*/
#include <limits.h>
-#include <dirent.h>
/* put before sys/types.h to silence glibc warnings */
#ifdef MAJOR_IN_MKDEV
#include "brw_context.h"
#include "brw_defines.h"
#include "brw_performance_query.h"
-#include "brw_oa_metrics.h"
#include "intel_batchbuffer.h"
+#include "perf/gen_perf.h"
+
#define FILE_DEBUG_FLAG DEBUG_PERFMON
#define OAREPORT_REASON_MASK 0x3f
/******************************************************************************/
-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;
- while ((n = read(fd, buf, sizeof (buf) - 1)) < 0 &&
- errno == EINTR);
- close(fd);
- if (n < 0)
- return false;
-
- buf[n] = '\0';
- *val = strtoull(buf, NULL, 0);
-
- return true;
-}
-
-static bool
-read_sysfs_drm_device_file_uint64(struct brw_context *brw,
- const char *file,
- uint64_t *value)
-{
- char buf[512];
- int len;
-
- len = snprintf(buf, sizeof(buf), "%s/%s",
- brw->perfquery.sysfs_dev_dir, file);
- if (len < 0 || len >= sizeof(buf)) {
- DBG("Failed to concatenate sys filename to read u64 from\n");
- return false;
- }
-
- return read_file_uint64(buf, value);
-}
-
-/******************************************************************************/
-
static bool
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 brw_perf_query_info *query)
+ 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 == OA_COUNTERS)
+ if (query->kind == GEN_PERF_QUERY_TYPE_OA)
return query->oa_metrics_set_id;
- assert(query->kind == OA_COUNTERS_RAW);
+ 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 !=
return query->oa_metrics_set_id;
}
- char metric_id_file[280];
- snprintf(metric_id_file, sizeof(metric_id_file),
- "%s/metrics/%s/id", brw->perfquery.sysfs_dev_dir, query->guid);
-
- struct brw_perf_query_info *raw_query = (struct brw_perf_query_info *)query;
- if (!read_file_uint64(metric_id_file, &raw_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 {
struct brw_perf_query_object *obj = query_void;
switch (obj->query->kind) {
- case OA_COUNTERS:
- case OA_COUNTERS_RAW:
+ 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,",
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 OA_COUNTERS_RAW:
+ 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;
}
}
+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");
+ }
+}
+
/**
* Driver hook for glGetPerfCounterInfoINTEL().
*/
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;
+ *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,
const uint32_t *start,
const uint32_t *end)
{
- const struct brw_perf_query_info *query = obj->query;
+ const struct gen_perf_query_info *query = obj->query;
uint64_t *accumulator = obj->oa.accumulator;
int idx = 0;
int i;
switch (query->oa_format) {
case I915_OA_FORMAT_A32u40_A4u32_B8_C8:
- brw_perf_query_accumulate_uint32(start + 1, end + 1, accumulator + idx++); /* timestamp */
- brw_perf_query_accumulate_uint32(start + 3, end + 3, accumulator + idx++); /* clock */
+ gen_perf_query_accumulate_uint32(start + 1, end + 1, accumulator + idx++); /* timestamp */
+ gen_perf_query_accumulate_uint32(start + 3, end + 3, accumulator + idx++); /* clock */
/* 32x 40bit A counters... */
for (i = 0; i < 32; i++)
- brw_perf_query_accumulate_uint40(i, start, end, accumulator + idx++);
+ gen_perf_query_accumulate_uint40(i, start, end, accumulator + idx++);
/* 4x 32bit A counters... */
for (i = 0; i < 4; i++)
- brw_perf_query_accumulate_uint32(start + 36 + i, end + 36 + i,
+ gen_perf_query_accumulate_uint32(start + 36 + i, end + 36 + i,
accumulator + idx++);
/* 8x 32bit B counters + 8x 32bit C counters... */
for (i = 0; i < 16; i++)
- brw_perf_query_accumulate_uint32(start + 48 + i, end + 48 + i,
+ gen_perf_query_accumulate_uint32(start + 48 + i, end + 48 + i,
accumulator + idx++);
break;
case I915_OA_FORMAT_A45_B8_C8:
- brw_perf_query_accumulate_uint32(start + 1, end + 1, accumulator); /* timestamp */
