freedreno/ir3: try to avoid syncs

Update postsched to be better aware of where costly (ss) syncs would
result.  Sometimes it is better to allow a nop or two, to avoid a
sync quickly after an SFU.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4071>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4071>

diff --git a/src/freedreno/ir3/ir3_postsched.c b/src/freedreno/ir3/ir3_postsched.c
index 1e706f75a24844ff34325f0884221a8ee830fd1c..b283af486fb58203c4c9070e511d64a7bf181f44 100644 (file)
--- a/src/freedreno/ir3/ir3_postsched.c
+++ b/src/freedreno/ir3/ir3_postsched.c
@@ -62,6 +62,8 @@ struct ir3_postsched_ctx {
        struct ir3_instruction *pred;      /* current p0.x user, if any */
 
        bool error;
+
+       int sfu_delay;
 };
 
 struct ir3_postsched_node {
@@ -97,6 +99,12 @@ schedule(struct ir3_postsched_ctx *ctx, struct ir3_instruction *instr)
        list_addtail(&instr->node, &instr->block->instr_list);
        ctx->scheduled = instr;
 
+       if (is_sfu(instr)) {
+               ctx->sfu_delay = 8;
+       } else if (ctx->sfu_delay > 0) {
+               ctx->sfu_delay--;
+       }
+
        struct ir3_postsched_node *n = instr->data;
        dag_prune_head(ctx->dag, &n->dag);
 }
@@ -119,6 +127,24 @@ dump_state(struct ir3_postsched_ctx *ctx)
        }
 }
 
+/* Determine if this is an instruction that we'd prefer not to schedule
+ * yet, in order to avoid an (ss) sync.  This is limited by the sfu_delay
+ * counter, ie. the more cycles it has been since the last SFU, the less
+ * costly a sync would be.
+ */
+static bool
+would_sync(struct ir3_postsched_ctx *ctx, struct ir3_instruction *instr)
+{
+       if (ctx->sfu_delay) {
+               struct ir3_register *reg;
+               foreach_src (reg, instr)
+                       if (reg->instr && is_sfu(reg->instr))
+                               return true;
+       }
+
+       return false;
+}
+
 /* find instruction to schedule: */
 static struct ir3_instruction *
 choose_instr(struct ir3_postsched_ctx *ctx)
@@ -195,7 +221,35 @@ choose_instr(struct ir3_postsched_ctx *ctx)
                return chosen->instr;
        }
 
-       /* First try to find a ready leader w/ soft delay (ie. including extra
+       /*
+        * Sometimes be better to take a nop, rather than scheduling an
+        * instruction that would require an (ss) shortly after another
+        * SFU..  ie. if last SFU was just one or two instr ago, and we
+        * could choose between taking a nop and then scheduling
+        * something else, vs scheduling the immed avail instruction that
+        * would require (ss), we are better with the nop.
+        */
+       for (unsigned delay = 0; delay < 4; delay++) {
+               foreach_sched_node (n, &ctx->dag->heads) {
+                       if (would_sync(ctx, n->instr))
+                               continue;
+
+                       unsigned d = ir3_delay_calc(ctx->block, n->instr, true, false);
+
+                       if (d > delay)
+                               continue;
+
+                       if (!chosen || (chosen->max_delay < n->max_delay))
+                               chosen = n;
+               }
+
+               if (chosen) {
+                       di(chosen->instr, "csp: chose (soft ready, delay=%u)", delay);
+                       return chosen->instr;
+               }
+       }
+
+       /* Next try to find a ready leader w/ soft delay (ie. including extra
         * delay for things like tex fetch which can be synchronized w/ sync
         * bit (but we probably do want to schedule some other instructions
         * while we wait)