#define MI_RPC_BO_END_OFFSET_BYTES (MI_RPC_BO_SIZE / 2)
#define MI_FREQ_END_OFFSET_BYTES (3076)
+#define INTEL_MASK(high, low) (((1u<<((high)-(low)+1))-1)<<(low))
+
+#define GEN7_RPSTAT1 0xA01C
+#define GEN7_RPSTAT1_CURR_GT_FREQ_SHIFT 7
+#define GEN7_RPSTAT1_CURR_GT_FREQ_MASK INTEL_MASK(13, 7)
+#define GEN7_RPSTAT1_PREV_GT_FREQ_SHIFT 0
+#define GEN7_RPSTAT1_PREV_GT_FREQ_MASK INTEL_MASK(6, 0)
+
+#define GEN9_RPSTAT0 0xA01C
+#define GEN9_RPSTAT0_CURR_GT_FREQ_SHIFT 23
+#define GEN9_RPSTAT0_CURR_GT_FREQ_MASK INTEL_MASK(31, 23)
+#define GEN9_RPSTAT0_PREV_GT_FREQ_SHIFT 0
+#define GEN9_RPSTAT0_PREV_GT_FREQ_MASK INTEL_MASK(8, 0)
+
+#define GEN6_SO_PRIM_STORAGE_NEEDED 0x2280
+#define GEN7_SO_PRIM_STORAGE_NEEDED(n) (0x5240 + (n) * 8)
+#define GEN6_SO_NUM_PRIMS_WRITTEN 0x2288
+#define GEN7_SO_NUM_PRIMS_WRITTEN(n) (0x5200 + (n) * 8)
+
#define MAP_READ (1 << 0)
#define MAP_WRITE (1 << 1)
+/**
+ * Periodic OA samples are read() into these buffer structures via the
+ * i915 perf kernel interface and appended to the
+ * perf_ctx->sample_buffers linked list. When we process the
+ * results of an OA metrics query we need to consider all the periodic
+ * samples between the Begin and End MI_REPORT_PERF_COUNT command
+ * markers.
+ *
+ * 'Periodic' is a simplification as there are other automatic reports
+ * written by the hardware also buffered here.
+ *
+ * Considering three queries, A, B and C:
+ *
+ * Time ---->
+ * ________________A_________________
+ * | |
+ * | ________B_________ _____C___________
+ * | | | | | |
+ *
+ * And an illustration of sample buffers read over this time frame:
+ * [HEAD ][ ][ ][ ][ ][ ][ ][ ][TAIL ]
+ *
+ * These nodes may hold samples for query A:
+ * [ ][ ][ A ][ A ][ A ][ A ][ A ][ ][ ]
+ *
+ * These nodes may hold samples for query B:
+ * [ ][ ][ B ][ B ][ B ][ ][ ][ ][ ]
+ *
+ * These nodes may hold samples for query C:
+ * [ ][ ][ ][ ][ ][ C ][ C ][ C ][ ]
+ *
+ * The illustration assumes we have an even distribution of periodic
+ * samples so all nodes have the same size plotted against time:
+ *
+ * Note, to simplify code, the list is never empty.
+ *
+ * With overlapping queries we can see that periodic OA reports may
+ * relate to multiple queries and care needs to be take to keep
+ * track of sample buffers until there are no queries that might
+ * depend on their contents.
+ *
+ * We use a node ref counting system where a reference ensures that a
+ * node and all following nodes can't be freed/recycled until the
+ * reference drops to zero.
+ *
+ * E.g. with a ref of one here:
+ * [ 0 ][ 0 ][ 1 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ]
+ *
+ * These nodes could be freed or recycled ("reaped"):
+ * [ 0 ][ 0 ]
+ *
+ * These must be preserved until the leading ref drops to zero:
+ * [ 1 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ]
+ *
+ * When a query starts we take a reference on the current tail of
+ * the list, knowing that no already-buffered samples can possibly
+ * relate to the newly-started query. A pointer to this node is
+ * also saved in the query object's ->oa.samples_head.
+ *
+ * E.g. starting query A while there are two nodes in .sample_buffers:
+ * ________________A________
+ * |
+ *
+ * [ 0 ][ 1 ]
+ * ^_______ Add a reference and store pointer to node in
+ * A->oa.samples_head
+ *
+ * Moving forward to when the B query starts with no new buffer nodes:
+ * (for reference, i915 perf reads() are only done when queries finish)
+ * ________________A_______
+ * | ________B___
+ * | |
+ *
+ * [ 0 ][ 2 ]
+ * ^_______ Add a reference and store pointer to
+ * node in B->oa.samples_head
+ *
+ * Once a query is finished, after an OA query has become 'Ready',
+ * once the End OA report has landed and after we we have processed
+ * all the intermediate periodic samples then we drop the
+ * ->oa.samples_head reference we took at the start.
