#include "dev/gen_debug.h"
#include "dev/gen_device_info.h"
#include "util/bitscan.h"
+#include "util/u_math.h"
#define FILE_DEBUG_FLAG DEBUG_PERFMON
+#define MI_RPC_BO_SIZE 4096
+#define MI_FREQ_START_OFFSET_BYTES (3072)
+#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)
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)
raw_query->oa_metrics_set_id = 0;
}
}
+
+bool
+gen_perf_open(struct gen_perf_context *perf_ctx,
+ int metrics_set_id,
+ int report_format,
+ int period_exponent,
+ int drm_fd,
+ uint32_t ctx_id)
+{
+ uint64_t properties[] = {
+ /* Single context sampling */
+ DRM_I915_PERF_PROP_CTX_HANDLE, ctx_id,
+
+ /* Include OA reports in samples */
+ DRM_I915_PERF_PROP_SAMPLE_OA, true,
+
+ /* OA unit configuration */
+ DRM_I915_PERF_PROP_OA_METRICS_SET, metrics_set_id,
+ DRM_I915_PERF_PROP_OA_FORMAT, report_format,
+ DRM_I915_PERF_PROP_OA_EXPONENT, period_exponent,
+ };
+ struct drm_i915_perf_open_param param = {
+ .flags = I915_PERF_FLAG_FD_CLOEXEC |
+ I915_PERF_FLAG_FD_NONBLOCK |
+ I915_PERF_FLAG_DISABLED,
+ .num_properties = ARRAY_SIZE(properties) / 2,
+ .properties_ptr = (uintptr_t) properties,
+ };
+ int fd = gen_ioctl(drm_fd, DRM_IOCTL_I915_PERF_OPEN, ¶m);
+ if (fd == -1) {
+ DBG("Error opening gen perf OA stream: %m\n");
+ return false;
+ }
+
+ perf_ctx->oa_stream_fd = fd;
+
+ perf_ctx->current_oa_metrics_set_id = metrics_set_id;
+ perf_ctx->current_oa_format = report_format;
+
+ return true;
+}
+
+bool
+gen_perf_inc_n_users(struct gen_perf_context *perf_ctx)
+{
+ if (perf_ctx->n_oa_users == 0 &&
+ gen_ioctl(perf_ctx->oa_stream_fd, I915_PERF_IOCTL_ENABLE, 0) < 0)
+ {
+ return false;
+ }
+ ++perf_ctx->n_oa_users;
+
+ return true;
+}
+
+void
+gen_perf_dec_n_users(struct gen_perf_context *perf_ctx)
+{
+ /* Disabling the i915 perf stream will effectively disable the OA
+ * counters. Note it's important to be sure there are no outstanding
+ * MI_RPC commands at this point since they could stall the CS
+ * indefinitely once OACONTROL is disabled.
+ */
+ --perf_ctx->n_oa_users;
+ if (perf_ctx->n_oa_users == 0 &&
+ gen_ioctl(perf_ctx->oa_stream_fd, I915_PERF_IOCTL_DISABLE, 0) < 0)
+ {
+ DBG("WARNING: Error disabling gen perf stream: %m\n");
+ }
+}
+
+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,
+ void * ctx, /* driver context (eg, brw_context) */
+ void * bufmgr, /* eg brw_bufmgr */
+ const struct gen_device_info *devinfo,
+ uint32_t hw_ctx,
+ int drm_fd)
+{
+ perf_ctx->perf = perf_cfg;
+ perf_ctx->ctx = ctx;
+ perf_ctx->bufmgr = bufmgr;
+ perf_ctx->drm_fd = drm_fd;
+ perf_ctx->hw_ctx = hw_ctx;
+ perf_ctx->devinfo = devinfo;
+
+ perf_ctx->unaccumulated =
+ ralloc_array(ctx, struct gen_perf_query_object *, 2);
+ perf_ctx->unaccumulated_elements = 0;
+ perf_ctx->unaccumulated_array_size = 2;
+
+ exec_list_make_empty(&perf_ctx->sample_buffers);
+ exec_list_make_empty(&perf_ctx->free_sample_buffers);
+
+ /* It's convenient to guarantee that this linked list of sample
+ * buffers is never empty so we add an empty head so when we
+ * Begin an OA query we can always take a reference on a buffer
+ * in this list.
