*/
#include <limits.h>
-#include <dirent.h>
/* put before sys/types.h to silence glibc warnings */
#ifdef MAJOR_IN_MKDEV
#include <sys/ioctl.h>
#include <xf86drm.h>
-#include <i915_drm.h>
+#include "drm-uapi/i915_drm.h"
#include "main/hash.h"
#include "main/macros.h"
#include "util/ralloc.h"
#include "util/hash_table.h"
#include "util/list.h"
+#include "util/u_math.h"
#include "brw_context.h"
#include "brw_defines.h"
#include "brw_performance_query.h"
-#include "brw_oa_hsw.h"
#include "intel_batchbuffer.h"
+#include "perf/gen_perf.h"
+#include "perf/gen_perf_mdapi.h"
+
#define FILE_DEBUG_FLAG DEBUG_PERFMON
-/*
- * The largest OA format we can use on Haswell includes:
- * 1 timestamp, 45 A counters, 8 B counters and 8 C counters.
- */
-#define MAX_OA_REPORT_COUNTERS 62
+#define OAREPORT_REASON_MASK 0x3f
+#define OAREPORT_REASON_SHIFT 19
+#define OAREPORT_REASON_TIMER (1<<0)
+#define OAREPORT_REASON_TRIGGER1 (1<<1)
+#define OAREPORT_REASON_TRIGGER2 (1<<2)
+#define OAREPORT_REASON_CTX_SWITCH (1<<3)
+#define OAREPORT_REASON_GO_TRANSITION (1<<4)
#define I915_PERF_OA_SAMPLE_SIZE (8 + /* drm_i915_perf_record_header */ \
256) /* OA counter report */
int refcount;
int len;
uint8_t buf[I915_PERF_OA_SAMPLE_SIZE * 10];
-};
-
-/**
- * i965 representation of a performance query object.
- *
- * NB: We want to keep this structure relatively lean considering that
- * applications may expect to allocate enough objects to be able to
- * query around all draw calls in a frame.
- */
-struct brw_perf_query_object
-{
- struct gl_perf_query_object base;
-
- const struct brw_perf_query_info *query;
-
- /* See query->kind to know which state below is in use... */
- union {
- struct {
-
- /**
- * BO containing OA counter snapshots at query Begin/End time.
- */
- struct brw_bo *bo;
-
- /**
- * The MI_REPORT_PERF_COUNT command lets us specify a unique
- * ID that will be reflected in the resulting OA report
- * that's written by the GPU. This is the ID we're expecting
- * in the begin report and the the end report should be
- * @begin_report_id + 1.
- */
- int begin_report_id;
-
- /**
- * Reference the head of the brw->perfquery.sample_buffers
- * list at the time that the query started (so we only need
- * to look at nodes after this point when looking for samples
- * related to this query)
- *
- * (See struct brw_oa_sample_buf description for more details)
- */
- struct exec_node *samples_head;
-
- /**
- * Storage for the final accumulated OA counters.
- */
- uint64_t accumulator[MAX_OA_REPORT_COUNTERS];
-
- /**
- * false while in the unaccumulated_elements list, and set to
- * true when the final, end MI_RPC snapshot has been
- * accumulated.
- */
- bool results_accumulated;
-
- } oa;
-
- struct {
- /**
- * BO containing starting and ending snapshots for the
- * statistics counters.
- */
- struct brw_bo *bo;
- } pipeline_stats;
- };
+ uint32_t last_timestamp;
};
/** Downcasting convenience macro. */
return (struct brw_perf_query_object *) o;
}
-#define STATS_BO_SIZE 4096
-#define STATS_BO_END_OFFSET_BYTES (STATS_BO_SIZE / 2)
-#define MAX_STAT_COUNTERS (STATS_BO_END_OFFSET_BYTES / 8)
-
#define MI_RPC_BO_SIZE 4096
#define MI_RPC_BO_END_OFFSET_BYTES (MI_RPC_BO_SIZE / 2)
+#define MI_FREQ_START_OFFSET_BYTES (3072)
+#define MI_FREQ_END_OFFSET_BYTES (3076)
/******************************************************************************/
brw_is_perf_query_ready(struct gl_context *ctx,
struct gl_perf_query_object *o);
+static uint64_t
+brw_perf_query_get_metric_id(struct brw_context *brw,
+ const struct gen_perf_query_info *query)
+{
+ /* These queries are know not to ever change, their config ID has been
+ * loaded upon the first query creation. No need to look them up again.
+ */
+ if (query->kind == GEN_PERF_QUERY_TYPE_OA)
+ return query->oa_metrics_set_id;
+
+ assert(query->kind == GEN_PERF_QUERY_TYPE_RAW);
+
+ /* Raw queries can be reprogrammed up by an external application/library.
+ * When a raw query is used for the first time it's id is set to a value !=
+ * 0. When it stops being used the id returns to 0. No need to reload the
+ * ID when it's already loaded.