+ *
+ * So when the B query has finished we have:
+ * ________________A________
+ * | ______B___________
+ * | | |
+ * [ 0 ][ 1 ][ 0 ][ 0 ][ 0 ]
+ * ^_______ Drop B->oa.samples_head reference
+ *
+ * We still can't free these due to the A->oa.samples_head ref:
+ * [ 1 ][ 0 ][ 0 ][ 0 ]
+ *
+ * When the A query finishes: (note there's a new ref for C's samples_head)
+ * ________________A_________________
+ * | |
+ * | _____C_________
+ * | | |
+ * [ 0 ][ 0 ][ 0 ][ 0 ][ 1 ][ 0 ][ 0 ]
+ * ^_______ Drop A->oa.samples_head reference
+ *
+ * And we can now reap these nodes up to the C->oa.samples_head:
+ * [ X ][ X ][ X ][ X ]
+ * keeping -> [ 1 ][ 0 ][ 0 ]
+ *
+ * We reap old sample buffers each time we finish processing an OA
+ * query by iterating the sample_buffers list from the head until we
+ * find a referenced node and stop.
+ *
+ * Reaped buffers move to a perfquery.free_sample_buffers list and
+ * when we come to read() we first look to recycle a buffer from the
+ * free_sample_buffers list before allocating a new buffer.
+ */
+struct oa_sample_buf {
+ struct exec_node link;
+ int refcount;
+ int len;
+ uint8_t buf[I915_PERF_OA_SAMPLE_SIZE * 10];
+ uint32_t last_timestamp;
+};
+
+struct gen_perf_query_object *
+gen_perf_new_query(struct gen_perf_context *perf_ctx, unsigned query_index)
+{
+ const struct gen_perf_query_info *query =
+ &perf_ctx->perf->queries[query_index];
+ struct gen_perf_query_object *obj =
+ calloc(1, sizeof(struct gen_perf_query_object));
+
+ if (!obj)
+ return NULL;
+
+ obj->queryinfo = query;
+
+ perf_ctx->n_query_instances++;
+ return obj;
+}
+
static bool
get_sysfs_dev_dir(struct gen_perf_config *perf, int fd)
{
return read_file_uint64(buf, value);
}
+static inline struct gen_perf_query_info *
+append_query_info(struct gen_perf_config *perf, int max_counters)
+{
+ struct gen_perf_query_info *query;
+
+ perf->queries = reralloc(perf, perf->queries,
+ struct gen_perf_query_info,
+ ++perf->n_queries);
+ query = &perf->queries[perf->n_queries - 1];
+ memset(query, 0, sizeof(*query));
+
+ if (max_counters > 0) {
+ query->max_counters = max_counters;
+ query->counters =
+ rzalloc_array(perf, struct gen_perf_query_counter, max_counters);
+ }
+
+ return query;
+}
+
static void
register_oa_config(struct gen_perf_config *perf,
const struct gen_perf_query_info *query,
uint64_t config_id)
{
- struct gen_perf_query_info *registred_query =
- gen_perf_query_append_query_info(perf, 0);
+ struct gen_perf_query_info *registred_query = append_query_info(perf, 0);
*registred_query = *query;
registred_query->oa_metrics_set_id = config_id;
return NULL;
}
-bool
-gen_perf_load_oa_metrics(struct gen_perf_config *perf, int fd,
+static inline void
+add_stat_reg(struct gen_perf_query_info *query, uint32_t reg,
+ uint32_t numerator, uint32_t denominator,
+ const char *name, const char *description)
+{
+ struct gen_perf_query_counter *counter;
+
+ assert(query->n_counters < query->max_counters);
+
+ counter = &query->counters[query->n_counters];
+ counter->name = name;
+ counter->desc = description;
+ counter->type = GEN_PERF_COUNTER_TYPE_RAW;
+ counter->data_type = GEN_PERF_COUNTER_DATA_TYPE_UINT64;
+ counter->offset = sizeof(uint64_t) * query->n_counters;
+ counter->pipeline_stat.reg = reg;
+ counter->pipeline_stat.numerator = numerator;
+ counter->pipeline_stat.denominator = denominator;
+
+ query->n_counters++;
+}
+
+static inline void
+add_basic_stat_reg(struct gen_perf_query_info *query,
+ uint32_t reg, const char *name)
+{
+ add_stat_reg(query, reg, 1, 1, name, name);
+}
+
+static void
