better

tenstorrent · Dec 17, 2024 · 02a91ff · 02a91ff
1 parent dc96a4f
commit 02a91ff
Showing 1 changed file with 4 additions and 3 deletions.
diff --git a/tt_metal/impl/dispatch/kernels/cq_dispatch.cpp b/tt_metal/impl/dispatch/kernels/cq_dispatch.cpp
@@ -574,6 +574,9 @@ void process_write_packed(
     uint32_t writes = 0;
     uint32_t mcasts = 0;
     auto wait_for_barrier = [&]() {
+        if (!mcast) {
+            return;
+        }
         noc_nonposted_writes_num_issued[noc_index] += writes;
         noc_nonposted_writes_acked[noc_index] += mcasts;
         writes = 0;
@@ -757,7 +760,6 @@ void process_write_packed_large(
                 wait_for_barrier();
                 cq_noc_async_write_with_state_any_len(data_ptr, dst_addr, xfer_size, num_dests);
                 must_barrier = false;
-                writes += div_up(xfer_size, NOC_MAX_BURST_SIZE);
             } else {
                 xfer_size = length;
                 if (unlink) {
@@ -769,16 +771,15 @@ void process_write_packed_large(
                     uint32_t data_offset = xfer_size - rem_xfer_size;
                     cq_noc_async_write_with_state<CQ_NOC_SnDL, CQ_NOC_wait>(
                         data_ptr + data_offset, dst_addr + data_offset, rem_xfer_size, num_dests);
-                    writes += div_up(xfer_size, NOC_MAX_BURST_SIZE);
                     must_barrier = true;
                     // Later writes must barrier.
                 } else {
                     wait_for_barrier();
                     cq_noc_async_write_with_state_any_len(data_ptr, dst_addr, xfer_size, num_dests);
                     must_barrier = false;
-                    writes++;
                 }
             }
+            writes += div_up(xfer_size, NOC_MAX_BURST_SIZE);
             length -= xfer_size;
             data_ptr += xfer_size;
             dst_addr += xfer_size;