+ */
+ struct oa_sample_buf *buf = gen_perf_get_free_sample_buf(perf_ctx);
+ exec_list_push_head(&perf_ctx->sample_buffers, &buf->link);
+
+ perf_ctx->oa_stream_fd = -1;
+ perf_ctx->next_query_start_report_id = 1000;
+}
+
+/**
+ * Add a query to the global list of "unaccumulated queries."
+ *
+ * Queries are tracked here until all the associated OA reports have
+ * been accumulated via accumulate_oa_reports() after the end
+ * MI_REPORT_PERF_COUNT has landed in query->oa.bo.
+ */
+static void
+add_to_unaccumulated_query_list(struct gen_perf_context *perf_ctx,
+ struct gen_perf_query_object *obj)
+{
+ if (perf_ctx->unaccumulated_elements >=
+ perf_ctx->unaccumulated_array_size)
+ {
+ perf_ctx->unaccumulated_array_size *= 1.5;
+ perf_ctx->unaccumulated =
+ reralloc(perf_ctx->ctx, perf_ctx->unaccumulated,
+ struct gen_perf_query_object *,
+ perf_ctx->unaccumulated_array_size);
+ }
+
+ perf_ctx->unaccumulated[perf_ctx->unaccumulated_elements++] = obj;
+}
+
+bool
+gen_perf_begin_query(struct gen_perf_context *perf_ctx,
+ struct gen_perf_query_object *query)
+{
+ struct gen_perf_config *perf_cfg = perf_ctx->perf;
+ const struct gen_perf_query_info *queryinfo = query->queryinfo;
+
+ /* XXX: We have to consider that the command parser unit that parses batch
+ * buffer commands and is used to capture begin/end counter snapshots isn't
+ * implicitly synchronized with what's currently running across other GPU
+ * units (such as the EUs running shaders) that the performance counters are
+ * associated with.
+ *
+ * The intention of performance queries is to measure the work associated
+ * with commands between the begin/end delimiters and so for that to be the
+ * case we need to explicitly synchronize the parsing of commands to capture
+ * Begin/End counter snapshots with what's running across other parts of the
+ * GPU.
+ *
+ * When the command parser reaches a Begin marker it effectively needs to
+ * drain everything currently running on the GPU until the hardware is idle
+ * before capturing the first snapshot of counters - otherwise the results
+ * would also be measuring the effects of earlier commands.
+ *
+ * When the command parser reaches an End marker it needs to stall until
+ * everything currently running on the GPU has finished before capturing the
+ * end snapshot - otherwise the results won't be a complete representation
+ * of the work.
+ *
+ * Theoretically there could be opportunities to minimize how much of the
+ * GPU pipeline is drained, or that we stall for, when we know what specific
+ * units the performance counters being queried relate to but we don't
+ * currently attempt to be clever here.
+ *
+ * Note: with our current simple approach here then for back-to-back queries
+ * we will redundantly emit duplicate commands to synchronize the command
+ * streamer with the rest of the GPU pipeline, but we assume that in HW the
+ * second synchronization is effectively a NOOP.
+ *
+ * N.B. The final results are based on deltas of counters between (inside)
+ * Begin/End markers so even though the total wall clock time of the
+ * workload is stretched by larger pipeline bubbles the bubbles themselves
+ * are generally invisible to the query results. Whether that's a good or a
+ * bad thing depends on the use case. For a lower real-time impact while
+ * capturing metrics then periodic sampling may be a better choice than
+ * INTEL_performance_query.
+ *
+ *
+ * This is our Begin synchronization point to drain current work on the
+ * GPU before we capture our first counter snapshot...
