#include "iris_screen.h"
#include "iris_context.h"
+#include "iris_perf.h"
-#include "perf/gen_perf.h"
+struct iris_monitor_object {
+ int num_active_counters;
+ int *active_counters;
-int iris_get_monitor_info(struct pipe_screen *pscreen, unsigned index,
- struct pipe_driver_query_info *info)
+ size_t result_size;
+ unsigned char *result_buffer;
+
+ struct gen_perf_query_object *query;
+};
+
+int
+iris_get_monitor_info(struct pipe_screen *pscreen, unsigned index,
+ struct pipe_driver_query_info *info)
{
const struct iris_screen *screen = (struct iris_screen *)pscreen;
assert(screen->monitor_cfg);
return 0;
const struct iris_monitor_config *monitor_cfg = screen->monitor_cfg;
- if (!info)
+
+ if (!info) {
/* return the number of metrics */
return monitor_cfg->num_counters;
+ }
+
const struct gen_perf_config *perf_cfg = monitor_cfg->perf_cfg;
const int group = monitor_cfg->counters[index].group;
const int counter_index = monitor_cfg->counters[index].counter;
- info->group_id = group;
struct gen_perf_query_counter *counter =
&perf_cfg->queries[group].counters[counter_index];
+
+ info->group_id = group;
info->name = counter->name;
info->query_type = PIPE_QUERY_DRIVER_SPECIFIC + index;
return 1;
}
-typedef void (*bo_unreference_t)(void *);
-typedef void *(*bo_map_t)(void *, void *, unsigned flags);
-typedef void (*bo_unmap_t)(void *);
-typedef void (* emit_mi_report_t)(void *, void *, uint32_t, uint32_t);
-typedef void (*emit_mi_flush_t)(void *);
-typedef void (*capture_frequency_stat_register_t)(void *, void *,
- uint32_t );
-typedef void (*store_register_mem64_t)(void *ctx, void *bo,
- uint32_t reg, uint32_t offset);
-typedef bool (*batch_references_t)(void *batch, void *bo);
-typedef void (*bo_wait_rendering_t)(void *bo);
-typedef int (*bo_busy_t)(void *bo);
-
-static void *
-iris_oa_bo_alloc(void *bufmgr,
- const char *name,
- uint64_t size)
-{
- return iris_bo_alloc(bufmgr, name, size, IRIS_MEMZONE_OTHER);
-}
-
-static void
-iris_monitor_emit_mi_flush(struct iris_context *ice)
-{
- const int flags = PIPE_CONTROL_RENDER_TARGET_FLUSH |
- PIPE_CONTROL_INSTRUCTION_INVALIDATE |
- PIPE_CONTROL_CONST_CACHE_INVALIDATE |
- PIPE_CONTROL_DATA_CACHE_FLUSH |
- PIPE_CONTROL_DEPTH_CACHE_FLUSH |
- PIPE_CONTROL_VF_CACHE_INVALIDATE |
- PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
- PIPE_CONTROL_CS_STALL;
- iris_emit_pipe_control_flush(&ice->batches[IRIS_BATCH_RENDER],
- "OA metrics",
- flags);
-}
-
-static void
-iris_monitor_emit_mi_report_perf_count(void *c,
- void *bo,
- uint32_t offset_in_bytes,
- uint32_t report_id)
-{
- struct iris_context *ice = c;
- ice->vtbl.emit_mi_report_perf_count(&ice->batches[IRIS_BATCH_RENDER],
- bo,
- offset_in_bytes,
- report_id);
-}
-
-static void
-iris_monitor_batchbuffer_flush(void *c, const char *file, int line)
-{
- struct iris_context *ice = c;
- _iris_batch_flush(&ice->batches[IRIS_BATCH_RENDER], __FILE__, __LINE__);
-}
-
-static void
-iris_monitor_capture_frequency_stat_register(void *ctx,
- void *bo,
- uint32_t bo_offset)
-{
- struct iris_context *ice = ctx;
- ice->vtbl.store_register_mem32(&ice->batches[IRIS_BATCH_RENDER],
- GEN9_RPSTAT0, bo, bo_offset, false);
-}
-
-static void
-iris_monitor_store_register_mem64(void *ctx, void *bo,
- uint32_t reg, uint32_t offset)
-{
- struct iris_context *ice = ctx;
- ice->vtbl.store_register_mem64(&ice->batches[IRIS_BATCH_RENDER], reg, bo,
- offset, false);
-}
-
-
static bool
iris_monitor_init_metrics(struct iris_screen *screen)
{
monitor_cfg->perf_cfg = perf_cfg;