+load_pipeline_statistic_metrics(struct gen_perf_config *perf_cfg,
+ const struct gen_device_info *devinfo)
+{
+ struct gen_perf_query_info *query =
+ append_query_info(perf_cfg, MAX_STAT_COUNTERS);
+
+ query->kind = GEN_PERF_QUERY_TYPE_PIPELINE;
+ query->name = "Pipeline Statistics Registers";
+
+ add_basic_stat_reg(query, IA_VERTICES_COUNT,
+ "N vertices submitted");
+ add_basic_stat_reg(query, IA_PRIMITIVES_COUNT,
+ "N primitives submitted");
+ add_basic_stat_reg(query, VS_INVOCATION_COUNT,
+ "N vertex shader invocations");
+
+ if (devinfo->gen == 6) {
+ add_stat_reg(query, GEN6_SO_PRIM_STORAGE_NEEDED, 1, 1,
+ "SO_PRIM_STORAGE_NEEDED",
+ "N geometry shader stream-out primitives (total)");
+ add_stat_reg(query, GEN6_SO_NUM_PRIMS_WRITTEN, 1, 1,
+ "SO_NUM_PRIMS_WRITTEN",
+ "N geometry shader stream-out primitives (written)");
+ } else {
+ add_stat_reg(query, GEN7_SO_PRIM_STORAGE_NEEDED(0), 1, 1,
+ "SO_PRIM_STORAGE_NEEDED (Stream 0)",
+ "N stream-out (stream 0) primitives (total)");
+ add_stat_reg(query, GEN7_SO_PRIM_STORAGE_NEEDED(1), 1, 1,
+ "SO_PRIM_STORAGE_NEEDED (Stream 1)",
+ "N stream-out (stream 1) primitives (total)");
+ add_stat_reg(query, GEN7_SO_PRIM_STORAGE_NEEDED(2), 1, 1,
+ "SO_PRIM_STORAGE_NEEDED (Stream 2)",
+ "N stream-out (stream 2) primitives (total)");
+ add_stat_reg(query, GEN7_SO_PRIM_STORAGE_NEEDED(3), 1, 1,
+ "SO_PRIM_STORAGE_NEEDED (Stream 3)",
+ "N stream-out (stream 3) primitives (total)");
+ add_stat_reg(query, GEN7_SO_NUM_PRIMS_WRITTEN(0), 1, 1,
+ "SO_NUM_PRIMS_WRITTEN (Stream 0)",
+ "N stream-out (stream 0) primitives (written)");
+ add_stat_reg(query, GEN7_SO_NUM_PRIMS_WRITTEN(1), 1, 1,
+ "SO_NUM_PRIMS_WRITTEN (Stream 1)",
+ "N stream-out (stream 1) primitives (written)");
+ add_stat_reg(query, GEN7_SO_NUM_PRIMS_WRITTEN(2), 1, 1,
+ "SO_NUM_PRIMS_WRITTEN (Stream 2)",
+ "N stream-out (stream 2) primitives (written)");
+ add_stat_reg(query, GEN7_SO_NUM_PRIMS_WRITTEN(3), 1, 1,
+ "SO_NUM_PRIMS_WRITTEN (Stream 3)",
+ "N stream-out (stream 3) primitives (written)");
+ }
+
+ add_basic_stat_reg(query, HS_INVOCATION_COUNT,
+ "N TCS shader invocations");
+ add_basic_stat_reg(query, DS_INVOCATION_COUNT,
+ "N TES shader invocations");
+
+ add_basic_stat_reg(query, GS_INVOCATION_COUNT,
+ "N geometry shader invocations");
+ add_basic_stat_reg(query, GS_PRIMITIVES_COUNT,
+ "N geometry shader primitives emitted");
+
+ add_basic_stat_reg(query, CL_INVOCATION_COUNT,
+ "N primitives entering clipping");
+ add_basic_stat_reg(query, CL_PRIMITIVES_COUNT,
+ "N primitives leaving clipping");
+
+ if (devinfo->is_haswell || devinfo->gen == 8) {
+ add_stat_reg(query, PS_INVOCATION_COUNT, 1, 4,
+ "N fragment shader invocations",
+ "N fragment shader invocations");
+ } else {
+ add_basic_stat_reg(query, PS_INVOCATION_COUNT,
+ "N fragment shader invocations");
+ }
+
+ add_basic_stat_reg(query, PS_DEPTH_COUNT,
+ "N z-pass fragments");
+
+ if (devinfo->gen >= 7) {
+ add_basic_stat_reg(query, CS_INVOCATION_COUNT,
+ "N compute shader invocations");
+ }
+
+ query->data_size = sizeof(uint64_t) * query->n_counters;
+}
+
+static bool
+load_oa_metrics(struct gen_perf_config *perf, int fd,
const struct gen_device_info *devinfo)
{
perf_register_oa_queries_t oa_register = get_register_queries_function(devinfo);
result->hw_id = 0xffffffff; /* invalid */
}
+static void
+gen_perf_query_register_mdapi_statistic_query(struct gen_perf_config *perf_cfg,
+ const struct gen_device_info *devinfo)
+{
+ if (!(devinfo->gen >= 7 && devinfo->gen <= 11))
+ return;
+
+ struct gen_perf_query_info *query =
+ append_query_info(perf_cfg, MAX_STAT_COUNTERS);