+ */
+ if (query->oa_metrics_set_id != 0) {
+ DBG("Raw query '%s' guid=%s using cached ID: %"PRIu64"\n",
+ query->name, query->guid, query->oa_metrics_set_id);
+ return query->oa_metrics_set_id;
+ }
+
+ struct gen_perf_query_info *raw_query = (struct gen_perf_query_info *)query;
+ if (!gen_perf_load_metric_id(brw->perfquery.perf, query->guid,
+ &raw_query->oa_metrics_set_id)) {
+ DBG("Unable to read query guid=%s ID, falling back to test config\n", query->guid);
+ raw_query->oa_metrics_set_id = 1ULL;
+ } else {
+ DBG("Raw query '%s'guid=%s loaded ID: %"PRIu64"\n",
+ query->name, query->guid, query->oa_metrics_set_id);
+ }
+ return query->oa_metrics_set_id;
+}
+
static void
dump_perf_query_callback(GLuint id, void *query_void, void *brw_void)
{
struct brw_perf_query_object *obj = query_void;
switch (obj->query->kind) {
- case OA_COUNTERS:
+ case GEN_PERF_QUERY_TYPE_OA:
+ case GEN_PERF_QUERY_TYPE_RAW:
DBG("%4d: %-6s %-8s BO: %-4s OA data: %-10s %-15s\n",
id,
o->Used ? "Dirty," : "New,",
brw_is_perf_query_ready(ctx, o) ? "ready," : "not ready,",
obj->oa.results_accumulated ? "accumulated" : "not accumulated");
break;
- case PIPELINE_STATS:
+ case GEN_PERF_QUERY_TYPE_PIPELINE:
DBG("%4d: %-6s %-8s BO: %-4s\n",
id,
o->Used ? "Dirty," : "New,",
o->Active ? "Active," : (o->Ready ? "Ready," : "Pending,"),
obj->pipeline_stats.bo ? "yes" : "no");
break;
+ default:
+ unreachable("Unknown query type");
+ break;
}
}
GLuint *n_active)
{
struct brw_context *brw = brw_context(ctx);
- const struct brw_perf_query_info *query =
- &brw->perfquery.queries[query_index];
+ const struct gen_perf_query_info *query =
+ &brw->perfquery.perf->queries[query_index];
*name = query->name;
*data_size = query->data_size;
*n_counters = query->n_counters;
switch (query->kind) {
- case OA_COUNTERS:
+ case GEN_PERF_QUERY_TYPE_OA:
+ case GEN_PERF_QUERY_TYPE_RAW:
*n_active = brw->perfquery.n_active_oa_queries;
break;
- case PIPELINE_STATS:
+ case GEN_PERF_QUERY_TYPE_PIPELINE:
*n_active = brw->perfquery.n_active_pipeline_stats_queries;
break;
+
+ default:
+ unreachable("Unknown query type");
+ break;
+ }
+}
+
+static GLuint
+gen_counter_type_enum_to_gl_type(enum gen_perf_counter_type type)
+{
+ switch (type) {
+ case GEN_PERF_COUNTER_TYPE_EVENT: return GL_PERFQUERY_COUNTER_EVENT_INTEL;
+ case GEN_PERF_COUNTER_TYPE_DURATION_NORM: return GL_PERFQUERY_COUNTER_DURATION_NORM_INTEL;
+ case GEN_PERF_COUNTER_TYPE_DURATION_RAW: return GL_PERFQUERY_COUNTER_DURATION_RAW_INTEL;
+ case GEN_PERF_COUNTER_TYPE_THROUGHPUT: return GL_PERFQUERY_COUNTER_THROUGHPUT_INTEL;
+ case GEN_PERF_COUNTER_TYPE_RAW: return GL_PERFQUERY_COUNTER_RAW_INTEL;
+ case GEN_PERF_COUNTER_TYPE_TIMESTAMP: return GL_PERFQUERY_COUNTER_TIMESTAMP_INTEL;
+ default:
+ unreachable("Unknown counter type");
+ }
+}
+
+static GLuint
+gen_counter_data_type_to_gl_type(enum gen_perf_counter_data_type type)
+{
+ switch (type) {
+ case GEN_PERF_COUNTER_DATA_TYPE_BOOL32: return GL_PERFQUERY_COUNTER_DATA_BOOL32_INTEL;
+ case GEN_PERF_COUNTER_DATA_TYPE_UINT32: return GL_PERFQUERY_COUNTER_DATA_UINT32_INTEL;
+ case GEN_PERF_COUNTER_DATA_TYPE_UINT64: return GL_PERFQUERY_COUNTER_DATA_UINT64_INTEL;
+ case GEN_PERF_COUNTER_DATA_TYPE_FLOAT: return GL_PERFQUERY_COUNTER_DATA_FLOAT_INTEL;
+ case GEN_PERF_COUNTER_DATA_TYPE_DOUBLE: return GL_PERFQUERY_COUNTER_DATA_DOUBLE_INTEL;
+ default:
+ unreachable("Unknown counter data type");
}
}
GLuint64 *raw_max)
{
struct brw_context *brw = brw_context(ctx);
- const struct brw_perf_query_info *query =
- &brw->perfquery.queries[query_index];
- const struct brw_perf_query_counter *counter =
+ const struct gen_perf_query_info *query =
+ &brw->perfquery.perf->queries[query_index];
+ const struct gen_perf_query_counter *counter =
&query->counters[counter_index];
*name = counter->name;
*desc = counter->desc;
*offset = counter->offset;
- *data_size = counter->size;
- *type_enum = counter->type;
- *data_type_enum = counter->data_type;
+ *data_size = gen_perf_query_counter_get_size(counter);
+ *type_enum = gen_counter_type_enum_to_gl_type(counter->type);
+ *data_type_enum = gen_counter_data_type_to_gl_type(counter->data_type);
*raw_max = counter->raw_max;
}
struct brw_perf_query_object *obj,
uint32_t offset_in_bytes)
{
- const struct brw_perf_query_info *query = obj->query;
+ const struct gen_perf_query_info *query = obj->query;
const int n_counters = query->n_counters;
for (int i = 0; i < n_counters; i++) {
- const struct brw_perf_query_counter *counter = &query->counters[i];
+ const struct gen_perf_query_counter *counter = &query->counters[i];
- assert(counter->data_type == GL_PERFQUERY_COUNTER_DATA_UINT64_INTEL);
+ assert(counter->data_type == GEN_PERF_COUNTER_DATA_TYPE_UINT64);
brw_store_register_mem64(brw, obj->pipeline_stats.bo,
counter->pipeline_stat.reg,
}
}
-/**
- * Emit an MI_REPORT_PERF_COUNT command packet.
- *
- * This asks the GPU to write a report of the current OA counter
- * values into @bo at the given offset and containing the given
- * @report_id which we can cross-reference when parsing the report.
- */
-static void
-emit_mi_report_perf_count(struct brw_context *brw,
- struct brw_bo *bo,
- uint32_t offset_in_bytes,
- uint32_t report_id)
-{
- assert(offset_in_bytes % 64 == 0);
-
- BEGIN_BATCH(3);
- OUT_BATCH(GEN6_MI_REPORT_PERF_COUNT);
- OUT_RELOC(bo, I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
- offset_in_bytes);
- OUT_BATCH(report_id);
- ADVANCE_BATCH();
-}
-
/**
* Add a query to the global list of "unaccumulated queries."
*
reap_old_sample_buffers(brw);
}
-static uint64_t
-timebase_scale(struct brw_context *brw, uint32_t u32_time_delta)
-{
- uint64_t tmp = ((uint64_t)u32_time_delta) * 1000000000ull;
-
- return tmp ? tmp / brw->perfquery.sys_vars.timestamp_frequency : 0;
-}
-
-static void
-accumulate_uint32(const uint32_t *report0,
- const uint32_t *report1,
- uint64_t *accumulator)
-{
- *accumulator += (uint32_t)(*report1 - *report0);
-}
-
-/**
- * Given pointers to starting and ending OA snapshots, add the deltas for each
- * counter to the results.