+ gen_perf_query_accumulate_uint32(start + 1, end + 1, accumulator); /* timestamp */
for (i = 0; i < 61; i++)
- brw_perf_query_accumulate_uint32(start + 3 + i, end + 3 + i, accumulator + 1 + i);
+ gen_perf_query_accumulate_uint32(start + 3 + i, end + 3 + i, accumulator + 1 + i);
break;
default:
static void
close_perf(struct brw_context *brw,
- const struct brw_perf_query_info *query)
+ 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 == OA_COUNTERS_RAW) {
- struct brw_perf_query_info *raw_query =
- (struct brw_perf_query_info *) query;
+ 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;
}
}
{
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 OA_COUNTERS_RAW: {
+ 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
a_counter_in_bits = 40;
uint64_t overflow_period = pow(2, a_counter_in_bits) /
- (brw->perfquery.sys_vars.n_eus *
+ (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.sys_vars.n_eus);
+ overflow_period, overflow_period / 1000000ul, brw->perfquery.perf->sys_vars.n_eus);
int period_exponent = 0;
uint64_t prev_sample_period, next_sample_period;
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;
brw_emit_mi_flush(brw);
switch (obj->query->kind) {
- case OA_COUNTERS:
- case OA_COUNTERS_RAW:
+ 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
*/
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;
assert(!o->Ready);
switch (obj->query->kind) {
- case OA_COUNTERS:
- case OA_COUNTERS_RAW:
+ 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;
* we need to wait for all the reports to come in before we can
* read them.
*/
- if (obj->query->kind == OA_COUNTERS ||
- obj->query->kind == OA_COUNTERS_RAW) {
+ 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))
;
}
return true;
switch (obj->query->kind) {
- case OA_COUNTERS:
- case OA_COUNTERS_RAW:
+ 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) &&
read_oa_samples_for_query(brw, obj)));
- case PIPELINE_STATS:
+ 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));
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;
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;
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,
+ *out_uint64 = counter->oa_counter_read_uint64(perf, query,
obj->oa.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,
+ *out_float = counter->oa_counter_read_float(perf, query,
obj->oa.accumulator);
break;
default:
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;
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:
- case OA_COUNTERS_RAW:
+ 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);
brw_bo_unmap(obj->oa.bo);
obj->oa.map = NULL;
}
- if (obj->query->kind == OA_COUNTERS)
+ if (obj->query->kind == GEN_PERF_QUERY_TYPE_OA)
written = get_oa_counter_data(brw, obj, data_size, (uint8_t *)data);
else
written = brw_perf_query_get_mdapi_oa_data(brw, obj, data_size, (uint8_t *)data);
break;
- case PIPELINE_STATS:
+ case GEN_PERF_QUERY_TYPE_PIPELINE:
written = get_pipeline_stats_data(brw, obj, data_size, (uint8_t *)data);
break;
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 OA_COUNTERS_RAW:
+ 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;
init_pipeline_statistic_query_registers(struct brw_context *brw)
{
const struct gen_device_info *devinfo = &brw->screen->devinfo;
- struct brw_perf_query_info *query = brw_perf_query_append_query_info(brw);
+ 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);
- brw_perf_query_info_add_basic_stat_reg(query, IA_VERTICES_COUNT,
- "N vertices submitted");
- brw_perf_query_info_add_basic_stat_reg(query, IA_PRIMITIVES_COUNT,
- "N primitives submitted");
- brw_perf_query_info_add_basic_stat_reg(query, VS_INVOCATION_COUNT,
- "N vertex shader invocations");
+ 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) {