+
+ query->kind = GEN_PERF_QUERY_TYPE_PIPELINE;
+ query->name = "Intel_Raw_Pipeline_Statistics_Query";
+
+ /* The order has to match mdapi_pipeline_metrics. */
+ add_basic_stat_reg(query, IA_VERTICES_COUNT,
+ "N vertices submitted");
+ add_basic_stat_reg(query, IA_PRIMITIVES_COUNT,
+ "N primitives submitted");
+ add_basic_stat_reg(query, VS_INVOCATION_COUNT,
+ "N vertex shader invocations");
+ add_basic_stat_reg(query, GS_INVOCATION_COUNT,
+ "N geometry shader invocations");
+ add_basic_stat_reg(query, GS_PRIMITIVES_COUNT,
+ "N geometry shader primitives emitted");
+ add_basic_stat_reg(query, CL_INVOCATION_COUNT,
+ "N primitives entering clipping");
+ add_basic_stat_reg(query, CL_PRIMITIVES_COUNT,
+ "N primitives leaving clipping");
+ if (devinfo->is_haswell || devinfo->gen == 8) {
+ add_stat_reg(query, PS_INVOCATION_COUNT, 1, 4,
+ "N fragment shader invocations",
+ "N fragment shader invocations");
+ } else {
+ add_basic_stat_reg(query, PS_INVOCATION_COUNT,
+ "N fragment shader invocations");
+ }
+ add_basic_stat_reg(query, HS_INVOCATION_COUNT,
+ "N TCS shader invocations");
+ add_basic_stat_reg(query, DS_INVOCATION_COUNT,
+ "N TES shader invocations");
+ if (devinfo->gen >= 7) {
+ add_basic_stat_reg(query, CS_INVOCATION_COUNT,
+ "N compute shader invocations");
+ }
+
+ if (devinfo->gen >= 10) {
+ /* Reuse existing CS invocation register until we can expose this new
+ * one.
+ */
+ add_basic_stat_reg(query, CS_INVOCATION_COUNT,
+ "Reserved1");
+ }
+
+ query->data_size = sizeof(uint64_t) * query->n_counters;
+}
+
static void
fill_mdapi_perf_query_counter(struct gen_perf_query_info *query,
const char *name,
sizeof(struct_name.field_name[0]), \
GEN_PERF_COUNTER_DATA_TYPE_##type_name)
-void
-gen_perf_query_register_mdapi_oa_query(const struct gen_device_info *devinfo,
- struct gen_perf_config *perf)
+static void
+register_mdapi_oa_query(const struct gen_device_info *devinfo,
+ struct gen_perf_config *perf)
{
struct gen_perf_query_info *query = NULL;
switch (devinfo->gen) {
case 7: {
- query = gen_perf_query_append_query_info(perf, 1 + 45 + 16 + 7);
+ query = append_query_info(perf, 1 + 45 + 16 + 7);
query->oa_format = I915_OA_FORMAT_A45_B8_C8;
struct gen7_mdapi_metrics metric_data;
break;
}
case 8: {
- query = gen_perf_query_append_query_info(perf, 2 + 36 + 16 + 16);
+ query = append_query_info(perf, 2 + 36 + 16 + 16);
query->oa_format = I915_OA_FORMAT_A32u40_A4u32_B8_C8;
struct gen8_mdapi_metrics metric_data;
case 9:
case 10:
case 11: {
- query = gen_perf_query_append_query_info(perf, 2 + 36 + 16 + 16 + 16 + 2);
+ query = append_query_info(perf, 2 + 36 + 16 + 16 + 16 + 2);
query->oa_format = I915_OA_FORMAT_A32u40_A4u32_B8_C8;
struct gen9_mdapi_metrics metric_data;
}
}
-void
-gen_perf_query_register_mdapi_statistic_query(const struct gen_device_info *devinfo,
- struct gen_perf_config *perf)
-{
- if (!(devinfo->gen >= 7 && devinfo->gen <= 11))
- return;
-
- struct gen_perf_query_info *query =
- gen_perf_query_append_query_info(perf, MAX_STAT_COUNTERS);
-
- query->kind = GEN_PERF_QUERY_TYPE_PIPELINE;
- query->name = "Intel_Raw_Pipeline_Statistics_Query";
-
- /* The order has to match mdapi_pipeline_metrics. */
- gen_perf_query_info_add_basic_stat_reg(query, IA_VERTICES_COUNT,
- "N vertices submitted");
- gen_perf_query_info_add_basic_stat_reg(query, IA_PRIMITIVES_COUNT,
- "N primitives submitted");
- gen_perf_query_info_add_basic_stat_reg(query, VS_INVOCATION_COUNT,
- "N vertex shader invocations");
- gen_perf_query_info_add_basic_stat_reg(query, GS_INVOCATION_COUNT,
- "N geometry shader invocations");
