On 09/05/18 18:48, Lionel Landwerlin wrote:
@@ -1953,10 +1992,26 @@ static int gen8_emit_bb_start(struct i915_request *rq,
rq->ctx->ppgtt->pd_dirty_rings &= ~intel_engine_flag(rq->engine);
}
- cs = intel_ring_begin(rq, 6);
+ cs = intel_ring_begin(rq, rq->engine->id == RCS ? 10 : 6);
if (IS_ERR(cs))
return PTR_ERR(cs);
+ if (rq->engine->id == RCS) {
+ /*
+ * Leave some instructions to be written with an
+ * MI_BATCH_BUFFER_START to the i915/perf NOA reprogramming
+ * batchbuffer. We only turn those MI_NOOP into
+ * MI_BATCH_BUFFER_START when we detect a SSEU powergating
+ * configuration change that might affect NOA. This is only
+ * for the RCS.
+ */
+ rq->perf_prog = intel_ring_offset(rq, cs);
+ *cs++ = MI_NOOP;
+ *cs++ = MI_NOOP;
+ *cs++ = MI_NOOP;
+ *cs++ = MI_NOOP; /* Aligning to 2 dwords */
+ }
+
I just realized that isn't going to work if a request is preempted, then
later resubmitted after another context with a different powergating
config...
This reprog bb won't be executed because the CS pointer should be past
that point already.
It seems to make this approach unworkable?
Would a per-ctx-wa-bb call into the reprogramming buffer under
MI_PREDICATE work?
LOAD rpcs into predicate_reg0
LOAD engine storage for last rpcs into prediate_reg1
PREDICATE reg0 == reg1
MI_LRI noa registers
PREDICATE unset
STORE rpcs into engine storage for last rpcs
Thanks,
-
Lionel
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
