drm/i915/guc: use for_each_engine_id() where appropriate
authorDave Gordon <david.s.gordon@intel.com>
Tue, 9 Aug 2016 14:19:22 +0000 (15:19 +0100)
committerTvrtko Ursulin <tvrtko.ursulin@intel.com>
Wed, 10 Aug 2016 09:40:05 +0000 (10:40 +0100)
Now that host structures are indexed by host engine-id rather than
guc_id, we can usefully convert some for_each_engine() loops to use
for_each_engine_id() and avoid multiple dereferences of engine->id.

Also a few related tweaks to cache structure members locally wherever
they're used more than once or twice, hopefully eliminating memory
references.

Signed-off-by: Dave Gordon <david.s.gordon@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
drivers/gpu/drm/i915/i915_debugfs.c
drivers/gpu/drm/i915/i915_guc_submission.c

index 0627e170ea2599d472448ad7118ad74e02122bb2..b84fa93124ce3bb381bd12b74813c66bf2c3bbcc 100644 (file)
@@ -2544,6 +2544,7 @@ static void i915_guc_client_info(struct seq_file *m,
                                 struct i915_guc_client *client)
 {
        struct intel_engine_cs *engine;
+       enum intel_engine_id id;
        uint64_t tot = 0;
 
        seq_printf(m, "\tPriority %d, GuC ctx index: %u, PD offset 0x%x\n",
@@ -2558,11 +2559,11 @@ static void i915_guc_client_info(struct seq_file *m,
        seq_printf(m, "\tFailed doorbell: %u\n", client->b_fail);
        seq_printf(m, "\tLast submission result: %d\n", client->retcode);
 
-       for_each_engine(engine, dev_priv) {
+       for_each_engine_id(engine, dev_priv, id) {
+               u64 submissions = client->submissions[id];
+               tot += submissions;
                seq_printf(m, "\tSubmissions: %llu %s\n",
-                               client->submissions[engine->id],
-                               engine->name);
-               tot += client->submissions[engine->id];
+                               submissions, engine->name);
        }
        seq_printf(m, "\tTotal: %llu\n", tot);
 }
@@ -2575,6 +2576,7 @@ static int i915_guc_info(struct seq_file *m, void *data)
        struct intel_guc guc;
        struct i915_guc_client client = {};
        struct intel_engine_cs *engine;
+       enum intel_engine_id id;
        u64 total = 0;
 
        if (!HAS_GUC_SCHED(dev_priv))
@@ -2601,11 +2603,11 @@ static int i915_guc_info(struct seq_file *m, void *data)
        seq_printf(m, "GuC last action error code: %d\n", guc.action_err);
 
        seq_printf(m, "\nGuC submissions:\n");
-       for_each_engine(engine, dev_priv) {
+       for_each_engine_id(engine, dev_priv, id) {
+               u64 submissions = guc.submissions[id];
+               total += submissions;
                seq_printf(m, "\t%-24s: %10llu, last seqno 0x%08x\n",
-                       engine->name, guc.submissions[engine->id],
-                       guc.last_seqno[engine->id]);
-               total += guc.submissions[engine->id];
+                       engine->name, submissions, guc.last_seqno[id]);
        }
        seq_printf(m, "\t%s: %llu\n", "Total", total);
 
index 405e445fba729768a04abb0bab281b9bdb8aceab..6831321a9c8c478eb7a65d2875c8cdf38426b3f4 100644 (file)
@@ -342,7 +342,8 @@ static void guc_init_ctx_desc(struct intel_guc *guc,
 
        for_each_engine_masked(engine, dev_priv, client->engines) {
                struct intel_context *ce = &ctx->engine[engine->id];
-               struct guc_execlist_context *lrc = &desc.lrc[engine->guc_id];
+               uint32_t guc_engine_id = engine->guc_id;
+               struct guc_execlist_context *lrc = &desc.lrc[guc_engine_id];
                struct drm_i915_gem_object *obj;
 
                /* TODO: We have a design issue to be solved here. Only when we
@@ -361,7 +362,7 @@ static void guc_init_ctx_desc(struct intel_guc *guc,
                gfx_addr = i915_gem_obj_ggtt_offset(ce->state);
                lrc->ring_lcra = gfx_addr + LRC_STATE_PN * PAGE_SIZE;
                lrc->context_id = (client->ctx_index << GUC_ELC_CTXID_OFFSET) |
-                               (engine->guc_id << GUC_ELC_ENGINE_OFFSET);
+                               (guc_engine_id << GUC_ELC_ENGINE_OFFSET);
 
                obj = ce->ring->obj;
                gfx_addr = i915_gem_obj_ggtt_offset(obj);
@@ -371,7 +372,7 @@ static void guc_init_ctx_desc(struct intel_guc *guc,
                lrc->ring_next_free_location = gfx_addr;
                lrc->ring_current_tail_pointer_value = 0;
 
-               desc.engines_used |= (1 << engine->guc_id);
+               desc.engines_used |= (1 << guc_engine_id);
        }
 
        DRM_DEBUG_DRIVER("Host engines 0x%x => GuC engines used 0x%x\n",
@@ -459,6 +460,7 @@ static void guc_add_workqueue_item(struct i915_guc_client *gc,
        /* wqi_len is in DWords, and does not include the one-word header */
        const size_t wqi_size = sizeof(struct guc_wq_item);
        const u32 wqi_len = wqi_size/sizeof(u32) - 1;
+       struct intel_engine_cs *engine = rq->engine;
        struct guc_process_desc *desc;
        struct guc_wq_item *wqi;
        void *base;
@@ -500,12 +502,11 @@ static void guc_add_workqueue_item(struct i915_guc_client *gc,
        /* Now fill in the 4-word work queue item */
        wqi->header = WQ_TYPE_INORDER |
                        (wqi_len << WQ_LEN_SHIFT) |
-                       (rq->engine->guc_id << WQ_TARGET_SHIFT) |
+                       (engine->guc_id << WQ_TARGET_SHIFT) |
                        WQ_NO_WCFLUSH_WAIT;
 
        /* The GuC wants only the low-order word of the context descriptor */
-       wqi->context_desc = (u32)intel_lr_context_descriptor(rq->ctx,
-                                                            rq->engine);
+       wqi->context_desc = (u32)intel_lr_context_descriptor(rq->ctx, engine);
 
        wqi->ring_tail = tail << WQ_RING_TAIL_SHIFT;
        wqi->fence_id = rq->fence.seqno;