- gen_perf_query_info_add_basic_stat_reg(query, GS_PRIMITIVES_COUNT,
- "N geometry shader primitives emitted");
- gen_perf_query_info_add_basic_stat_reg(query, CL_INVOCATION_COUNT,
- "N primitives entering clipping");
- gen_perf_query_info_add_basic_stat_reg(query, CL_PRIMITIVES_COUNT,
- "N primitives leaving clipping");
- if (devinfo->is_haswell || devinfo->gen == 8) {
- gen_perf_query_info_add_stat_reg(query, PS_INVOCATION_COUNT, 1, 4,
- "N fragment shader invocations",
- "N fragment shader invocations");
- } else {
- gen_perf_query_info_add_basic_stat_reg(query, PS_INVOCATION_COUNT,
- "N fragment shader invocations");
- }
- gen_perf_query_info_add_basic_stat_reg(query, HS_INVOCATION_COUNT,
- "N TCS shader invocations");
- gen_perf_query_info_add_basic_stat_reg(query, DS_INVOCATION_COUNT,
- "N TES shader invocations");
- if (devinfo->gen >= 7) {
- gen_perf_query_info_add_basic_stat_reg(query, CS_INVOCATION_COUNT,
- "N compute shader invocations");
- }
-
- if (devinfo->gen >= 10) {
- /* Reuse existing CS invocation register until we can expose this new
- * one.
- */
- gen_perf_query_info_add_basic_stat_reg(query, CS_INVOCATION_COUNT,
- "Reserved1");
- }
-
- query->data_size = sizeof(uint64_t) * query->n_counters;
-}
-
uint64_t
gen_perf_query_get_metric_id(struct gen_perf_config *perf,
const struct gen_perf_query_info *query)
}
}
+void
+gen_perf_init_metrics(struct gen_perf_config *perf_cfg,
+ const struct gen_device_info *devinfo,
+ int drm_fd)
+{
+ load_pipeline_statistic_metrics(perf_cfg, devinfo);
+ gen_perf_query_register_mdapi_statistic_query(perf_cfg, devinfo);
+ if (load_oa_metrics(perf_cfg, drm_fd, devinfo))
+ register_mdapi_oa_query(devinfo, perf_cfg);
+}
+
void
gen_perf_init_context(struct gen_perf_context *perf_ctx,
struct gen_perf_config *perf_cfg,
return false;
}
+
+/**
+ * Remove a query from the global list of unaccumulated queries once
+ * after successfully accumulating the OA reports associated with the
+ * query in accumulate_oa_reports() or when discarding unwanted query
+ * results.
+ */
+static void
+drop_from_unaccumulated_query_list(struct gen_perf_context *perf_ctx,
+ struct gen_perf_query_object *query)
+{
+ for (int i = 0; i < perf_ctx->unaccumulated_elements; i++) {
+ if (perf_ctx->unaccumulated[i] == query) {
+ int last_elt = --perf_ctx->unaccumulated_elements;
+
+ if (i == last_elt)
+ perf_ctx->unaccumulated[i] = NULL;
+ else {
+ perf_ctx->unaccumulated[i] =
+ perf_ctx->unaccumulated[last_elt];
+ }
+
+ break;
+ }
+ }
+
+ /* Drop our samples_head reference so that associated periodic
+ * sample data buffers can potentially be reaped if they aren't
+ * referenced by any other queries...
+ */
+
+ struct oa_sample_buf *buf =
+ exec_node_data(struct oa_sample_buf, query->oa.samples_head, link);
+
+ assert(buf->refcount > 0);
+ buf->refcount--;
+
+ query->oa.samples_head = NULL;
+
+ gen_perf_reap_old_sample_buffers(perf_ctx);
+}
+
+/* In general if we see anything spurious while accumulating results,
+ * we don't try and continue accumulating the current query, hoping
+ * for the best, we scrap anything outstanding, and then hope for the
+ * best with new queries.
+ */
+static void
+discard_all_queries(struct gen_perf_context *perf_ctx)
+{
+ while (perf_ctx->unaccumulated_elements) {
+ struct gen_perf_query_object *query = perf_ctx->unaccumulated[0];
+
+ query->oa.results_accumulated = true;
+ drop_from_unaccumulated_query_list(perf_ctx, query);
+
+ gen_perf_dec_n_users(perf_ctx);
+ }
+}
+
+/**
+ * Accumulate raw OA counter values based on deltas between pairs of
+ * OA reports.