- */
-static void
-add_deltas(struct brw_context *brw,
- struct brw_perf_query_object *obj,
- const uint32_t *start,
- const uint32_t *end)
-{
- const struct brw_perf_query_info *query = obj->query;
- uint64_t *accumulator = obj->oa.accumulator;
- int i;
-
- switch (query->oa_format) {
- case I915_OA_FORMAT_A45_B8_C8:
- accumulate_uint32(start + 1, end + 1, accumulator); /* timestamp */
-
- for (i = 0; i < 61; i++)
- accumulate_uint32(start + 3 + i, end + 3 + i, accumulator + 1 + i);
-
- break;
- default:
- unreachable("Can't accumulate OA counters in unknown format");
- }
-}
-
static bool
inc_n_oa_users(struct brw_context *brw)
{
}
}
-static bool
-read_oa_samples(struct brw_context *brw)
+enum OaReadStatus {
+ OA_READ_STATUS_ERROR,
+ OA_READ_STATUS_UNFINISHED,
+ OA_READ_STATUS_FINISHED,
+};
+
+static enum OaReadStatus
+read_oa_samples_until(struct brw_context *brw,
+ uint32_t start_timestamp,
+ uint32_t end_timestamp)
{
+ struct exec_node *tail_node =
+ exec_list_get_tail(&brw->perfquery.sample_buffers);
+ struct brw_oa_sample_buf *tail_buf =
+ exec_node_data(struct brw_oa_sample_buf, tail_node, link);
+ uint32_t last_timestamp = tail_buf->last_timestamp;
+
while (1) {
struct brw_oa_sample_buf *buf = get_free_sample_buf(brw);
+ uint32_t offset;
int len;
while ((len = read(brw->perfquery.oa_stream_fd, buf->buf,
if (len < 0) {
if (errno == EAGAIN)
- return true;
+ return ((last_timestamp - start_timestamp) >=
+ (end_timestamp - start_timestamp)) ?
+ OA_READ_STATUS_FINISHED :
+ OA_READ_STATUS_UNFINISHED;
else {
DBG("Error reading i915 perf samples: %m\n");
- return false;
}
- } else {
+ } else
DBG("Spurious EOF reading i915 perf samples\n");
- return false;
- }
+
+ return OA_READ_STATUS_ERROR;
}
buf->len = len;
exec_list_push_tail(&brw->perfquery.sample_buffers, &buf->link);
+
+ /* Go through the reports and update the last timestamp. */
+ offset = 0;
+ while (offset < buf->len) {
+ const struct drm_i915_perf_record_header *header =
+ (const struct drm_i915_perf_record_header *) &buf->buf[offset];
+ uint32_t *report = (uint32_t *) (header + 1);
+
+ if (header->type == DRM_I915_PERF_RECORD_SAMPLE)
+ last_timestamp = report[1];
+
+ offset += header->size;
+ }
+
+ buf->last_timestamp = last_timestamp;
}
unreachable("not reached");
+ return OA_READ_STATUS_ERROR;
+}
+
+/**
+ * Try to read all the reports until either the delimiting timestamp
+ * or an error arises.
+ */
+static bool
+read_oa_samples_for_query(struct brw_context *brw,
+ struct brw_perf_query_object *obj)
+{
+ uint32_t *start;
+ uint32_t *last;
+ uint32_t *end;
+
+ /* We need the MI_REPORT_PERF_COUNT to land before we can start
+ * accumulate. */
+ assert(!brw_batch_references(&brw->batch, obj->oa.bo) &&
+ !brw_bo_busy(obj->oa.bo));
+
+ /* Map the BO once here and let accumulate_oa_reports() unmap
+ * it. */
+ if (obj->oa.map == NULL)
+ obj->oa.map = brw_bo_map(brw, obj->oa.bo, MAP_READ);
+
+ start = last = obj->oa.map;
+ end = obj->oa.map + MI_RPC_BO_END_OFFSET_BYTES;
+
+ if (start[0] != obj->oa.begin_report_id) {
+ DBG("Spurious start report id=%"PRIu32"\n", start[0]);
+ return true;
+ }
+ if (end[0] != (obj->oa.begin_report_id + 1)) {
+ DBG("Spurious end report id=%"PRIu32"\n", end[0]);
+ return true;
+ }
+
+ /* Read the reports until the end timestamp. */
+ switch (read_oa_samples_until(brw, start[1], end[1])) {
+ case OA_READ_STATUS_ERROR:
+ /* Fallthrough and let accumulate_oa_reports() deal with the
+ * error. */
+ case OA_READ_STATUS_FINISHED:
+ return true;
+ case OA_READ_STATUS_UNFINISHED:
+ return false;
+ }
+
+ unreachable("invalid read status");
return false;
}
/**
- * Accumulate raw OA counter values based on deltas between pairs
- * of OA reports.
+ * 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
*
* 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.
+ * 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 brw_context *brw,
struct brw_perf_query_object *obj)
{
+ const struct gen_device_info *devinfo = &brw->screen->devinfo;
struct gl_perf_query_object *o = &obj->base;
- uint32_t *query_buffer;
uint32_t *start;
uint32_t *last;
uint32_t *end;
struct exec_node *first_samples_node;
+ bool in_ctx = true;
+ int out_duration = 0;
assert(o->Ready);
+ assert(obj->oa.map != NULL);
- /* Collect the latest periodic OA reports from i915 perf */
- if (!read_oa_samples(brw))
- goto error;
-
- brw_bo_map(obj->oa.bo, false);
- query_buffer = obj->oa.bo->virtual;
-
- start = last = query_buffer;
- end = query_buffer + (MI_RPC_BO_END_OFFSET_BYTES / sizeof(uint32_t));
+ start = last = obj->oa.map;
+ end = obj->oa.map + MI_RPC_BO_END_OFFSET_BYTES;
if (start[0] != obj->oa.begin_report_id) {
DBG("Spurious start report id=%"PRIu32"\n", start[0]);
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 (timebase_scale(brw, report[1] - start[1]) > 5000000000)
+ 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 (timebase_scale(brw, report[1] - end[1]) <= 5000000000)
+ if (gen_device_info_timebase_scale(devinfo,
+ report[1] - end[1]) <= 5000000000) {
goto end;
+ }
- add_deltas(brw, obj, last, report);
+ /* 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] != obj->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] == obj->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] == obj->oa.result.hw_id);