- perf_cfg->vtbl.bo_alloc = iris_oa_bo_alloc;
- perf_cfg->vtbl.bo_unreference = (bo_unreference_t)iris_bo_unreference;
- perf_cfg->vtbl.bo_map = (bo_map_t)iris_bo_map;
- perf_cfg->vtbl.bo_unmap = (bo_unmap_t)iris_bo_unmap;
- perf_cfg->vtbl.emit_mi_flush = (emit_mi_flush_t)iris_monitor_emit_mi_flush;
-
- perf_cfg->vtbl.emit_mi_report_perf_count =
- (emit_mi_report_t)iris_monitor_emit_mi_report_perf_count;
- perf_cfg->vtbl.batchbuffer_flush = iris_monitor_batchbuffer_flush;
- perf_cfg->vtbl.capture_frequency_stat_register =
- (capture_frequency_stat_register_t) iris_monitor_capture_frequency_stat_register;
- perf_cfg->vtbl.store_register_mem64 =
- (store_register_mem64_t) iris_monitor_store_register_mem64;
- perf_cfg->vtbl.batch_references = (batch_references_t)iris_batch_references;
- perf_cfg->vtbl.bo_wait_rendering =
- (bo_wait_rendering_t)iris_bo_wait_rendering;
- perf_cfg->vtbl.bo_busy = (bo_busy_t)iris_bo_busy;
+ iris_perf_init_vtbl(perf_cfg);
gen_perf_init_metrics(perf_cfg, &screen->devinfo, screen->fd);
screen->monitor_cfg = monitor_cfg;
for (int existing_counter = 0;
existing_counter < perf_cfg->queries[existing_group].n_counters && !duplicate;
++existing_counter) {
- const char *current_name = perf_cfg->queries[group].counters[counter].name;
+ const char *current_name =
+ perf_cfg->queries[group].counters[counter].name;
const char *existing_name =
perf_cfg->queries[existing_group].counters[existing_counter].name;
if (strcmp(current_name, existing_name) == 0) {
return false;
}
-int iris_get_monitor_group_info(struct pipe_screen *pscreen,
- unsigned group_index,
- struct pipe_driver_query_group_info *info)
+int
+iris_get_monitor_group_info(struct pipe_screen *pscreen,
+ unsigned group_index,
+ struct pipe_driver_query_group_info *info)
{
struct iris_screen *screen = (struct iris_screen *)pscreen;
if (!screen->monitor_cfg) {
const struct iris_monitor_config *monitor_cfg = screen->monitor_cfg;
const struct gen_perf_config *perf_cfg = monitor_cfg->perf_cfg;
- if (!info)
+
+ if (!info) {
/* return the count that can be queried */
return perf_cfg->n_queries;
+ }
- if (group_index >= perf_cfg->n_queries)
+ if (group_index >= perf_cfg->n_queries) {
/* out of range */
return 0;
+ }
struct gen_perf_query_info *query = &perf_cfg->queries[group_index];
+
info->name = query->name;
info->max_active_queries = query->n_counters;
info->num_queries = query->n_counters;
+
return 1;
}
+
+static void
+iris_init_monitor_ctx(struct iris_context *ice)
+{
+ struct iris_screen *screen = (struct iris_screen *) ice->ctx.screen;
+ struct iris_monitor_config *monitor_cfg = screen->monitor_cfg;
+
+ ice->perf_ctx = gen_perf_new_context(ice);
+ if (unlikely(!ice->perf_ctx))
+ return;
+
+ struct gen_perf_context *perf_ctx = ice->perf_ctx;
+ struct gen_perf_config *perf_cfg = monitor_cfg->perf_cfg;
+ gen_perf_init_context(perf_ctx,
+ perf_cfg,
+ ice,
+ screen->bufmgr,
+ &screen->devinfo,
+ ice->batches[IRIS_BATCH_RENDER].hw_ctx_id,
+ screen->fd);
+}
+
+/* entry point for GenPerfMonitorsAMD */
+struct iris_monitor_object *
+iris_create_monitor_object(struct iris_context *ice,
+ unsigned num_queries,
+ unsigned *query_types)
+{
+ struct iris_screen *screen = (struct iris_screen *) ice->ctx.screen;
+ struct iris_monitor_config *monitor_cfg = screen->monitor_cfg;
+ struct gen_perf_config *perf_cfg = monitor_cfg->perf_cfg;
+ struct gen_perf_query_object *query_obj = NULL;
+
+ /* initialize perf context if this has not already been done. This
+ * function is the first entry point that carries the gl context.