+ */
+ perf_cfg->vtbl.emit_mi_flush(perf_ctx->ctx);
+
+ switch (queryinfo->kind) {
+ 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
+ * specific layout/format so we can't begin any OA based queries that
+ * 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 = gen_perf_query_get_metric_id(perf_ctx->perf, queryinfo);
+
+ if (perf_ctx->oa_stream_fd != -1 &&
+ perf_ctx->current_oa_metrics_set_id != metric_id) {
+
+ if (perf_ctx->n_oa_users != 0) {
+ DBG("WARNING: Begin failed already using perf config=%i/%"PRIu64"\n",
+ perf_ctx->current_oa_metrics_set_id, metric_id);
+ return false;
+ } else
+ gen_perf_close(perf_ctx, queryinfo);
+ }
+
+ /* If the OA counters aren't already on, enable them. */
+ if (perf_ctx->oa_stream_fd == -1) {
+ const struct gen_device_info *devinfo = perf_ctx->devinfo;
+
+ /* The period_exponent gives a sampling period as follows:
+ * sample_period = timestamp_period * 2^(period_exponent + 1)
+ *
+ * The timestamps increments every 80ns (HSW), ~52ns (GEN9LP) or
+ * ~83ns (GEN8/9).
+ *
+ * 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)
+ *
+ * (E.g. 40 EUs @ 1GHz = ~53ms)
+ *
+ * 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.
+ */
+
+ 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) / (perf_cfg->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, perf_cfg->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 (!gen_perf_open(perf_ctx, metric_id, queryinfo->oa_format,
+ period_exponent, perf_ctx->drm_fd,
+ perf_ctx->hw_ctx))
+ return false;
+ } else {
+ assert(perf_ctx->current_oa_metrics_set_id == metric_id &&
+ perf_ctx->current_oa_format == queryinfo->oa_format);
+ }
+
+ if (!gen_perf_inc_n_users(perf_ctx)) {
+ DBG("WARNING: Error enabling i915 perf stream: %m\n");
+ return false;
+ }
+
+ if (query->oa.bo) {
+ perf_cfg->vtbl.bo_unreference(query->oa.bo);
+ query->oa.bo = NULL;
+ }
+
+ query->oa.bo = perf_cfg->vtbl.bo_alloc(perf_ctx->bufmgr,
+ "perf. query OA MI_RPC bo",
+ MI_RPC_BO_SIZE);
+#ifdef DEBUG
+ /* Pre-filling the BO helps debug whether writes landed. */
+ void *map = perf_cfg->vtbl.bo_map(perf_ctx->ctx, query->oa.bo, MAP_WRITE);
+ memset(map, 0x80, MI_RPC_BO_SIZE);
+ perf_cfg->vtbl.bo_unmap(query->oa.bo);
+#endif
+
+ query->oa.begin_report_id = perf_ctx->next_query_start_report_id;
+ perf_ctx->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.
+ */
+ perf_cfg->vtbl.batchbuffer_flush(perf_ctx->ctx, __FILE__, __LINE__);
+
+ /* Take a starting OA counter snapshot. */
+ perf_cfg->vtbl.emit_mi_report_perf_count(perf_ctx->ctx, query->oa.bo, 0,
+ query->oa.begin_report_id);
+ perf_cfg->vtbl.capture_frequency_stat_register(perf_ctx->ctx, query->oa.bo,
+ MI_FREQ_START_OFFSET_BYTES);
+
+ ++perf_ctx->n_active_oa_queries;
+
+ /* No already-buffered samples can possibly be associated with this query
+ * so create a marker within the list of sample buffers enabling us to
+ * easily ignore earlier samples when processing this query after
+ * completion.
+ */
+ assert(!exec_list_is_empty(&perf_ctx->sample_buffers));
+ query->oa.samples_head = exec_list_get_tail(&perf_ctx->sample_buffers);
+
+ struct oa_sample_buf *buf =
+ exec_node_data(struct oa_sample_buf, query->oa.samples_head, link);
+
+ /* This reference will ensure that future/following sample
+ * buffers (that may relate to this query) can't be freed until
+ * this drops to zero.