+ DBG("i915 perf: Continuation IN\n");
+ } else {
+ assert(report[2] != obj->oa.result.hw_id);
+ DBG("i915 perf: Continuation OUT\n");
+ add = false;
+ out_duration++;
+ }
+ }
+
+ if (add) {
+ gen_perf_query_result_accumulate(&obj->oa.result, obj->query,
+ last, report);
+ }
last = report;
end:
- add_deltas(brw, obj, last, end);
+ gen_perf_query_result_accumulate(&obj->oa.result, obj->query,
+ last, end);
DBG("Marking %d accumulated - results gathered\n", o->Id);
- brw_bo_unmap(obj->oa.bo);
obj->oa.results_accumulated = true;
drop_from_unaccumulated_query_list(brw, obj);
dec_n_oa_users(brw);
error:
- brw_bo_unmap(obj->oa.bo);
discard_all_queries(brw);
}
I915_PERF_FLAG_FD_NONBLOCK |
I915_PERF_FLAG_DISABLED,
.num_properties = ARRAY_SIZE(properties) / 2,
- .properties_ptr = (uint64_t)properties
+ .properties_ptr = (uintptr_t) properties,
};
int fd = drmIoctl(drm_fd, DRM_IOCTL_I915_PERF_OPEN, ¶m);
if (fd == -1) {
}
static void
-close_perf(struct brw_context *brw)
+close_perf(struct brw_context *brw,
+ const struct gen_perf_query_info *query)
{
if (brw->perfquery.oa_stream_fd != -1) {
close(brw->perfquery.oa_stream_fd);
brw->perfquery.oa_stream_fd = -1;
}
+ if (query->kind == GEN_PERF_QUERY_TYPE_RAW) {
+ struct gen_perf_query_info *raw_query =
+ (struct gen_perf_query_info *) query;
+ raw_query->oa_metrics_set_id = 0;
+ }
+}
+
+static void
+capture_frequency_stat_register(struct brw_context *brw,
+ struct brw_bo *bo,
+ uint32_t bo_offset)
+{
+ const struct gen_device_info *devinfo = &brw->screen->devinfo;
+
+ if (devinfo->gen >= 7 && devinfo->gen <= 8 &&
+ !devinfo->is_baytrail && !devinfo->is_cherryview) {
+ brw_store_register_mem32(brw, bo, GEN7_RPSTAT1, bo_offset);
+ } else if (devinfo->gen >= 9) {
+ brw_store_register_mem32(brw, bo, GEN9_RPSTAT0, bo_offset);
+ }
}
/**
{
struct brw_context *brw = brw_context(ctx);
struct brw_perf_query_object *obj = brw_perf_query(o);
- const struct brw_perf_query_info *query = obj->query;
+ const struct gen_perf_query_info *query = obj->query;
/* We can assume the frontend hides mistaken attempts to Begin a
* query object multiple times before its End. Similarly if an
brw_emit_mi_flush(brw);
switch (query->kind) {
- case OA_COUNTERS:
+ case GEN_PERF_QUERY_TYPE_OA:
+ case GEN_PERF_QUERY_TYPE_RAW: {
/* Opening an i915 perf stream implies exclusive access to the OA unit
* which will generate counter reports for a specific counter set with a
* require a different counter set or format unless we get an opportunity
* to close the stream and open a new one...
*/
+ uint64_t metric_id = brw_perf_query_get_metric_id(brw, query);
+
if (brw->perfquery.oa_stream_fd != -1 &&
- brw->perfquery.current_oa_metrics_set_id !=
- query->oa_metrics_set_id) {
+ brw->perfquery.current_oa_metrics_set_id != metric_id) {
- if (brw->perfquery.n_oa_users != 0)
+ if (brw->perfquery.n_oa_users != 0) {
+ DBG("WARNING: Begin(%d) failed already using perf config=%i/%"PRIu64"\n",
+ o->Id, brw->perfquery.current_oa_metrics_set_id, metric_id);
return false;
- else
- close_perf(brw);
+ } else
+ close_perf(brw, query);
}
/* If the OA counters aren't already on, enable them. */
if (brw->perfquery.oa_stream_fd == -1) {
__DRIscreen *screen = brw->screen->driScrnPriv;
- int period_exponent;
+ const struct gen_device_info *devinfo = &brw->screen->devinfo;
- /* The timestamp for HSW+ increments every 80ns
+ /* The period_exponent gives a sampling period as follows:
+ * sample_period = timestamp_period * 2^(period_exponent + 1)
*
- * The period_exponent gives a sampling period as follows:
- * sample_period = 80ns * 2^(period_exponent + 1)
+ * The timestamps increments every 80ns (HSW), ~52ns (GEN9LP) or
+ * ~83ns (GEN8/9).
*
- * The overflow period for Haswell can be calculated as:
+ * The counter overflow period is derived from the EuActive counter
+ * which reads a counter that increments by the number of clock
+ * cycles multiplied by the number of EUs. It can be calculated as:
+ *
+ * 2^(number of bits in A counter) / (n_eus * max_gen_freq * 2)
*
- * 2^32 / (n_eus * max_gen_freq * 2)
* (E.g. 40 EUs @ 1GHz = ~53ms)
*
- * We currently sample every 42 milliseconds...
+ * We select a sampling period inferior to that overflow period to
+ * ensure we cannot see more than 1 counter overflow, otherwise we
+ * could loose information.
*/
- period_exponent = 18;
+
+ int a_counter_in_bits = 32;
+ if (devinfo->gen >= 8)
+ a_counter_in_bits = 40;
+
+ uint64_t overflow_period = pow(2, a_counter_in_bits) /
+ (brw->perfquery.perf->sys_vars.n_eus *
+ /* drop 1GHz freq to have units in nanoseconds */
+ 2);
+
+ DBG("A counter overflow period: %"PRIu64"ns, %"PRIu64"ms (n_eus=%"PRIu64")\n",
+ overflow_period, overflow_period / 1000000ul, brw->perfquery.perf->sys_vars.n_eus);
+
+ int period_exponent = 0;
+ uint64_t prev_sample_period, next_sample_period;
+ for (int e = 0; e < 30; e++) {
+ prev_sample_period = 1000000000ull * pow(2, e + 1) / devinfo->timestamp_frequency;
+ next_sample_period = 1000000000ull * pow(2, e + 2) / devinfo->timestamp_frequency;
+
+ /* Take the previous sampling period, lower than the overflow
+ * period.
