From 4cb1ebd9b286660257ad235153e6e6540e3cc4ea Mon Sep 17 00:00:00 2001 From: jasha64 Date: Mon, 15 Jun 2026 15:50:19 +0800 Subject: [PATCH] perftest: fix reducescatter latency size reporting Avoid shadowing num_elems in for loop so the timed kernel launches use the current message size instead of repeatedly measuring the initial element count. Signed-off-by: jasha64 --- perftest/device/coll/reducescatter_latency.cu | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perftest/device/coll/reducescatter_latency.cu b/perftest/device/coll/reducescatter_latency.cu index c0928088..6a1845be 100644 --- a/perftest/device/coll/reducescatter_latency.cu +++ b/perftest/device/coll/reducescatter_latency.cu @@ -107,7 +107,7 @@ CALL_RDXN_OPS_ALL_TG(int64, int64_t) \ nvshmem_barrier_all(); \ j = 0; \ - for (size_t num_elems = min_elems; num_elems < ELEM_COMP; num_elems *= step_factor) { \ + for (num_elems = min_elems; num_elems < ELEM_COMP; num_elems *= step_factor) { \ CALL_RDXN_KERNEL(TYPENAME, OP, GROUP, num_blocks, nvshm_test_num_tpb, skip_arg_list, \ stream) \ CUDA_CHECK(cudaStreamSynchronize(stream)); \