+ */
+ if (ice->perf_ctx == NULL) {
+ iris_init_monitor_ctx(ice);
+ }
+ struct gen_perf_context *perf_ctx = ice->perf_ctx;
+
+ assert(num_queries > 0);
+ int query_index = query_types[0] - PIPE_QUERY_DRIVER_SPECIFIC;
+ assert(query_index <= monitor_cfg->num_counters);
+ const int group = monitor_cfg->counters[query_index].group;
+
+ struct iris_monitor_object *monitor =
+ calloc(1, sizeof(struct iris_monitor_object));
+ if (unlikely(!monitor))
+ goto allocation_failure;
+
+ monitor->num_active_counters = num_queries;
+ monitor->active_counters = calloc(num_queries, sizeof(int));
+ if (unlikely(!monitor->active_counters))
+ goto allocation_failure;
+
+ for (int i = 0; i < num_queries; ++i) {
+ unsigned current_query = query_types[i];
+ unsigned current_query_index = current_query - PIPE_QUERY_DRIVER_SPECIFIC;
+
+ /* all queries must be in the same group */
+ assert(current_query_index <= monitor_cfg->num_counters);
+ assert(monitor_cfg->counters[current_query_index].group == group);
+ monitor->active_counters[i] =
+ monitor_cfg->counters[current_query_index].counter;
+ }
+
+ /* create the gen_perf_query */
+ query_obj = gen_perf_new_query(perf_ctx, group);
+ if (unlikely(!query_obj))
+ goto allocation_failure;
+
+ monitor->query = query_obj;
+ monitor->result_size = perf_cfg->queries[group].data_size;
+ monitor->result_buffer = calloc(1, monitor->result_size);
+ if (unlikely(!monitor->result_buffer))
+ goto allocation_failure;
+
+ return monitor;
+
+allocation_failure:
+ if (monitor) {
+ free(monitor->active_counters);
+ free(monitor->result_buffer);
+ }
+ free(query_obj);
+ free(monitor);
+ return NULL;
+}
+
+void
+iris_destroy_monitor_object(struct pipe_context *ctx,
+ struct iris_monitor_object *monitor)
+{
+ struct iris_context *ice = (struct iris_context *)ctx;
+
+ gen_perf_delete_query(ice->perf_ctx, monitor->query);
+ free(monitor->result_buffer);
+ monitor->result_buffer = NULL;
+ free(monitor->active_counters);
+ monitor->active_counters = NULL;
+ free(monitor);
+}
+
+bool
+iris_begin_monitor(struct pipe_context *ctx,
+ struct iris_monitor_object *monitor)
+{
+ struct iris_context *ice = (void *) ctx;
+ struct gen_perf_context *perf_ctx = ice->perf_ctx;
+
+ return gen_perf_begin_query(perf_ctx, monitor->query);
+}
+
+bool
+iris_end_monitor(struct pipe_context *ctx,
+ struct iris_monitor_object *monitor)
+{
+ struct iris_context *ice = (void *) ctx;
+ struct gen_perf_context *perf_ctx = ice->perf_ctx;
+
+ gen_perf_end_query(perf_ctx, monitor->query);
+ return true;
+}
+
+bool
+iris_get_monitor_result(struct pipe_context *ctx,
+ struct iris_monitor_object *monitor,
+ bool wait,
+ union pipe_numeric_type_union *result)
+{
+ struct iris_context *ice = (void *) ctx;
+ struct gen_perf_context *perf_ctx = ice->perf_ctx;
+ struct iris_batch *batch = &ice->batches[IRIS_BATCH_RENDER];
+
+ bool monitor_ready =
+ gen_perf_is_query_ready(perf_ctx, monitor->query, batch);
+
+ if (!monitor_ready) {
+ if (!wait)
+ return false;
+ gen_perf_wait_query(perf_ctx, monitor->query, batch);
+ }
+
+ assert(gen_perf_is_query_ready(perf_ctx, monitor->query, batch));
+
+ unsigned bytes_written;
+ gen_perf_get_query_data(perf_ctx, monitor->query,
+ monitor->result_size,
+ (unsigned*) monitor->result_buffer,
+ &bytes_written);
+ if (bytes_written != monitor->result_size)
+ return false;
+
+ /* copy metrics into the batch result */
+ for (int i = 0; i < monitor->num_active_counters; ++i) {
+ int current_counter = monitor->active_counters[i];
+ const struct gen_perf_query_info *info =
+ gen_perf_query_info(monitor->query);
+ const struct gen_perf_query_counter *counter =
+ &info->counters[current_counter];
+ assert(gen_perf_query_counter_get_size(counter));
+ switch (counter->data_type) {
+ case GEN_PERF_COUNTER_DATA_TYPE_UINT64:
+ result[i].u64 = *(uint64_t*)(monitor->result_buffer + counter->offset);
+ break;
+ case GEN_PERF_COUNTER_DATA_TYPE_FLOAT:
+ result[i].f = *(float*)(monitor->result_buffer + counter->offset);
+ break;
+ case GEN_PERF_COUNTER_DATA_TYPE_UINT32:
+ case GEN_PERF_COUNTER_DATA_TYPE_BOOL32:
+ result[i].u64 = *(uint32_t*)(monitor->result_buffer + counter->offset);
+ break;
+ case GEN_PERF_COUNTER_DATA_TYPE_DOUBLE: {
+ double v = *(double*)(monitor->result_buffer + counter->offset);
+ result[i].f = v;
+ break;
+ }
+ default:
+ unreachable("unexpected counter data type");
+ }
+ }
+ return true;
+}