+ */
+ if (prev_sample_period < overflow_period &&
+ next_sample_period > overflow_period)
+ period_exponent = e + 1;
+ }
+
+ if (period_exponent == 0) {
+ DBG("WARNING: enable to find a sampling exponent\n");
+ return false;
+ }
+
+ DBG("OA sampling exponent: %i ~= %"PRIu64"ms\n", period_exponent,
+ prev_sample_period / 1000000ul);
if (!open_i915_perf_oa_stream(brw,
- query->oa_metrics_set_id,
+ metric_id,
query->oa_format,
period_exponent,
screen->fd, /* drm fd */
brw->hw_ctx))
return false;
} else {
- assert(brw->perfquery.current_oa_metrics_set_id ==
- query->oa_metrics_set_id &&
- brw->perfquery.current_oa_format ==
- query->oa_format);
+ assert(brw->perfquery.current_oa_metrics_set_id == metric_id &&
+ brw->perfquery.current_oa_format == query->oa_format);
}
if (!inc_n_oa_users(brw)) {
}
obj->oa.bo =
- brw_bo_alloc(brw->bufmgr, "perf. query OA MI_RPC bo",
- MI_RPC_BO_SIZE, 64);
+ brw_bo_alloc(brw->bufmgr, "perf. query OA MI_RPC bo", MI_RPC_BO_SIZE,
+ BRW_MEMZONE_OTHER);
#ifdef DEBUG
/* Pre-filling the BO helps debug whether writes landed. */
- brw_bo_map(obj->oa.bo, true);
- memset((char *) obj->oa.bo->virtual, 0x80, MI_RPC_BO_SIZE);
+ void *map = brw_bo_map(brw, obj->oa.bo, MAP_WRITE);
+ memset(map, 0x80, MI_RPC_BO_SIZE);
brw_bo_unmap(obj->oa.bo);
#endif
obj->oa.begin_report_id = brw->perfquery.next_query_start_report_id;
brw->perfquery.next_query_start_report_id += 2;
+ /* We flush the batchbuffer here to minimize the chances that MI_RPC
+ * delimiting commands end up in different batchbuffers. If that's the
+ * case, the measurement will include the time it takes for the kernel
+ * scheduler to load a new request into the hardware. This is manifested in
+ * tools like frameretrace by spikes in the "GPU Core Clocks" counter.
+ */
+ intel_batchbuffer_flush(brw);
+
/* Take a starting OA counter snapshot. */
- emit_mi_report_perf_count(brw, obj->oa.bo, 0,
- obj->oa.begin_report_id);
+ brw->vtbl.emit_mi_report_perf_count(brw, obj->oa.bo, 0,
+ obj->oa.begin_report_id);
+ capture_frequency_stat_register(brw, obj->oa.bo, MI_FREQ_START_OFFSET_BYTES);
+
++brw->perfquery.n_active_oa_queries;
/* No already-buffered samples can possibly be associated with this query
*/
buf->refcount++;
- memset(obj->oa.accumulator, 0, sizeof(obj->oa.accumulator));
+ gen_perf_query_result_clear(&obj->oa.result);
obj->oa.results_accumulated = false;
add_to_unaccumulated_query_list(brw, obj);
break;
+ }
- case PIPELINE_STATS:
+ case GEN_PERF_QUERY_TYPE_PIPELINE:
if (obj->pipeline_stats.bo) {
brw_bo_unreference(obj->pipeline_stats.bo);
obj->pipeline_stats.bo = NULL;
obj->pipeline_stats.bo =
brw_bo_alloc(brw->bufmgr, "perf. query pipeline stats bo",
- STATS_BO_SIZE, 64);
+ STATS_BO_SIZE, BRW_MEMZONE_OTHER);
/* Take starting snapshots. */
snapshot_statistics_registers(brw, obj, 0);
++brw->perfquery.n_active_pipeline_stats_queries;
break;
+
+ default:
+ unreachable("Unknown query type");
+ break;
}
if (INTEL_DEBUG & DEBUG_PERFMON)
brw_emit_mi_flush(brw);
switch (obj->query->kind) {
- case OA_COUNTERS:
+ case GEN_PERF_QUERY_TYPE_OA:
+ case GEN_PERF_QUERY_TYPE_RAW:
/* NB: It's possible that the query will have already been marked
* as 'accumulated' if an error was seen while reading samples
*/
if (!obj->oa.results_accumulated) {
/* Take an ending OA counter snapshot. */
- emit_mi_report_perf_count(brw, obj->oa.bo,
- MI_RPC_BO_END_OFFSET_BYTES,
- obj->oa.begin_report_id + 1);
+ capture_frequency_stat_register(brw, obj->oa.bo, MI_FREQ_END_OFFSET_BYTES);
+ brw->vtbl.emit_mi_report_perf_count(brw, obj->oa.bo,
+ MI_RPC_BO_END_OFFSET_BYTES,
+ obj->oa.begin_report_id + 1);
}
--brw->perfquery.n_active_oa_queries;
*/
break;
- case PIPELINE_STATS:
+ case GEN_PERF_QUERY_TYPE_PIPELINE:
snapshot_statistics_registers(brw, obj,
STATS_BO_END_OFFSET_BYTES);
--brw->perfquery.n_active_pipeline_stats_queries;
break;
+
+ default:
+ unreachable("Unknown query type");
+ break;
}
}
assert(!o->Ready);
switch (obj->query->kind) {
- case OA_COUNTERS:
+ case GEN_PERF_QUERY_TYPE_OA:
+ case GEN_PERF_QUERY_TYPE_RAW:
bo = obj->oa.bo;
break;
- case PIPELINE_STATS:
+ case GEN_PERF_QUERY_TYPE_PIPELINE:
bo = obj->pipeline_stats.bo;
break;
+
+ default:
+ unreachable("Unknown query type");
+ break;
}
if (bo == NULL)
if (brw_batch_references(&brw->batch, bo))
intel_batchbuffer_flush(brw);
- if (unlikely(brw->perf_debug)) {
- if (brw_bo_busy(bo))
- perf_debug("Stalling GPU waiting for a performance query object.\n");
- }
-
brw_bo_wait_rendering(bo);
+
+ /* Due to a race condition between the OA unit signaling report
+ * availability and the report actually being written into memory,
+ * we need to wait for all the reports to come in before we can
+ * read them.