- brw_perf_query_info_add_stat_reg(query, GEN6_SO_PRIM_STORAGE_NEEDED, 1, 1,
+ 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)");
- brw_perf_query_info_add_stat_reg(query, GEN6_SO_NUM_PRIMS_WRITTEN, 1, 1,
+ 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 {
- brw_perf_query_info_add_stat_reg(query, GEN7_SO_PRIM_STORAGE_NEEDED(0), 1, 1,
+ 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)");
- brw_perf_query_info_add_stat_reg(query, GEN7_SO_PRIM_STORAGE_NEEDED(1), 1, 1,
+ 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)");
- brw_perf_query_info_add_stat_reg(query, GEN7_SO_PRIM_STORAGE_NEEDED(2), 1, 1,
+ 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)");
- brw_perf_query_info_add_stat_reg(query, GEN7_SO_PRIM_STORAGE_NEEDED(3), 1, 1,
+ 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)");
- brw_perf_query_info_add_stat_reg(query, GEN7_SO_NUM_PRIMS_WRITTEN(0), 1, 1,
+ 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)");
- brw_perf_query_info_add_stat_reg(query, GEN7_SO_NUM_PRIMS_WRITTEN(1), 1, 1,
+ 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)");
- brw_perf_query_info_add_stat_reg(query, GEN7_SO_NUM_PRIMS_WRITTEN(2), 1, 1,
+ 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)");
- brw_perf_query_info_add_stat_reg(query, GEN7_SO_NUM_PRIMS_WRITTEN(3), 1, 1,
+ 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)");
}
- brw_perf_query_info_add_basic_stat_reg(query, HS_INVOCATION_COUNT,
+ gen_perf_query_info_add_basic_stat_reg(query, HS_INVOCATION_COUNT,
"N TCS shader invocations");
- brw_perf_query_info_add_basic_stat_reg(query, DS_INVOCATION_COUNT,
+ gen_perf_query_info_add_basic_stat_reg(query, DS_INVOCATION_COUNT,
"N TES shader invocations");
- brw_perf_query_info_add_basic_stat_reg(query, GS_INVOCATION_COUNT,
+ gen_perf_query_info_add_basic_stat_reg(query, GS_INVOCATION_COUNT,
"N geometry shader invocations");
- brw_perf_query_info_add_basic_stat_reg(query, GS_PRIMITIVES_COUNT,
+ gen_perf_query_info_add_basic_stat_reg(query, GS_PRIMITIVES_COUNT,
"N geometry shader primitives emitted");
- brw_perf_query_info_add_basic_stat_reg(query, CL_INVOCATION_COUNT,
+ gen_perf_query_info_add_basic_stat_reg(query, CL_INVOCATION_COUNT,
"N primitives entering clipping");
- brw_perf_query_info_add_basic_stat_reg(query, CL_PRIMITIVES_COUNT,
+ gen_perf_query_info_add_basic_stat_reg(query, CL_PRIMITIVES_COUNT,
"N primitives leaving clipping");
- if (devinfo->is_haswell || devinfo->gen == 8)
- brw_perf_query_info_add_stat_reg(query, PS_INVOCATION_COUNT, 1, 4,
+ 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
- brw_perf_query_info_add_basic_stat_reg(query, PS_INVOCATION_COUNT,
+ } else {
+ gen_perf_query_info_add_basic_stat_reg(query, PS_INVOCATION_COUNT,
"N fragment shader invocations");
-
- brw_perf_query_info_add_basic_stat_reg(query, PS_DEPTH_COUNT, "N z-pass fragments");
-
- if (devinfo->gen >= 7)
- brw_perf_query_info_add_basic_stat_reg(query, CS_INVOCATION_COUNT,
- "N compute shader invocations");
-
- query->data_size = sizeof(uint64_t) * query->n_counters;
-}
-
-static void
-register_oa_config(struct brw_context *brw,
- const struct brw_perf_query_info *query,
- uint64_t config_id)
-{
- struct brw_perf_query_info *registred_query =
- brw_perf_query_append_query_info(brw);
-
- *registred_query = *query;
- registred_query->oa_metrics_set_id = config_id;
- DBG("metric set registred: id = %" PRIu64", guid = %s\n",
- registred_query->oa_metrics_set_id, query->guid);
-}
-
-static void
-enumerate_sysfs_metrics(struct brw_context *brw)
-{
- char buf[256];
- DIR *metricsdir = NULL;
- struct dirent *metric_entry;
- int len;
-
- len = snprintf(buf, sizeof(buf), "%s/metrics", brw->perfquery.sysfs_dev_dir);
- if (len < 0 || len >= sizeof(buf)) {
- DBG("Failed to concatenate path to sysfs metrics/ directory\n");
- return;
- }
-
