diff options
| -rw-r--r-- | SEL/host/app.c | 26 |
1 files changed, 24 insertions, 2 deletions
diff --git a/SEL/host/app.c b/SEL/host/app.c index a2ec8b1..611732a 100644 --- a/SEL/host/app.c +++ b/SEL/host/app.c @@ -29,6 +29,9 @@ #include <dpu_probe.h> #endif +#include <dpu_management.h> +#include <dpu_target_macros.h> + // Pointer declaration static T* A; static T* C; @@ -71,6 +74,8 @@ int main(int argc, char **argv) { // Timer declaration Timer timer; + int numa_node_rank = -2; + // Allocate DPUs and load binary #if !WITH_ALLOC_OVERHEAD DPU_ASSERT(dpu_alloc(NR_DPUS, NULL, &dpu_set)); @@ -143,6 +148,23 @@ int main(int argc, char **argv) { assert(nr_of_dpus == NR_DPUS); #endif + // int prev_rank_id = -1; + int rank_id = -1; + DPU_FOREACH (dpu_set, dpu) { + rank_id = dpu_get_rank_id(dpu_get_rank(dpu_from_set(dpu))) & DPU_TARGET_MASK; + if ((numa_node_rank != -2) && numa_node_rank != dpu_get_rank_numa_node(dpu_get_rank(dpu_from_set(dpu)))) { + numa_node_rank = -1; + } else { + numa_node_rank = dpu_get_rank_numa_node(dpu_get_rank(dpu_from_set(dpu))); + } + /* + if (rank_id != prev_rank_id) { + printf("/dev/dpu_rank%d @ NUMA node %d\n", rank_id, numa_node_rank); + prev_rank_id = rank_id; + } + */ + } + // Compute output on CPU (performance comparison and verification purposes) if(rep >= p.n_warmup) start(&timer, 2, 0); @@ -297,8 +319,8 @@ int main(int argc, char **argv) { if (rep >= p.n_warmup) { printf("[::] SEL UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d e_type=%s block_size_B=%d n_elements=%d n_elements_per_dpu=%d", NR_DPUS, nr_of_ranks, NR_TASKLETS, XSTR(T), BLOCK_SIZE, input_size, input_size_dpu_round); - printf(" b_with_alloc_overhead=%d b_with_load_overhead=%d b_with_free_overhead=%d ", - WITH_ALLOC_OVERHEAD, WITH_LOAD_OVERHEAD, WITH_FREE_OVERHEAD); + printf(" b_with_alloc_overhead=%d b_with_load_overhead=%d b_with_free_overhead=%d numa_node_rank=%d ", + WITH_ALLOC_OVERHEAD, WITH_LOAD_OVERHEAD, WITH_FREE_OVERHEAD, numa_node_rank); printf("| latency_alloc_us=%f latency_load_us=%f latency_cpu_us=%f latency_write_us=%f latency_kernel_us=%f latency_read_us=%f latency_free_us=%f", timer.time[0], timer.time[1], |