+ */
+ buf->refcount++;
+
+ gen_perf_query_result_clear(&query->oa.result);
+ query->oa.results_accumulated = false;
+
+ add_to_unaccumulated_query_list(perf_ctx, query);
+ 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;
+ }
+
+ query->pipeline_stats.bo =
+ perf_cfg->vtbl.bo_alloc(perf_ctx->bufmgr,
+ "perf. query pipeline stats bo",
+ STATS_BO_SIZE);
+
+ /* Take starting snapshots. */
+ gen_perf_snapshot_statistics_registers(perf_ctx->ctx , perf_cfg, query, 0);
+
+ ++perf_ctx->n_active_pipeline_stats_queries;
+ break;
+
+ default:
+ unreachable("Unknown query type");
+ break;
+ }
+
+ return true;
+}
+
+void
+gen_perf_end_query(struct gen_perf_context *perf_ctx,
+ struct gen_perf_query_object *query)
+{
+ struct gen_perf_config *perf_cfg = perf_ctx->perf;
+
+ /* Ensure that the work associated with the queried commands will have
+ * finished before taking our query end counter readings.
+ *
+ * For more details see comment in brw_begin_perf_query for
+ * corresponding flush.
+ */
+ perf_cfg->vtbl.emit_mi_flush(perf_ctx->ctx);
+
+ switch (query->queryinfo->kind) {
+ 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
+ * from perf. In this case we mustn't try and emit a closing
+ * MI_RPC command in case the OA unit has already been disabled
+ */
+ if (!query->oa.results_accumulated) {
+ /* Take an ending OA counter snapshot. */
+ perf_cfg->vtbl.capture_frequency_stat_register(perf_ctx->ctx, query->oa.bo,
+ MI_FREQ_END_OFFSET_BYTES);
+ perf_cfg->vtbl.emit_mi_report_perf_count(perf_ctx->ctx, query->oa.bo,
+ MI_RPC_BO_END_OFFSET_BYTES,
+ query->oa.begin_report_id + 1);
+ }
+
+ --perf_ctx->n_active_oa_queries;
+
+ /* NB: even though the query has now ended, it can't be accumulated
+ * until the end MI_REPORT_PERF_COUNT snapshot has been written
+ * to query->oa.bo
+ */
+ break;
+
+ case GEN_PERF_QUERY_TYPE_PIPELINE:
+ gen_perf_snapshot_statistics_registers(perf_ctx->ctx, perf_cfg, query,
+ STATS_BO_END_OFFSET_BYTES);
+ --perf_ctx->n_active_pipeline_stats_queries;
+ break;
+
+ default:
+ unreachable("Unknown query type");
+ break;
+ }
+}
+
+enum OaReadStatus {
+ OA_READ_STATUS_ERROR,
+ OA_READ_STATUS_UNFINISHED,
+ OA_READ_STATUS_FINISHED,
+};
+
+static enum OaReadStatus
+read_oa_samples_until(struct gen_perf_context *perf_ctx,
+ uint32_t start_timestamp,
+ uint32_t end_timestamp)
+{
+ struct exec_node *tail_node =
+ exec_list_get_tail(&perf_ctx->sample_buffers);
+ struct oa_sample_buf *tail_buf =
+ exec_node_data(struct oa_sample_buf, tail_node, link);
+ uint32_t last_timestamp = tail_buf->last_timestamp;
+
+ while (1) {
+ struct oa_sample_buf *buf = gen_perf_get_free_sample_buf(perf_ctx);
+ uint32_t offset;
+ int len;
+
+ while ((len = read(perf_ctx->oa_stream_fd, buf->buf,
+ sizeof(buf->buf))) < 0 && errno == EINTR)
+ ;
+
+ if (len <= 0) {
+ exec_list_push_tail(&perf_ctx->free_sample_buffers, &buf->link);
+
+ if (len < 0) {
+ if (errno == EAGAIN)
+ 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");
