[i-g-t,04/15] gem_wsim: Engine map load balance command

Commit Message

Tvrtko Ursulin May 22, 2019, 3:57 p.m. UTC

From: Tvrtko Ursulin <[e-mail]>

A new workload command for enabling a load balanced context map (aka
Virtual Engine). Example usage:

  B.1

This turns on load balancing for context one, assuming it has already been
configured with an engine map. Only DEFAULT engine specifier can be used
with load balanced engine maps.

v2:
 * Lift restriction to only use load balancer when enabled in context map.
   (Chris)

Signed-off-by: Tvrtko Ursulin <[e-mail]>
---
 benchmarks/gem_wsim.c  | 66 +++++++++++++++++++++++++++++++++++++-----
 benchmarks/wsim/README | 15 ++++++++++
 2 files changed, 74 insertions(+), 7 deletions(-)

Comments

Chris Wilson May 23, 2019, 1:25 p.m. UTC | #1

Quoting Tvrtko Ursulin (2019-05-22 16:57:09)
> From: Tvrtko Ursulin <[e-mail]>
> 
> A new workload command for enabling a load balanced context map (aka
> Virtual Engine). Example usage:
> 
>   B.1
> 
> This turns on load balancing for context one, assuming it has already been
> configured with an engine map. Only DEFAULT engine specifier can be used
> with load balanced engine maps.
> 
> v2:
>  * Lift restriction to only use load balancer when enabled in context map.
>    (Chris)

You didn't fancy going all out and say:
B.1.DEFAULT.VCS1|VCS2
?

If you are happy with the current code, it looks to do what you want,
Reviewed-by: Chris Wilson <[e-mail]>
-Chris

Tvrtko Ursulin May 23, 2019, 1:58 p.m. UTC | #2

On 23/05/2019 14:25, Chris Wilson wrote:
> Quoting Tvrtko Ursulin (2019-05-22 16:57:09)
>> From: Tvrtko Ursulin <[e-mail]>
>>
>> A new workload command for enabling a load balanced context map (aka
>> Virtual Engine). Example usage:
>>
>>    B.1
>>
>> This turns on load balancing for context one, assuming it has already been
>> configured with an engine map. Only DEFAULT engine specifier can be used
>> with load balanced engine maps.
>>
>> v2:
>>   * Lift restriction to only use load balancer when enabled in context map.
>>     (Chris)
> 
> You didn't fancy going all out and say:
> B.1.DEFAULT.VCS1|VCS2
> ?
> 
> If you are happy with the current code, it looks to do what you want,
> Reviewed-by: Chris Wilson <[e-mail]>

Thanks, I wouldn't say I am happy as in really proud of it happy. But as 
a tool by hackers for hackers, which grew organically always as a second 
priority thing, it seems to work for now and is able to exercise the new 
uAPI and scheduling paths.

So there is scope to tidy, and will certainly need more work in the 
future (not least per engine calibration), but I need to call it done 
for a while at some reasonable point and it feels like that should be now.

Proof in the pudding is that I think you found it useful when 
benchmarking the semaphore code and related issues. So it at least 
continues providing the same simulated workloads over the new uAPI.

Regards,

Tvrtko

Patch

diff --git a/benchmarks/gem_wsim.c b/benchmarks/gem_wsim.c
index 66832f74e34a..d645611cf8c2 100644
--- a/benchmarks/gem_wsim.c
+++ b/benchmarks/gem_wsim.c
@@ -83,7 +83,8 @@  enum w_type
 	SW_FENCE_SIGNAL,
 	CTX_PRIORITY,
 	PREEMPTION,
-	ENGINE_MAP
+	ENGINE_MAP,
+	LOAD_BALANCE,
 };
 
 struct deps
@@ -121,6 +122,7 @@  struct w_step
 			unsigned int engine_map_count;
 			enum intel_engine_id *engine_map;
 		};
+		bool load_balance;
 	};
 
 	/* Implementation details */
@@ -507,6 +509,25 @@  parse_workload(struct w_arg *arg, unsigned int flags, struct workload *app_w)
 
 				step.type = PREEMPTION;
 				goto add_step;
+			} else if (!strcmp(field, "B")) {
+				unsigned int nr = 0;
+				while ((field = strtok_r(fstart, ".", &fctx))) {
+					tmp = atoi(field);
+					check_arg(nr == 0 && tmp <= 0,
+						  "Invalid context at step %u!\n",
+						  nr_steps);
+					check_arg(nr > 0,
+						  "Invalid load balance format at step %u!\n",
+						  nr_steps);
+
+					step.context = tmp;
+					step.load_balance = true;
+
+					nr++;
+				}
+
+				step.type = LOAD_BALANCE;
+				goto add_step;
 			}
 
 			if (!field) {
@@ -841,7 +862,7 @@  find_engine_in_map(struct ctx *ctx, enum intel_engine_id engine)
 			return i + 1;
 	}
 
