;
}
+int
+gen_perf_active_queries(struct gen_perf_context *perf_ctx,
+ const struct gen_perf_query_info *query)
+{
+ assert(perf_ctx->n_active_oa_queries == 0 || perf_ctx->n_active_pipeline_stats_queries == 0);
+
+ switch (query->kind) {
+ case GEN_PERF_QUERY_TYPE_OA:
+ case GEN_PERF_QUERY_TYPE_RAW:
+ return perf_ctx->n_active_oa_queries;
+ break;
+
+ case GEN_PERF_QUERY_TYPE_PIPELINE:
+ return perf_ctx->n_active_pipeline_stats_queries;
+ break;
+
+ default:
+ unreachable("Unknown query type");
+ break;
+ }
+}
+
static bool
get_sysfs_dev_dir(struct gen_perf_config *perf, int fd)
{
diff --git a/src/intel/perf/gen_perf.h b/src/intel/perf/gen_perf.h
index d4178900714..1038d5018a9 100644
--- a/src/intel/perf/gen_perf.h
+++ b/src/intel/perf/gen_perf.h
@@ -321,70 +321,8 @@ struct gen_perf_query_object
};
};
-struct gen_perf_context {
- struct gen_perf_config *perf;
-
- void * ctx; /* driver context (eg, brw_context) */
- void * bufmgr;
- const struct gen_device_info *devinfo;
-
- uint32_t hw_ctx;
- int drm_fd;
-
- /* The i915 perf stream we open to setup + enable the OA counters */
- int oa_stream_fd;
-
- /* An i915 perf stream fd gives exclusive access to the OA unit that will
- * report counter snapshots for a specific counter set/profile in a
- * specific layout/format so we can only start OA queries that are
- * compatible with the currently open fd...
- */
- int current_oa_metrics_set_id;
- int current_oa_format;
-
- /* List of buffers containing OA reports */
- struct exec_list sample_buffers;
-
- /* Cached list of empty sample buffers */
- struct exec_list free_sample_buffers;
-
- int n_active_oa_queries;
- int n_active_pipeline_stats_queries;
-
- /* The number of queries depending on running OA counters which
- * extends beyond brw_end_perf_query() since we need to wait until
- * the last MI_RPC command has parsed by the GPU.
- *
- * Accurate accounting is important here as emitting an
- * MI_REPORT_PERF_COUNT command while the OA unit is disabled will
- * effectively hang the gpu.
- */
- int n_oa_users;
-
- /* To help catch an spurious problem with the hardware or perf
- * forwarding samples, we emit each MI_REPORT_PERF_COUNT command
- * with a unique ID that we can explicitly check for...
- */
- int next_query_start_report_id;
-
- /**
- * An array of queries whose results haven't yet been assembled
- * based on the data in buffer objects.
- *
- * These may be active, or have already ended. However, the
- * results have not been requested.
- */
- struct gen_perf_query_object **unaccumulated;
- int unaccumulated_elements;
- int unaccumulated_array_size;
-
- /* The total number of query objects so we can relinquish
- * our exclusive access to perf if the application deletes
- * all of its objects. (NB: We only disable perf while
- * there are no active queries)
- */
- int n_query_instances;
-};
+struct gen_perf_context;
+struct gen_perf_context *gen_perf_new_context(void *parent);
void gen_perf_init_metrics(struct gen_perf_config *perf_cfg,
const struct gen_device_info *devinfo,
@@ -397,6 +335,11 @@ void gen_perf_init_context(struct gen_perf_context *perf_ctx,
uint32_t hw_ctx,
int drm_fd);
+struct gen_perf_config *gen_perf_config(struct gen_perf_context *ctx);
+
+int gen_perf_active_queries(struct gen_perf_context *perf_ctx,
+ const struct gen_perf_query_info *query);
+
static inline size_t
gen_perf_query_counter_get_size(const struct gen_perf_query_counter *counter)
{