+ */
+ if (obj->query->kind == GEN_PERF_QUERY_TYPE_OA ||
+ obj->query->kind == GEN_PERF_QUERY_TYPE_RAW) {
+ while (!read_oa_samples_for_query(brw, obj))
+ ;
+ }
}
static bool
return true;
switch (obj->query->kind) {
- case OA_COUNTERS:
+ case GEN_PERF_QUERY_TYPE_OA:
+ case GEN_PERF_QUERY_TYPE_RAW:
return (obj->oa.results_accumulated ||
(obj->oa.bo &&
!brw_batch_references(&brw->batch, obj->oa.bo) &&
- !brw_bo_busy(obj->oa.bo)));
-
- case PIPELINE_STATS:
+ !brw_bo_busy(obj->oa.bo) &&
+ read_oa_samples_for_query(brw, obj)));
+ case GEN_PERF_QUERY_TYPE_PIPELINE:
return (obj->pipeline_stats.bo &&
!brw_batch_references(&brw->batch, obj->pipeline_stats.bo) &&
!brw_bo_busy(obj->pipeline_stats.bo));
+
+ default:
+ unreachable("Unknown query type");
+ break;
}
- unreachable("missing ready check for unknown query kind");
return false;
}
+static void
+read_slice_unslice_frequencies(struct brw_context *brw,
+ struct brw_perf_query_object *obj)
+{
+ const struct gen_device_info *devinfo = &brw->screen->devinfo;
+ uint32_t *begin_report = obj->oa.map, *end_report = obj->oa.map + MI_RPC_BO_END_OFFSET_BYTES;
+
+ gen_perf_query_result_read_frequencies(&obj->oa.result,
+ devinfo, begin_report, end_report);
+}
+
+static void
+read_gt_frequency(struct brw_context *brw,
+ struct brw_perf_query_object *obj)
+{
+ const struct gen_device_info *devinfo = &brw->screen->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 brw_context *brw,
struct brw_perf_query_object *obj,
size_t data_size,
uint8_t *data)
{
- const struct brw_perf_query_info *query = obj->query;
+ struct gen_perf *perf = brw->perfquery.perf;
+ const struct gen_perf_query_info *query = obj->query;
int n_counters = query->n_counters;
int written = 0;
- if (!obj->oa.results_accumulated) {
- accumulate_oa_reports(brw, obj);
- assert(obj->oa.results_accumulated);
- }
-
for (int i = 0; i < n_counters; i++) {
- const struct brw_perf_query_counter *counter = &query->counters[i];
+ const struct gen_perf_query_counter *counter = &query->counters[i];
uint64_t *out_uint64;
float *out_float;
+ size_t counter_size = gen_perf_query_counter_get_size(counter);
- if (counter->size) {
+ if (counter_size) {
switch (counter->data_type) {
- case GL_PERFQUERY_COUNTER_DATA_UINT64_INTEL:
+ case GEN_PERF_COUNTER_DATA_TYPE_UINT64:
out_uint64 = (uint64_t *)(data + counter->offset);
- *out_uint64 = counter->oa_counter_read_uint64(brw, query,
- obj->oa.accumulator);
+ *out_uint64 =
+ counter->oa_counter_read_uint64(perf, query,
+ obj->oa.result.accumulator);
break;
- case GL_PERFQUERY_COUNTER_DATA_FLOAT_INTEL:
+ case GEN_PERF_COUNTER_DATA_TYPE_FLOAT:
out_float = (float *)(data + counter->offset);
- *out_float = counter->oa_counter_read_float(brw, query,
- obj->oa.accumulator);
+ *out_float =
+ counter->oa_counter_read_float(perf, query,
+ obj->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;
+ written = counter->offset + counter_size;
}
}
uint8_t *data)
{
- const struct brw_perf_query_info *query = obj->query;
+ const struct gen_perf_query_info *query = obj->query;
int n_counters = obj->query->n_counters;
uint8_t *p = data;
- brw_bo_map(obj->pipeline_stats.bo, false);
- uint64_t *start = obj->pipeline_stats.bo->virtual;
+ uint64_t *start = brw_bo_map(brw, obj->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 brw_perf_query_counter *counter = &query->counters[i];
+ const struct gen_perf_query_counter *counter = &query->counters[i];
uint64_t value = end[i] - start[i];
if (counter->pipeline_stat.numerator !=
assert(o->Ready);
switch (obj->query->kind) {
- case OA_COUNTERS:
- written = get_oa_counter_data(brw, obj, data_size, (uint8_t *)data);
+ case GEN_PERF_QUERY_TYPE_OA:
+ case GEN_PERF_QUERY_TYPE_RAW:
+ if (!obj->oa.results_accumulated) {
+ read_gt_frequency(brw, obj);
+ read_slice_unslice_frequencies(brw, obj);
+ accumulate_oa_reports(brw, obj);
+ assert(obj->oa.results_accumulated);
+
+ brw_bo_unmap(obj->oa.bo);
+ obj->oa.map = NULL;
+ }
+ if (obj->query->kind == GEN_PERF_QUERY_TYPE_OA) {
+ written = get_oa_counter_data(brw, obj, data_size, (uint8_t *)data);
+ } else {
+ const struct gen_device_info *devinfo = &brw->screen->devinfo;
+
+ written = gen_perf_query_result_write_mdapi((uint8_t *)data, data_size,
+ devinfo, &obj->oa.result,
+ obj->oa.gt_frequency[0],
+ obj->oa.gt_frequency[1]);
+ }
break;
- case PIPELINE_STATS:
+ case GEN_PERF_QUERY_TYPE_PIPELINE:
written = get_pipeline_stats_data(brw, obj, data_size, (uint8_t *)data);
break;
+
+ default:
+ unreachable("Unknown query type");
+ break;
}
if (bytes_written)
brw_new_perf_query_object(struct gl_context *ctx, unsigned query_index)
{
struct brw_context *brw = brw_context(ctx);
- const struct brw_perf_query_info *query =
- &brw->perfquery.queries[query_index];
+ const struct gen_perf_query_info *query =
+ &brw->perfquery.perf->queries[query_index];
struct brw_perf_query_object *obj =
calloc(1, sizeof(struct brw_perf_query_object));
DBG("Delete(%d)\n", o->Id);
switch (obj->query->kind) {
- case OA_COUNTERS:
+ case GEN_PERF_QUERY_TYPE_OA:
+ case GEN_PERF_QUERY_TYPE_RAW:
if (obj->oa.bo) {
if (!obj->oa.results_accumulated) {
drop_from_unaccumulated_query_list(brw, obj);
obj->oa.results_accumulated = false;
break;
- case PIPELINE_STATS:
+ case GEN_PERF_QUERY_TYPE_PIPELINE:
if (obj->pipeline_stats.bo) {
brw_bo_unreference(obj->pipeline_stats.bo);
obj->pipeline_stats.bo = NULL;
}
break;
- }
- free(obj);
+ 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
*/
if (--brw->perfquery.n_query_instances == 0) {
free_sample_bufs(brw);
- close_perf(brw);
+ close_perf(brw, obj->query);
}
-}
-
-/******************************************************************************/
-
-static struct brw_perf_query_info *