-	igt_assert(0);
+	igt_assert(ctx->wants_balance);
 	return 0;
 }
 
@@ -1073,12 +1094,19 @@  prepare_workload(unsigned int id, struct workload *wrk, unsigned int flags)
 				wrk->ctx_list[j].engine_map = w->engine_map;
 				wrk->ctx_list[j].engine_map_count =
 					w->engine_map_count;
+			} else if (w->type == LOAD_BALANCE) {
+				if (!wrk->ctx_list[j].engine_map) {
+					wsim_err("Load balancing needs an engine map!\n");
+					return 1;
+				}
+				wrk->ctx_list[j].wants_balance =
+					w->load_balance;
 			}
 		}
 
 		wrk->ctx_list[j].targets_instance = targets;
 		if (flags & I915)
-			wrk->ctx_list[j].wants_balance = balance;
+			wrk->ctx_list[j].wants_balance |= balance;
 	}
 
 	/*
@@ -1092,7 +1120,9 @@  prepare_workload(unsigned int id, struct workload *wrk, unsigned int flags)
 			if (w->type != BATCH)
 				continue;
 
-			if (wrk->ctx_list[j].engine_map && w->engine == VCS) {
+			if (wrk->ctx_list[j].engine_map &&
+			    !wrk->ctx_list[j].wants_balance &&
+			    (w->engine == VCS || w->engine == DEFAULT)) {
 				wsim_err("Batches targetting engine maps must use explicit engines!\n");
 				return -1;
 			}
@@ -1140,7 +1170,8 @@  prepare_workload(unsigned int id, struct workload *wrk, unsigned int flags)
 				break;
 			}
 
-			if (!ctx->engine_map && !ctx->targets_instance)
+			if ((!ctx->engine_map && !ctx->targets_instance) ||
+			    (ctx->engine_map && ctx->wants_balance))
 				args.flags |=
 				     I915_CONTEXT_CREATE_FLAGS_SINGLE_TIMELINE;
 
@@ -1201,6 +1232,8 @@  prepare_workload(unsigned int id, struct workload *wrk, unsigned int flags)
 		if (ctx->engine_map) {
 			I915_DEFINE_CONTEXT_PARAM_ENGINES(set_engines,
 							  ctx->engine_map_count + 1);
+			I915_DEFINE_CONTEXT_ENGINES_LOAD_BALANCE(load_balance,
+								 ctx->engine_map_count);
 			struct drm_i915_gem_context_param param = {
 				.ctx_id = ctx_id,
 				.param = I915_CONTEXT_PARAM_ENGINES,
@@ -1208,7 +1241,25 @@  prepare_workload(unsigned int id, struct workload *wrk, unsigned int flags)
 				.value = to_user_pointer(&set_engines),
 			};
 
-			set_engines.extensions = 0;
+			if (ctx->wants_balance) {
+				set_engines.extensions =
+					to_user_pointer(&load_balance);
+
+				memset(&load_balance, 0, sizeof(load_balance));
+				load_balance.base.name =
+					I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE;
+				load_balance.num_siblings =
+					ctx->engine_map_count;
+
+				for (j = 0; j < ctx->engine_map_count; j++) {
+					load_balance.engines[j].engine_class =
+						I915_ENGINE_CLASS_VIDEO; /* FIXME */
+					load_balance.engines[j].engine_instance =
+						ctx->engine_map[j] - VCS1; /* FIXME */
+				}
+			} else {
+				set_engines.extensions = 0;
+			}
 
 			/* Reserve slot for virtual engine. */
 			set_engines.engines[0].engine_class =
@@ -2196,7 +2247,8 @@  static void *run_workload(void *data)
 				}
 				continue;
 			} else if (w->type == PREEMPTION ||
-				   w->type == ENGINE_MAP) {
+				   w->type == ENGINE_MAP ||
+				   w->type == LOAD_BALANCE) {
 				continue;
 			}
 
diff --git a/benchmarks/wsim/README b/benchmarks/wsim/README
index 53f814a73c73..2c085921e97b 100644
--- a/benchmarks/wsim/README
+++ b/benchmarks/wsim/README
@@ -3,6 +3,7 @@  Workload descriptor format
 
 ctx.engine.duration_us.dependency.wait,...
 <uint>.<str>.<uint>[-<uint>].<int <= 0>[/<int <= 0>][...].<0|1>,...
+B.<uint>
 M.<uint>.<str>[|<str>]...
 P|X.<uint>.<int>
 d|p|s|t|q|a.<int>,...
@@ -24,6 +25,7 @@  Additional workload steps are also supported:
  'q' - Throttle to n max queue depth.
  'f' - Create a sync fence.
  'a' - Advance the previously created sync fence.
+ 'B' - Turn on context load balancing.
  'M' - Set up engine map.
  'P' - Context priority.
  'X' - Context preemption control.
@@ -184,3 +186,16 @@  Example:
 M.1.VCS
 
 This sets up the engine map to all available VCS class engines.
+
+Context load balancing
+----------------------
+
+Context load balancing (aka Virtual Engine) is an i915 feature where the driver
+will pick the best engine (most idle) to submit to given previously configured
+engine map.
+
+Example:
+
+  B.1
+
+This enables load balancing for context number one.