- metricsdir = opendir(buf);
- if (!metricsdir) {
- DBG("Failed to open %s: %m\n", buf);
- return;
- }
-
- 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) {
- uint64_t id;
-
- len = snprintf(buf, sizeof(buf), "%s/metrics/%s/id",
- brw->perfquery.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;
- }
-
- register_oa_config(brw, (const struct brw_perf_query_info *)entry->data, id);
- } else
- DBG("metric set not known by mesa (skipping)\n");
}
- closedir(metricsdir);
-}
-
-static bool
-kernel_has_dynamic_config_support(struct brw_context *brw)
-{
- __DRIscreen *screen = brw->screen->driScrnPriv;
-
- hash_table_foreach(brw->perfquery.oa_metrics_table, entry) {
- struct brw_perf_query_info *query = entry->data;
- char config_path[280];
- uint64_t config_id;
+ gen_perf_query_info_add_basic_stat_reg(query, PS_DEPTH_COUNT,
+ "N z-pass fragments");
- snprintf(config_path, sizeof(config_path), "%s/metrics/%s/id",
- brw->perfquery.sysfs_dev_dir, query->guid);
-
- /* Look for the test config, which we know we can't replace. */
- if (read_file_uint64(config_path, &config_id) && config_id == 1) {
- return drmIoctl(screen->fd, DRM_IOCTL_I915_PERF_REMOVE_CONFIG,
- &config_id) < 0 && errno == ENOENT;
- }
+ if (devinfo->gen >= 7) {
+ gen_perf_query_info_add_basic_stat_reg(query, CS_INVOCATION_COUNT,
+ "N compute shader invocations");
}
- return false;
-}
-
-static void
-init_oa_configs(struct brw_context *brw)
-{
- __DRIscreen *screen = brw->screen->driScrnPriv;
-
- hash_table_foreach(brw->perfquery.oa_metrics_table, entry) {
- const struct brw_perf_query_info *query = entry->data;
- struct drm_i915_perf_oa_config config;
- char config_path[280];
- uint64_t config_id;
- int ret;
-
- snprintf(config_path, sizeof(config_path), "%s/metrics/%s/id",
- brw->perfquery.sysfs_dev_dir, query->guid);
-
- /* Don't recreate already loaded configs. */
- if (read_file_uint64(config_path, &config_id)) {
- DBG("metric set: %s (already loaded)\n", query->guid);
- register_oa_config(brw, query, config_id);
- continue;
- }
-
- memset(&config, 0, sizeof(config));
-
- memcpy(config.uuid, query->guid, sizeof(config.uuid));
-
- config.n_mux_regs = query->n_mux_regs;
- config.mux_regs_ptr = (uintptr_t) query->mux_regs;
-
- config.n_boolean_regs = query->n_b_counter_regs;
- config.boolean_regs_ptr = (uintptr_t) query->b_counter_regs;
-
- config.n_flex_regs = query->n_flex_regs;
- config.flex_regs_ptr = (uintptr_t) query->flex_regs;
-
- ret = drmIoctl(screen->fd, DRM_IOCTL_I915_PERF_ADD_CONFIG, &config);
- if (ret < 0) {
- DBG("Failed to load \"%s\" (%s) metrics set in kernel: %s\n",
- query->name, query->guid, strerror(errno));
- continue;
- }
-
- register_oa_config(brw, query, ret);
- DBG("metric set: %s (added)\n", query->guid);
- }
+ query->data_size = sizeof(uint64_t) * query->n_counters;
}
static bool
return true;
}
-static void
-compute_topology_builtins(struct brw_context *brw)
+static unsigned
+brw_init_perf_query_info(struct gl_context *ctx)
{
+ struct brw_context *brw = brw_context(ctx);
const struct gen_device_info *devinfo = &brw->screen->devinfo;
+ __DRIscreen *screen = brw->screen->driScrnPriv;
- brw->perfquery.sys_vars.slice_mask = devinfo->slice_masks;
- brw->perfquery.sys_vars.n_eu_slices = devinfo->num_slices;
-
- for (int i = 0; i < sizeof(devinfo->subslice_masks[i]); i++) {
- brw->perfquery.sys_vars.n_eu_sub_slices +=
- util_bitcount(devinfo->subslice_masks[i]);
- }
-
- for (int i = 0; i < sizeof(devinfo->eu_masks); i++)
- brw->perfquery.sys_vars.n_eus += util_bitcount(devinfo->eu_masks[i]);
-
- brw->perfquery.sys_vars.eu_threads_count =
- brw->perfquery.sys_vars.n_eus * devinfo->num_thread_per_eu;
-
- /* The subslice mask builtin contains bits for all slices. Prior to Gen11
- * it had groups of 3bits for each slice, on Gen11 it's 8bits for each
- * slice.