+ *
+ * Accumulation starts from the first report captured via
+ * MI_REPORT_PERF_COUNT (MI_RPC) by brw_begin_perf_query() until the
+ * last MI_RPC report requested by brw_end_perf_query(). Between these
+ * two reports there may also some number of periodically sampled OA
+ * reports collected via the i915 perf interface - depending on the
+ * duration of the query.
+ *
+ * These periodic snapshots help to ensure we handle counter overflow
+ * correctly by being frequent enough to ensure we don't miss multiple
+ * overflows of a counter between snapshots. For Gen8+ the i915 perf
+ * snapshots provide the extra context-switch reports that let us
+ * subtract out the progress of counters associated with other
+ * contexts running on the system.
+ */
+static void
+accumulate_oa_reports(struct gen_perf_context *perf_ctx,
+ struct gen_perf_query_object *query)
+{
+ const struct gen_device_info *devinfo = perf_ctx->devinfo;
+ uint32_t *start;
+ uint32_t *last;
+ uint32_t *end;
+ struct exec_node *first_samples_node;
+ bool in_ctx = true;
+ int out_duration = 0;
+
+ assert(query->oa.map != NULL);
+
+ start = last = query->oa.map;
+ end = query->oa.map + MI_RPC_BO_END_OFFSET_BYTES;
+
+ if (start[0] != query->oa.begin_report_id) {
+ DBG("Spurious start report id=%"PRIu32"\n", start[0]);
+ goto error;
+ }
+ if (end[0] != (query->oa.begin_report_id + 1)) {
+ DBG("Spurious end report id=%"PRIu32"\n", end[0]);
+ goto error;
+ }
+
+ /* See if we have any periodic reports to accumulate too... */
+
+ /* N.B. The oa.samples_head was set when the query began and
+ * pointed to the tail of the perf_ctx->sample_buffers list at
+ * the time the query started. Since the buffer existed before the
+ * first MI_REPORT_PERF_COUNT command was emitted we therefore know
+ * that no data in this particular node's buffer can possibly be
+ * associated with the query - so skip ahead one...
+ */
+ first_samples_node = query->oa.samples_head->next;
+
+ foreach_list_typed_from(struct oa_sample_buf, buf, link,
+ &perf_ctx.sample_buffers,
+ first_samples_node)
+ {
+ int offset = 0;
+
+ while (offset < buf->len) {
+ const struct drm_i915_perf_record_header *header =
+ (const struct drm_i915_perf_record_header *)(buf->buf + offset);
+
+ assert(header->size != 0);
+ assert(header->size <= buf->len);
+
+ offset += header->size;
+
+ switch (header->type) {
+ case DRM_I915_PERF_RECORD_SAMPLE: {
+ uint32_t *report = (uint32_t *)(header + 1);
+ bool add = true;
+
+ /* Ignore reports that come before the start marker.
+ * (Note: takes care to allow overflow of 32bit timestamps)
+ */
+ if (gen_device_info_timebase_scale(devinfo,
+ report[1] - start[1]) > 5000000000) {
+ continue;
+ }
+
+ /* Ignore reports that come after the end marker.
+ * (Note: takes care to allow overflow of 32bit timestamps)
+ */
+ if (gen_device_info_timebase_scale(devinfo,
+ report[1] - end[1]) <= 5000000000) {
+ goto end;
+ }
+
+ /* For Gen8+ since the counters continue while other
+ * contexts are running we need to discount any unrelated
+ * deltas. The hardware automatically generates a report
+ * on context switch which gives us a new reference point
+ * to continuing adding deltas from.
+ *
+ * For Haswell we can rely on the HW to stop the progress
+ * of OA counters while any other context is acctive.
+ */
+ if (devinfo->gen >= 8) {
+ if (in_ctx && report[2] != query->oa.result.hw_id) {
+ DBG("i915 perf: Switch AWAY (observed by ID change)\n");
+ in_ctx = false;
+ out_duration = 0;
+ } else if (in_ctx == false && report[2] == query->oa.result.hw_id) {
+ DBG("i915 perf: Switch TO\n");
+ in_ctx = true;
+
+ /* From experimentation in IGT, we found that the OA unit
+ * might label some report as "idle" (using an invalid
+ * context ID), right after a report for a given context.
+ * Deltas generated by those reports actually belong to the
+ * previous context, even though they're not labelled as
+ * such.
+ *
+ * We didn't *really* Switch AWAY in the case that we e.g.
+ * saw a single periodic report while idle...
