Perf optimization. I-cache aware?

tenstorrent · Feb 18, 2025 · a163057 · a163057
1 parent fefe478
commit a163057
Show file tree

Hide file tree

Showing 2 changed files with 4 additions and 5 deletions.
diff --git a/tt_fabric/hw/inc/tt_fabric.h b/tt_fabric/hw/inc/tt_fabric.h
@@ -494,7 +494,6 @@ typedef struct fvc_producer_state {
             next_routing_ptr =
                 (tt_routing*)(next_header_ptr + packet_header_routing_offset_dwords - dwords_before_wrap);
         } else {
-#pragma GCC unroll 12
             for (uint32_t i = 0; i < dwords_to_copy; i++) {
                 packet_header_ptr[i] = next_header_ptr[i];
             }

diff --git a/tt_fabric/hw/inc/tt_fabric_interface.h b/tt_fabric/hw/inc/tt_fabric_interface.h
@@ -140,8 +140,8 @@ constexpr uint32_t packet_header_routing_offset_dwords = offsetof(packet_header_
 void tt_fabric_add_header_checksum(packet_header_t* p_header) {
     uint16_t* ptr = (uint16_t*)p_header;
     uint32_t sum = 0;
-    for (uint32_t i = 2; i < sizeof(packet_header_t) / 2; i++) {
-        sum += ptr[i];
+    for (uint32_t i = 2; i + 1 < sizeof(packet_header_t) / 2; i += 2) {
+        sum += ptr[i] + ptr[i + 1];
     }
     sum = ~sum;
     sum += sum;
@@ -152,8 +152,8 @@ bool tt_fabric_is_header_valid(packet_header_t* p_header) {
 #ifdef TT_FABRIC_DEBUG
     uint16_t* ptr = (uint16_t*)p_header;
     uint32_t sum = 0;
-    for (uint32_t i = 2; i < sizeof(packet_header_t) / 2; i++) {
-        sum += ptr[i];
+    for (uint32_t i = 2; i + 1 < sizeof(packet_header_t) / 2; i += 2) {
+        sum += ptr[i] + ptr[i + 1];
     }
     sum = ~sum;
     sum += sum;