Calculate steps needed by tick of last output event

1 files changed, 14 insertions, 5 deletions

diff --git a/tpu.c b/tpu.c
index aac59ed..13db28d 100644
--- a/tpu.c
+++ b/tpu.c
@@ -181,6 +181,7 @@ tpu_io_port_init(struct tpu_io_port *io_port, char c, enum tpu_port_dir dir,
 	enum tpu_port_type type, int x, int y)
 {
 	tpu_port_init(&io_port->port, NULL);
+	io_port->io_step = 0;
 	io_port->type = type;
 	io_port->file = NULL;
 	io_port->dir = dir;
@@ -673,7 +674,6 @@ tis_step(struct tis *tis)
 
 	tis_communicate(tis);
 
-	running = false;
 	for (i = 0; i < TPU_MAP_BUCKETS; i++) {
 		link = tis->tpu_map.buckets[i];
 		for (; link; link = link->next)
@@ -682,15 +682,15 @@ tis_step(struct tis *tis)
 
 	for (i = 0; i < TPU_MAP_BUCKETS; i++) {
 		link = tis->tpu_map.buckets[i];
-		for (; link; link = link->next) {
+		for (; link; link = link->next)
 			tpu_update(link->tpu);
-			running |= !link->tpu->idle;
-		}
 	}
 
+	running = false;
 	for (i = 0; i < TPU_MAP_BUCKETS; i++) {
 		link = tis->tpu_map.buckets[i];
 		for (; link; link = link->next) {
+			running |= !link->tpu->idle;
 			if (link->tpu->pc < link->tpu->inst_cnt) {
 				link->tpu->steps += 1;
 				if (link->tpu->idle)
@@ -732,6 +732,7 @@ tis_communicate(struct tis *tis)
 				die("communicate: invalid input '%s'", buf);
 			io_port->port.out = val;
 			io_port->port.writing = true;
+			io_port->io_step = tis->steps;
 			break;
 		}
 	}
@@ -745,6 +746,7 @@ tis_communicate(struct tis *tis)
 			fprintf(io_port->file, "%i\n", io_port->port.in);
 		io_port->port.reset_in = true;
 		tpu_port_update(&io_port->port);
+		io_port->io_step = tis->steps;
 	}
 }
 
@@ -757,11 +759,18 @@ tis_gen_stats(struct tis *tis)
 	size_t all_steps;
 	int i;
 
-	stats.steps = tis->steps;
+	stats.steps = 0;
 	stats.nodes = 0;
 	stats.insts = 0;
 	stats.idle = 0;
 
+	for (i = 0; i < TIS_MAX_IO_PORTS; i++) {
+		if (tis->out_ports[i] && tis->out_ports[i]->io_step > 0) {
+			stats.steps = MAX(stats.steps, tis->out_ports[i]->io_step + 1);
+		}
+	}
+	if (!stats.steps) stats.steps = tis->steps;
+
 	all_steps = 0;
 	idle_steps = 0;
 	for (i = 0; i < TPU_MAP_BUCKETS; i++) {