- *
- * Ideally equations would be updated to have a slice/subslice query
- * function/operator.
- */
- brw->perfquery.sys_vars.subslice_mask = 0;
-
- int bits_per_subslice = devinfo->gen == 11 ? 8 : 3;
-
- for (int s = 0; s < util_last_bit(devinfo->slice_masks); s++) {
- for (int ss = 0; ss < (devinfo->subslice_slice_stride * 8); ss++) {
- if (gen_device_info_subslice_available(devinfo, s, ss))
- brw->perfquery.sys_vars.subslice_mask |= 1UL << (s * bits_per_subslice + ss);
- }
- }
-}
-
-static bool
-init_oa_sys_vars(struct brw_context *brw)
-{
- const struct gen_device_info *devinfo = &brw->screen->devinfo;
- uint64_t min_freq_mhz = 0, max_freq_mhz = 0;
+ if (brw->perfquery.perf)
+ return brw->perfquery.perf->n_queries;
- if (!read_sysfs_drm_device_file_uint64(brw, "gt_min_freq_mhz", &min_freq_mhz))
- return false;
+ brw->perfquery.perf = gen_perf_new(brw, drmIoctl);
- if (!read_sysfs_drm_device_file_uint64(brw, "gt_max_freq_mhz", &max_freq_mhz))
- return false;
+ init_pipeline_statistic_query_registers(brw);
+ brw_perf_query_register_mdapi_statistic_query(brw);
if (!query_topology(brw)) {
/* We need the i915 query uAPI on CNL+ (kernel 4.17+). */
}
}
- memset(&brw->perfquery.sys_vars, 0, sizeof(brw->perfquery.sys_vars));
- brw->perfquery.sys_vars.gt_min_freq = min_freq_mhz * 1000000;
- brw->perfquery.sys_vars.gt_max_freq = max_freq_mhz * 1000000;
- brw->perfquery.sys_vars.timestamp_frequency = devinfo->timestamp_frequency;
- brw->perfquery.sys_vars.revision = devinfo->revision;
- compute_topology_builtins(brw);
-
- return true;
-}
-
-static bool
-get_sysfs_dev_dir(struct brw_context *brw)
-{
- __DRIscreen *screen = brw->screen->driScrnPriv;
- struct stat sb;
- int min, maj;
- DIR *drmdir;
- struct dirent *drm_entry;
- int len;
-
- brw->perfquery.sysfs_dev_dir[0] = '\0';
-
- if (fstat(screen->fd, &sb)) {
- DBG("Failed to stat DRM fd\n");
- 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");
- return false;
- }
-
- len = snprintf(brw->perfquery.sysfs_dev_dir,
- sizeof(brw->perfquery.sysfs_dev_dir),
- "/sys/dev/char/%d:%d/device/drm", maj, min);
- if (len < 0 || len >= sizeof(brw->perfquery.sysfs_dev_dir)) {
- DBG("Failed to concatenate sysfs path to drm device\n");
- return false;
- }
-
- drmdir = opendir(brw->perfquery.sysfs_dev_dir);
- if (!drmdir) {
- DBG("Failed to open %s: %m\n", brw->perfquery.sysfs_dev_dir);
- 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(brw->perfquery.sysfs_dev_dir,
- sizeof(brw->perfquery.sysfs_dev_dir),
- "/sys/dev/char/%d:%d/device/drm/%s",
- maj, min, drm_entry->d_name);
- closedir(drmdir);
- if (len < 0 || len >= sizeof(brw->perfquery.sysfs_dev_dir))
- 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;
-}
-
-typedef void (*perf_register_oa_queries_t)(struct brw_context *);
-
-static perf_register_oa_queries_t