+ */
+ if (out_duration >= 1)
+ add = false;
+ } else if (in_ctx) {
+ assert(report[2] == query->oa.result.hw_id);
+ DBG("i915 perf: Continuation IN\n");
+ } else {
+ assert(report[2] != query->oa.result.hw_id);
+ DBG("i915 perf: Continuation OUT\n");
+ add = false;
+ out_duration++;
+ }
+ }
+
+ if (add) {
+ gen_perf_query_result_accumulate(&query->oa.result, query->queryinfo,
+ last, report);
+ }
+
+ last = report;
+
+ break;
+ }
+
+ case DRM_I915_PERF_RECORD_OA_BUFFER_LOST:
+ DBG("i915 perf: OA error: all reports lost\n");
+ goto error;
+ case DRM_I915_PERF_RECORD_OA_REPORT_LOST:
+ DBG("i915 perf: OA report lost\n");
+ break;
+ }
+ }
+ }
+
+end:
+
+ gen_perf_query_result_accumulate(&query->oa.result, query->queryinfo,
+ last, end);
+
+ query->oa.results_accumulated = true;
+ drop_from_unaccumulated_query_list(perf_ctx, query);
+ gen_perf_dec_n_users(perf_ctx);
+
+ return;
+
+error:
+
+ discard_all_queries(perf_ctx);
+}
+
+void
+gen_perf_delete_query(struct gen_perf_context *perf_ctx,
+ struct gen_perf_query_object *query)
+{
+ struct gen_perf_config *perf_cfg = perf_ctx->perf;
+
+ /* We can assume that the frontend waits for a query to complete
+ * before ever calling into here, so we don't have to worry about
+ * deleting an in-flight query object.
+ */
+ switch (query->queryinfo->kind) {
+ case GEN_PERF_QUERY_TYPE_OA:
+ case GEN_PERF_QUERY_TYPE_RAW:
+ if (query->oa.bo) {
+ if (!query->oa.results_accumulated) {
+ drop_from_unaccumulated_query_list(perf_ctx, query);
+ gen_perf_dec_n_users(perf_ctx);
+ }
+
+ perf_cfg->vtbl.bo_unreference(query->oa.bo);
+ query->oa.bo = NULL;
+ }
+
+ query->oa.results_accumulated = false;
+ break;
+
+ case GEN_PERF_QUERY_TYPE_PIPELINE:
+ if (query->pipeline_stats.bo) {
+ perf_cfg->vtbl.bo_unreference(query->pipeline_stats.bo);
+ query->pipeline_stats.bo = NULL;
+ }
+ break;
+
+ default:
+ unreachable("Unknown query type");
+ break;
+ }
+
+ /* As an indication that the INTEL_performance_query extension is no
+ * longer in use, it's a good time to free our cache of sample
+ * buffers and close any current i915-perf stream.
+ */
+ if (--perf_ctx->n_query_instances == 0) {
+ gen_perf_free_sample_bufs(perf_ctx);
+ gen_perf_close(perf_ctx, query->queryinfo);
+ }
+
+ free(query);
+}
+
+#define GET_FIELD(word, field) (((word) & field ## _MASK) >> field ## _SHIFT)
+
+static void
+read_gt_frequency(struct gen_perf_context *perf_ctx,
+ struct gen_perf_query_object *obj)
+{
+ const struct gen_device_info *devinfo = perf_ctx->devinfo;
+ uint32_t start = *((uint32_t *)(obj->oa.map + MI_FREQ_START_OFFSET_BYTES)),
+ end = *((uint32_t *)(obj->oa.map + MI_FREQ_END_OFFSET_BYTES));
+
+ switch (devinfo->gen) {
+ case 7:
+ case 8:
+ obj->oa.gt_frequency[0] = GET_FIELD(start, GEN7_RPSTAT1_CURR_GT_FREQ) * 50ULL;
+ obj->oa.gt_frequency[1] = GET_FIELD(end, GEN7_RPSTAT1_CURR_GT_FREQ) * 50ULL;
+ break;
+ case 9:
+ case 10:
+ case 11:
+ obj->oa.gt_frequency[0] = GET_FIELD(start, GEN9_RPSTAT0_CURR_GT_FREQ) * 50ULL / 3ULL;
+ obj->oa.gt_frequency[1] = GET_FIELD(end, GEN9_RPSTAT0_CURR_GT_FREQ) * 50ULL / 3ULL;
+ break;
+ default:
+ unreachable("unexpected gen");
+ }
+
+ /* Put the numbers into Hz. */
+ obj->oa.gt_frequency[0] *= 1000000ULL;
+ obj->oa.gt_frequency[1] *= 1000000ULL;
+}
+
+static int
+get_oa_counter_data(struct gen_perf_context *perf_ctx,
+ struct gen_perf_query_object *query,
+ size_t data_size,
+ uint8_t *data)
+{
+ struct gen_perf_config *perf_cfg = perf_ctx->perf;
