#15221: Post completion messages to dispatch_s

We never wait on the acks from these completion messages, so make them posted to avoid contention from a lot of replies being sent at once. In the case where every worker is sending them at the same time, this can halve the latency from 500ns to 250ns (on wormhole).
tenstorrent · Dec 19, 2024 · 3f54011 · 3f54011
1 parent 485a18d
commit 3f54011
Show file tree

Hide file tree

Showing 3 changed files with 21 additions and 4 deletions.
diff --git a/tt_metal/hw/firmware/src/brisc.cc b/tt_metal/hw/firmware/src/brisc.cc
@@ -406,7 +406,8 @@ int main() {
                     NOC_UNICAST_WRITE_VC,
                     1,
                     31 /*wrap*/,
-                    false /*linked*/);
+                    false /*linked*/,
+                    true /*posted*/);
             }
         }
 
@@ -529,7 +530,8 @@ int main() {
                     NOC_UNICAST_WRITE_VC,
                     1,
                     31 /*wrap*/,
-                    false /*linked*/);
+                    false /*linked*/,
+                    true /*posted*/);
                 mailboxes->launch_msg_rd_ptr = (launch_msg_rd_ptr + 1) & (launch_msg_buffer_num_entries - 1);
             }
         }

diff --git a/tt_metal/hw/firmware/src/idle_erisc.cc b/tt_metal/hw/firmware/src/idle_erisc.cc
@@ -176,7 +176,15 @@ int main() {
                     DISPATCH_MESSAGE_ADDR + mailboxes->go_message.dispatch_message_offset);
                 DEBUG_SANITIZE_NOC_ADDR(noc_index, dispatch_addr, 4);
                 CLEAR_PREVIOUS_LAUNCH_MESSAGE_ENTRY_FOR_WATCHER();
-                noc_fast_atomic_increment(noc_index, NCRISC_AT_CMD_BUF, dispatch_addr, NOC_UNICAST_WRITE_VC, 1, 31 /*wrap*/, false /*linked*/);
+                noc_fast_atomic_increment(
+                    noc_index,
+                    NCRISC_AT_CMD_BUF,
+                    dispatch_addr,
+                    NOC_UNICAST_WRITE_VC,
+                    1,
+                    31 /*wrap*/,
+                    false /*linked*/,
+                    true /*posted*/);
                 mailboxes->launch_msg_rd_ptr = (launch_msg_rd_ptr + 1) & (launch_msg_buffer_num_entries - 1);
             }
 

diff --git a/tt_metal/hw/inc/ethernet/tunneling.h b/tt_metal/hw/inc/ethernet/tunneling.h
@@ -109,7 +109,14 @@ void notify_dispatch_core_done(uint64_t dispatch_addr) {
     }
     DEBUG_SANITIZE_NOC_ADDR(noc_index, dispatch_addr, 4);
     noc_fast_atomic_increment(
-        noc_index, NCRISC_AT_CMD_BUF, dispatch_addr, NOC_UNICAST_WRITE_VC, 1, 31 /*wrap*/, false /*linked*/);
+        noc_index,
+        NCRISC_AT_CMD_BUF,
+        dispatch_addr,
+        NOC_UNICAST_WRITE_VC,
+        1,
+        31 /*wrap*/,
+        false /*linked*/,
+        true /*posted*/);
 }
 
 }  // namespace internal_