+ }
+ } else
+ DBG("Spurious EOF reading i915 perf samples\n");
+
+ return OA_READ_STATUS_ERROR;
+ }
+
+ buf->len = len;
+ exec_list_push_tail(&perf_ctx->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 gen_perf_context *perf_ctx,
+ struct gen_perf_query_object *query,
+ void *current_batch)
+{
+ uint32_t *start;
+ uint32_t *last;
+ uint32_t *end;
+ struct gen_perf_config *perf_cfg = perf_ctx->perf;
+
+ /* We need the MI_REPORT_PERF_COUNT to land before we can start
+ * accumulate. */
+ assert(!perf_cfg->vtbl.batch_references(current_batch, query->oa.bo) &&
+ !perf_cfg->vtbl.bo_busy(query->oa.bo));
+
+ /* Map the BO once here and let accumulate_oa_reports() unmap
+ * it. */
+ if (query->oa.map == NULL)
+ query->oa.map = perf_cfg->vtbl.bo_map(perf_ctx->ctx, query->oa.bo, MAP_READ);
+
+ 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]);
+ return true;
+ }
+ if (end[0] != (query->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(perf_ctx, 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;
+}
+
+void
+gen_perf_wait_query(struct gen_perf_context *perf_ctx,
+ struct gen_perf_query_object *query,
+ void *current_batch)
+{
+ struct gen_perf_config *perf_cfg = perf_ctx->perf;
+ struct brw_bo *bo = NULL;
+
+ switch (query->queryinfo->kind) {
+ case GEN_PERF_QUERY_TYPE_OA:
+ case GEN_PERF_QUERY_TYPE_RAW:
+ bo = query->oa.bo;
+ break;
+
+ case GEN_PERF_QUERY_TYPE_PIPELINE:
+ bo = query->pipeline_stats.bo;
+ break;
+
+ default:
+ unreachable("Unknown query type");
+ break;
+ }
+
+ if (bo == NULL)
+ return;
+
+ /* If the current batch references our results bo then we need to
+ * flush first...
+ */
+ if (perf_cfg->vtbl.batch_references(current_batch, bo))
+ perf_cfg->vtbl.batchbuffer_flush(perf_ctx->ctx, __FILE__, __LINE__);
+
+ perf_cfg->vtbl.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 (query->queryinfo->kind == GEN_PERF_QUERY_TYPE_OA ||
+ query->queryinfo->kind == GEN_PERF_QUERY_TYPE_RAW) {
+ while (!read_oa_samples_for_query(perf_ctx, query, current_batch))
+ ;
+ }
+}
+
+bool
+gen_perf_is_query_ready(struct gen_perf_context *perf_ctx,
+ struct gen_perf_query_object *query,
+ void *current_batch)
+{
+ struct gen_perf_config *perf_cfg = perf_ctx->perf;
+
+ switch (query->queryinfo->kind) {
+ case GEN_PERF_QUERY_TYPE_OA:
+ case GEN_PERF_QUERY_TYPE_RAW:
+ return (query->oa.results_accumulated ||
+ (query->oa.bo &&
+ !perf_cfg->vtbl.batch_references(current_batch, query->oa.bo) &&
+ !perf_cfg->vtbl.bo_busy(query->oa.bo) &&
+ read_oa_samples_for_query(perf_ctx, query, current_batch)));
+ case GEN_PERF_QUERY_TYPE_PIPELINE:
+ return (query->pipeline_stats.bo &&
+ !perf_cfg->vtbl.batch_references(current_batch, query->pipeline_stats.bo) &&
+ !perf_cfg->vtbl.bo_busy(query->pipeline_stats.bo));
+
+ default:
+ unreachable("Unknown query type");
+ break;
+ }
+
+ 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;
+}
projects / mesa.git / blobdiff