-append_query_info(struct brw_context *brw)
-{
- brw->perfquery.queries =
- reralloc(brw, brw->perfquery.queries,
- struct brw_perf_query_info, ++brw->perfquery.n_queries);
- return &brw->perfquery.queries[brw->perfquery.n_queries - 1];
-}
-
-static void
-add_stat_reg(struct brw_perf_query_info *query,
- uint32_t reg,
- uint32_t numerator,
- uint32_t denominator,
- const char *name,
- const char *description)
-{
- struct brw_perf_query_counter *counter;
-
- assert(query->n_counters < MAX_STAT_COUNTERS);
-
- counter = &query->counters[query->n_counters];
- counter->name = name;
- counter->desc = description;
- counter->type = GL_PERFQUERY_COUNTER_RAW_INTEL;
- counter->data_type = GL_PERFQUERY_COUNTER_DATA_UINT64_INTEL;
- counter->size = sizeof(uint64_t);
- 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++;
+ free(obj);
}
-static void
-add_basic_stat_reg(struct brw_perf_query_info *query,
- uint32_t reg, const char *name)
-{
- add_stat_reg(query, reg, 1, 1, name, name);
-}
+/******************************************************************************/
static void
init_pipeline_statistic_query_registers(struct brw_context *brw)
{
- struct brw_perf_query_info *query = append_query_info(brw);
+ const struct gen_device_info *devinfo = &brw->screen->devinfo;
+ struct gen_perf *perf = brw->perfquery.perf;
+ struct gen_perf_query_info *query =
+ gen_perf_query_append_query_info(perf, MAX_STAT_COUNTERS);
- query->kind = PIPELINE_STATS;
+ query->kind = GEN_PERF_QUERY_TYPE_PIPELINE;
query->name = "Pipeline Statistics Registers";
- query->n_counters = 0;
- query->counters =
- rzalloc_array(brw, struct brw_perf_query_counter, MAX_STAT_COUNTERS);
-
- 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 (brw->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 (brw->is_haswell || brw->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 (brw->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
-read_file_uint64(const char *file, uint64_t *val)
-{
- char buf[32];
- int fd, n;
-
- fd = open(file, 0);
- if (fd < 0)
- return false;
- n = read(fd, buf, sizeof (buf) - 1);
- close(fd);
- if (n < 0)
- return false;
-
- buf[n] = '\0';
- *val = strtoull(buf, NULL, 0);
-
- return true;
-}
-
-static void
-enumerate_sysfs_metrics(struct brw_context *brw, const char *sysfs_dev_dir)
-{
- char buf[256];
- DIR *metricsdir = NULL;
- struct dirent *metric_entry;
- int len;
-
- len = snprintf(buf, sizeof(buf), "%s/metrics", sysfs_dev_dir);
- if (len < 0 || len >= sizeof(buf)) {
- DBG("Failed to concatenate path to sysfs metrics/ directory\n");
- return;
+ 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");
+
+ if (devinfo->gen == 6) {
+ gen_perf_query_info_add_stat_reg(query, GEN6_SO_PRIM_STORAGE_NEEDED, 1, 1,
+ "SO_PRIM_STORAGE_NEEDED",
+ "N geometry shader stream-out primitives (total)");
+ gen_perf_query_info_add_stat_reg(query, GEN6_SO_NUM_PRIMS_WRITTEN, 1, 1,
+ "SO_NUM_PRIMS_WRITTEN",
+ "N geometry shader stream-out primitives (written)");
+ } else {
+ gen_perf_query_info_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)");
+ gen_perf_query_info_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)");
+ gen_perf_query_info_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)");
+ gen_perf_query_info_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)");
+ gen_perf_query_info_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)");
+ gen_perf_query_info_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)");
+ gen_perf_query_info_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)");
+ gen_perf_query_info_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)");
}
- metricsdir = opendir(buf);
- if (!metricsdir) {
- DBG("Failed to open %s: %m\n", buf);
- return;
+ 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");
+
+ 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");
}
- while ((metric_entry = readdir(metricsdir))) {
- struct hash_entry *entry;
-
- if ((metric_entry->d_type != DT_DIR &&
- metric_entry->d_type != DT_LNK) ||
- metric_entry->d_name[0] == '.')
- continue;
-
- DBG("metric set: %s\n", metric_entry->d_name);
- entry = _mesa_hash_table_search(brw->perfquery.oa_metrics_table,
- metric_entry->d_name);
- if (entry) {
- struct brw_perf_query_info *query;
- uint64_t id;
-
- len = snprintf(buf, sizeof(buf), "%s/metrics/%s/id",
- sysfs_dev_dir, metric_entry->d_name);
- if (len < 0 || len >= sizeof(buf)) {
- DBG("Failed to concatenate path to sysfs metric id file\n");
- continue;
- }
-
- if (!read_file_uint64(buf, &id)) {
- DBG("Failed to read metric set id from %s: %m", buf);
- continue;
- }
-
- query = append_query_info(brw);
- *query = *(struct brw_perf_query_info *)entry->data;
- query->oa_metrics_set_id = id;
+ gen_perf_query_info_add_basic_stat_reg(query, PS_DEPTH_COUNT,
+ "N z-pass fragments");
- DBG("metric set known by mesa: id = %" PRIu64"\n",
- query->oa_metrics_set_id);
- } else
- DBG("metric set not known by mesa (skipping)\n");
+ if (devinfo->gen >= 7) {
+ gen_perf_query_info_add_basic_stat_reg(query, CS_INVOCATION_COUNT,
+ "N compute shader invocations");
}
- closedir(metricsdir);
+ query->data_size = sizeof(uint64_t) * query->n_counters;
}
static bool
-read_sysfs_drm_device_file_uint64(struct brw_context *brw,
- const char *sysfs_dev_dir,
- const char *file,
- uint64_t *value)
+query_topology(struct brw_context *brw)
{
- char buf[512];
- int len;
+ __DRIscreen *screen = brw->screen->driScrnPriv;
+ struct drm_i915_query_item item = {
+ .query_id = DRM_I915_QUERY_TOPOLOGY_INFO,
+ };
+ struct drm_i915_query query = {
+ .num_items = 1,
+ .items_ptr = (uintptr_t) &item,
+ };
- len = snprintf(buf, sizeof(buf), "%s/%s", sysfs_dev_dir, file);
