OA_READ_STATUS_FINISHED,
};
-static enum OaReadStatus
-read_oa_samples_until(struct brw_context *brw,
- uint32_t start_timestamp,
- uint32_t end_timestamp)
-{
- struct gen_perf_context *perf_ctx = &brw->perf_ctx;
- 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 brw_context *brw,
- struct gen_perf_query_object *obj)
-{
- uint32_t *start;
- uint32_t *last;
- uint32_t *end;
-
- struct gen_perf_config *perf_cfg = brw->perf_ctx.perf;
-
- /* We need the MI_REPORT_PERF_COUNT to land before we can start
- * accumulate. */
- assert(!perf_cfg->vtbl.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 = perf_cfg->vtbl.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.
struct brw_context *brw = brw_context(ctx);
struct brw_perf_query_object *brw_query = brw_perf_query(o);
struct gen_perf_query_object *obj = brw_query->query;
- struct brw_bo *bo = NULL;
- struct gen_perf_config *perf_cfg = brw->perf_ctx.perf;
assert(!o->Ready);
- switch (obj->queryinfo->kind) {
- case GEN_PERF_QUERY_TYPE_OA:
- case GEN_PERF_QUERY_TYPE_RAW:
- bo = obj->oa.bo;
- break;
-
- case GEN_PERF_QUERY_TYPE_PIPELINE:
- bo = obj->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(&brw->batch, bo))
- perf_cfg->vtbl.batchbuffer_flush(brw, __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 (obj->queryinfo->kind == GEN_PERF_QUERY_TYPE_OA ||
- obj->queryinfo->kind == GEN_PERF_QUERY_TYPE_RAW) {
- while (!read_oa_samples_for_query(brw, obj))
- ;
- }
+ gen_perf_wait_query(&brw->perf_ctx, obj, &brw->batch);
}
static bool
struct brw_context *brw = brw_context(ctx);
struct brw_perf_query_object *brw_query = brw_perf_query(o);
struct gen_perf_query_object *obj = brw_query->query;
- struct gen_perf_config *perf_cfg = brw->perf_ctx.perf;
if (o->Ready)
return true;
- switch (obj->queryinfo->kind) {
- case GEN_PERF_QUERY_TYPE_OA:
- case GEN_PERF_QUERY_TYPE_RAW:
- return (obj->oa.results_accumulated ||
- (obj->oa.bo &&
- !perf_cfg->vtbl.batch_references(&brw->batch, obj->oa.bo) &&
- !brw_bo_busy(obj->oa.bo) &&
- read_oa_samples_for_query(brw, obj)));
- case GEN_PERF_QUERY_TYPE_PIPELINE:
- return (obj->pipeline_stats.bo &&
- !perf_cfg->vtbl.batch_references(&brw->batch, obj->pipeline_stats.bo) &&
- !brw_bo_busy(obj->pipeline_stats.bo));
-
- default:
- unreachable("Unknown query type");
- break;
- }
-
- return false;
+ return gen_perf_is_query_ready(&brw->perf_ctx, obj, &brw->batch);
}
static void
struct gl_perf_query_object *o)
{
struct brw_context *brw = brw_context(ctx);
- struct gen_perf_config *perf_cfg = brw->perf_ctx.perf;
struct brw_perf_query_object *brw_query = brw_perf_query(o);
struct gen_perf_query_object *obj = brw_query->query;
struct gen_perf_context *perf_ctx = &brw->perf_ctx;
DBG("Delete(%d)\n", o->Id);
- switch (obj->queryinfo->kind) {
- 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);
- gen_perf_dec_n_users(perf_ctx);
- }
-
- perf_cfg->vtbl.bo_unreference(obj->oa.bo);
- obj->oa.bo = NULL;
- }
-
- obj->oa.results_accumulated = false;
- break;
-
- case GEN_PERF_QUERY_TYPE_PIPELINE:
- if (obj->pipeline_stats.bo) {
- perf_cfg->vtbl.bo_unreference(obj->pipeline_stats.bo);
- obj->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, obj->queryinfo);
- }
-
- free(obj);
+ gen_perf_delete_query(perf_ctx, obj);
free(brw_query);
}
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 unsigned
brw_init_perf_query_info(struct gl_context *ctx)
(store_register_mem64_t) brw_store_register_mem64;
perf_cfg->vtbl.batch_references = (batch_references_t)brw_batch_references;
perf_cfg->vtbl.bo_wait_rendering = (bo_wait_rendering_t)brw_bo_wait_rendering;
+ perf_cfg->vtbl.bo_busy = (bo_busy_t)brw_bo_busy;
gen_perf_init_context(perf_ctx, perf_cfg, brw, brw->bufmgr, devinfo,
brw->hw_ctx, brw->screen->driScrnPriv->fd);
projects / mesa.git / blobdiff