-get_register_queries_function(const struct gen_device_info *devinfo)
-{
- if (devinfo->is_haswell)
- return brw_oa_register_queries_hsw;
- if (devinfo->is_cherryview)
- return brw_oa_register_queries_chv;
- if (devinfo->is_broadwell)
- return brw_oa_register_queries_bdw;
- if (devinfo->is_broxton)
- return brw_oa_register_queries_bxt;
- if (devinfo->is_skylake) {
- if (devinfo->gt == 2)
- return brw_oa_register_queries_sklgt2;
- if (devinfo->gt == 3)
- return brw_oa_register_queries_sklgt3;
- if (devinfo->gt == 4)
- return brw_oa_register_queries_sklgt4;
- }
- if (devinfo->is_kabylake) {
- if (devinfo->gt == 2)
- return brw_oa_register_queries_kblgt2;
- if (devinfo->gt == 3)
- return brw_oa_register_queries_kblgt3;
- }
- if (devinfo->is_geminilake)
- return brw_oa_register_queries_glk;
- if (devinfo->is_coffeelake) {
- if (devinfo->gt == 2)
- return brw_oa_register_queries_cflgt2;
- if (devinfo->gt == 3)
- return brw_oa_register_queries_cflgt3;
- }
- if (devinfo->is_cannonlake)
- return brw_oa_register_queries_cnl;
- if (devinfo->gen == 11)
- return brw_oa_register_queries_icl;
-
- return NULL;
-}
-
-static unsigned
-brw_init_perf_query_info(struct gl_context *ctx)
-{
- struct brw_context *brw = brw_context(ctx);
- const struct gen_device_info *devinfo = &brw->screen->devinfo;
- bool i915_perf_oa_available = false;
- struct stat sb;
- perf_register_oa_queries_t oa_register;
-
- if (brw->perfquery.n_queries)
- return brw->perfquery.n_queries;
-
- init_pipeline_statistic_query_registers(brw);
- brw_perf_query_register_mdapi_statistic_query(brw);
-
- oa_register = get_register_queries_function(devinfo);
-
- /* The existence of this sysctl parameter implies the kernel supports
- * the i915 perf interface.
- */
- if (stat("/proc/sys/dev/i915/perf_stream_paranoid", &sb) == 0) {
-
- /* If _paranoid == 1 then on Gen8+ we won't be able to access OA
- * metrics unless running as root.
- */
- if (devinfo->is_haswell)
- i915_perf_oa_available = true;
- else {
- uint64_t paranoid = 1;
-
- read_file_uint64("/proc/sys/dev/i915/perf_stream_paranoid", ¶noid);
-
- if (paranoid == 0 || geteuid() == 0)
- i915_perf_oa_available = true;
- }
- }
-
- if (i915_perf_oa_available &&
- oa_register &&
- get_sysfs_dev_dir(brw) &&
- init_oa_sys_vars(brw))
- {
- brw->perfquery.oa_metrics_table =
- _mesa_hash_table_create(NULL, _mesa_key_hash_string,
- _mesa_key_string_equal);
-
- /* Index all the metric sets mesa knows about before looking to see what
- * the kernel is advertising.
- */
- oa_register(brw);
-
- if (likely((INTEL_DEBUG & DEBUG_NO_OACONFIG) == 0) &&
- kernel_has_dynamic_config_support(brw))
- init_oa_configs(brw);
- else
- enumerate_sysfs_metrics(brw);
-
+ 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.next_query_start_report_id = 1000;
- return brw->perfquery.n_queries;
+ return brw->perfquery.perf->n_queries;
}
void