- if (len < 0 || len >= sizeof(buf)) {
- DBG("Failed to concatenate sys filename to read u64 from\n");
+ if (drmIoctl(screen->fd, DRM_IOCTL_I915_QUERY, &query))
return false;
- }
- return read_file_uint64(buf, value);
-}
+ struct drm_i915_query_topology_info *topo_info =
+ (struct drm_i915_query_topology_info *) calloc(1, item.length);
+ item.data_ptr = (uintptr_t) topo_info;
-static bool
-init_oa_sys_vars(struct brw_context *brw, const char *sysfs_dev_dir)
-{
- uint64_t min_freq_mhz = 0, max_freq_mhz = 0;
-
- if (!read_sysfs_drm_device_file_uint64(brw, sysfs_dev_dir,
- "gt_min_freq_mhz",
- &min_freq_mhz))
+ if (drmIoctl(screen->fd, DRM_IOCTL_I915_QUERY, &query) ||
+ item.length <= 0)
return false;
- if (!read_sysfs_drm_device_file_uint64(brw, sysfs_dev_dir,
- "gt_max_freq_mhz",
- &max_freq_mhz))
- return false;
+ gen_device_info_update_from_topology(&brw->screen->devinfo,
+ topo_info);
- brw->perfquery.sys_vars.gt_min_freq = min_freq_mhz * 1000000;
- brw->perfquery.sys_vars.gt_max_freq = max_freq_mhz * 1000000;
-
- if (brw->is_haswell) {
- const struct gen_device_info *info = &brw->screen->devinfo;
-
- brw->perfquery.sys_vars.timestamp_frequency = 12500000;
-
- if (info->gt == 1) {
- brw->perfquery.sys_vars.n_eus = 10;
- brw->perfquery.sys_vars.n_eu_slices = 1;
- brw->perfquery.sys_vars.subslice_mask = 0x1;
- } else if (info->gt == 2) {
- brw->perfquery.sys_vars.n_eus = 20;
- brw->perfquery.sys_vars.n_eu_slices = 1;
- brw->perfquery.sys_vars.subslice_mask = 0x3;
- } else if (info->gt == 3) {
- brw->perfquery.sys_vars.n_eus = 40;
- brw->perfquery.sys_vars.n_eu_slices = 2;
- brw->perfquery.sys_vars.subslice_mask = 0xf;
- } else
- unreachable("not reached");
+ free(topo_info);
- return true;
- } else
- return false;
+ return true;
}
static bool
-get_sysfs_dev_dir(struct brw_context *brw,
- char *path_buf,
- int path_buf_len)
+getparam_topology(struct brw_context *brw)
{
__DRIscreen *screen = brw->screen->driScrnPriv;
- struct stat sb;
- int min, maj;
- DIR *drmdir;
- struct dirent *drm_entry;
- int len;
-
- assert(path_buf);
- assert(path_buf_len);
- path_buf[0] = '\0';
-
- if (fstat(screen->fd, &sb)) {
- DBG("Failed to stat DRM fd\n");
+ drm_i915_getparam_t gp;
+ int ret;
+
+ int slice_mask = 0;
+ gp.param = I915_PARAM_SLICE_MASK;
+ gp.value = &slice_mask;
+ ret = drmIoctl(screen->fd, DRM_IOCTL_I915_GETPARAM, &gp);
+ if (ret)
return false;
- }
- maj = major(sb.st_rdev);
- min = minor(sb.st_rdev);
-
- if (!S_ISCHR(sb.st_mode)) {
- DBG("DRM fd is not a character device as expected\n");
+ int subslice_mask = 0;
+ gp.param = I915_PARAM_SUBSLICE_MASK;
+ gp.value = &subslice_mask;
+ ret = drmIoctl(screen->fd, DRM_IOCTL_I915_GETPARAM, &gp);
+ if (ret)
return false;
- }
- len = snprintf(path_buf, path_buf_len,
- "/sys/dev/char/%d:%d/device/drm", maj, min);
- if (len < 0 || len >= path_buf_len) {
- DBG("Failed to concatenate sysfs path to drm device\n");
+ if (!gen_device_info_update_from_masks(&brw->screen->devinfo,
+ slice_mask,
+ subslice_mask,
+ brw->screen->eu_total))
return false;
- }
-
- drmdir = opendir(path_buf);
- if (!drmdir) {
- DBG("Failed to open %s: %m\n", path_buf);
- return false;
- }
- while ((drm_entry = readdir(drmdir))) {
- if ((drm_entry->d_type == DT_DIR ||
- drm_entry->d_type == DT_LNK) &&
- strncmp(drm_entry->d_name, "card", 4) == 0)
- {
- len = snprintf(path_buf, path_buf_len,
- "/sys/dev/char/%d:%d/device/drm/%s",
- maj, min, drm_entry->d_name);
- closedir(drmdir);
- if (len < 0 || len >= path_buf_len)
- return false;
- else
- return true;
- }
- }
-
- closedir(drmdir);
-
- DBG("Failed to find cardX directory under /sys/dev/char/%d:%d/device/drm\n",
- maj, min);
-
- return false;
+ return true;
}
static unsigned
brw_init_perf_query_info(struct gl_context *ctx)
{
struct brw_context *brw = brw_context(ctx);
- struct stat sb;
- char sysfs_dev_dir[128];
+ const struct gen_device_info *devinfo = &brw->screen->devinfo;
+ __DRIscreen *screen = brw->screen->driScrnPriv;
- if (brw->perfquery.n_queries)
- return brw->perfquery.n_queries;
+ if (brw->perfquery.perf)
+ return brw->perfquery.perf->n_queries;
+
+ brw->perfquery.perf = gen_perf_new(brw, drmIoctl);
init_pipeline_statistic_query_registers(brw);
+ brw_perf_query_register_mdapi_statistic_query(brw);
- /* The existence of this sysctl parameter implies the kernel supports
- * the i915 perf interface.
- */
- if (brw->is_haswell &&
- stat("/proc/sys/dev/i915/perf_stream_paranoid", &sb) == 0 &&
- get_sysfs_dev_dir(brw, sysfs_dev_dir, sizeof(sysfs_dev_dir)) &&
- init_oa_sys_vars(brw, sysfs_dev_dir))
- {
- brw->perfquery.oa_metrics_table =
- _mesa_hash_table_create(NULL, _mesa_key_hash_string,
- _mesa_key_string_equal);
+ if (!query_topology(brw)) {
+ /* We need the i915 query uAPI on CNL+ (kernel 4.17+). */
+ if (devinfo->gen >= 10)
+ return false;
- /* Index all the metric sets mesa knows about before looking to
- * see what the kernel is advertising.
- */
- brw_oa_register_queries_hsw(brw);
+ if (!getparam_topology(brw)) {
+ /* We need the SLICE_MASK/SUBSLICE_MASK on gen8+ (kernel 4.13+). */
+ if (devinfo->gen >= 8)
+ return false;
- enumerate_sysfs_metrics(brw, sysfs_dev_dir);
+ /* On Haswell, the values are already computed for us in
+ * gen_device_info.
+ */
+ }
}
+ if (gen_perf_load_oa_metrics(brw->perfquery.perf, screen->fd, devinfo))
+ brw_perf_query_register_mdapi_oa_query(brw);
+
brw->perfquery.unaccumulated =
ralloc_array(brw, struct brw_perf_query_object *, 2);
brw->perfquery.unaccumulated_elements = 0;
brw->perfquery.next_query_start_report_id = 1000;
- return brw->perfquery.n_queries;
+ return brw->perfquery.perf->n_queries;
}
void
projects / mesa.git / blobdiff