+ const struct gen_perf_query_info *queryinfo = query->queryinfo;
+ int n_counters = queryinfo->n_counters;
+ int written = 0;
+
+ for (int i = 0; i < n_counters; i++) {
+ const struct gen_perf_query_counter *counter = &queryinfo->counters[i];
+ uint64_t *out_uint64;
+ float *out_float;
+ size_t counter_size = gen_perf_query_counter_get_size(counter);
+
+ if (counter_size) {
+ switch (counter->data_type) {
+ case GEN_PERF_COUNTER_DATA_TYPE_UINT64:
+ out_uint64 = (uint64_t *)(data + counter->offset);
+ *out_uint64 =
+ counter->oa_counter_read_uint64(perf_cfg, queryinfo,
+ query->oa.result.accumulator);
+ break;
+ case GEN_PERF_COUNTER_DATA_TYPE_FLOAT:
+ out_float = (float *)(data + counter->offset);
+ *out_float =
+ counter->oa_counter_read_float(perf_cfg, queryinfo,
+ query->oa.result.accumulator);
+ break;
+ default:
+ /* So far we aren't using uint32, double or bool32... */
+ unreachable("unexpected counter data type");
+ }
+ written = counter->offset + counter_size;
+ }
+ }
+
+ return written;
+}
+
+static int
+get_pipeline_stats_data(struct gen_perf_context *perf_ctx,
+ struct gen_perf_query_object *query,
+ size_t data_size,
+ uint8_t *data)
+
+{
+ struct gen_perf_config *perf_cfg = perf_ctx->perf;
+ const struct gen_perf_query_info *queryinfo = query->queryinfo;
+ int n_counters = queryinfo->n_counters;
+ uint8_t *p = data;
+
+ uint64_t *start = perf_cfg->vtbl.bo_map(perf_ctx->ctx, query->pipeline_stats.bo, MAP_READ);
+ uint64_t *end = start + (STATS_BO_END_OFFSET_BYTES / sizeof(uint64_t));
+
+ for (int i = 0; i < n_counters; i++) {
+ const struct gen_perf_query_counter *counter = &queryinfo->counters[i];
+ uint64_t value = end[i] - start[i];
+
+ if (counter->pipeline_stat.numerator !=
+ counter->pipeline_stat.denominator) {
+ value *= counter->pipeline_stat.numerator;
+ value /= counter->pipeline_stat.denominator;
+ }
+
+ *((uint64_t *)p) = value;
+ p += 8;
+ }
+
+ perf_cfg->vtbl.bo_unmap(query->pipeline_stats.bo);
+
+ return p - data;
+}
+
+void
+gen_perf_get_query_data(struct gen_perf_context *perf_ctx,
+ struct gen_perf_query_object *query,
+ int data_size,
+ unsigned *data,
+ unsigned *bytes_written)
+{
+ struct gen_perf_config *perf_cfg = perf_ctx->perf;
+ int written = 0;
+
+ switch (query->queryinfo->kind) {
+ case GEN_PERF_QUERY_TYPE_OA:
+ case GEN_PERF_QUERY_TYPE_RAW:
+ if (!query->oa.results_accumulated) {
+ read_gt_frequency(perf_ctx, query);
+ uint32_t *begin_report = query->oa.map;
+ uint32_t *end_report = query->oa.map + MI_RPC_BO_END_OFFSET_BYTES;
+ gen_perf_query_result_read_frequencies(&query->oa.result,
+ perf_ctx->devinfo,
+ begin_report,
+ end_report);
+ accumulate_oa_reports(perf_ctx, query);
+ assert(query->oa.results_accumulated);
+
+ perf_cfg->vtbl.bo_unmap(query->oa.bo);
+ query->oa.map = NULL;
+ }
+ if (query->queryinfo->kind == GEN_PERF_QUERY_TYPE_OA) {
+ written = get_oa_counter_data(perf_ctx, query, data_size, (uint8_t *)data);
+ } else {
+ const struct gen_device_info *devinfo = perf_ctx->devinfo;
+
+ written = gen_perf_query_result_write_mdapi((uint8_t *)data, data_size,
+ devinfo, &query->oa.result,
+ query->oa.gt_frequency[0],
+ query->oa.gt_frequency[1]);
+ }
+ break;
+
+ case GEN_PERF_QUERY_TYPE_PIPELINE:
+ written = get_pipeline_stats_data(perf_ctx, query, data_size, (uint8_t *)data);
+ break;
+
+ default:
+ unreachable("Unknown query type");
+ break;
+ }
+
+ if (bytes_written)
+ *bytes_written = written;
+}