benchdnn: add ref run report status

oneapi-src · Feb 13, 2025 · fb09328 · fb09328
1 parent 94279b6
commit fb09328
Show file tree

Hide file tree

Showing 3 changed files with 22 additions and 14 deletions.
diff --git a/tests/benchdnn/dnnl_common.hpp b/tests/benchdnn/dnnl_common.hpp
@@ -594,6 +594,12 @@ void check_correctness(const prb_t *prb, const std::vector<data_kind_t> &kinds,
         TIME_COMPARE(check_buffer_overwrite(args.dnn_mem(i), args.arg(i), res));
     }
 
+    // Report prim_ref run status for easier distinguishing between GPU failures
+    // and ref CPU failures.
+    if (prim_ref) {
+        BENCHDNN_PRINT(1, "run ref: %s\n", res->prim_ref_repro.c_str());
+    }
+
     TIME_REF(compute_ref(prb, ref_args, prim_ref));
 
     for (const auto &kind : kinds) {
@@ -621,15 +627,6 @@ void check_correctness(const prb_t *prb, const std::vector<data_kind_t> &kinds,
                 cpu_cache_args.L2_size, cpu_cache_args.L3_size,
                 benchdnn_get_max_threads(),
                 query_impl_info(query_pd(prim_ref)).c_str());
-
-        // Replace engine kind for repro line from GPU to CPU.
-        const auto eng_pos = res->prim_ref_repro.find("engine=gpu");
-        if (eng_pos != std::string::npos)
-            // Replace `g` in `gpu` with `c`
-            res->prim_ref_repro[eng_pos + 7] = 'c';
-
-        BENCHDNN_PRINT(
-                0, "[PRIM_REF][REPRO]: %s\n", res->prim_ref_repro.c_str());
     }
 }
 
@@ -1063,7 +1060,15 @@ int init_prim_ref_common(benchdnn_dnnl_wrapper_t<dnnl_primitive_t> &prim_ref,
 
     BENCHDNN_PRINT(5, "CPU reference oneDNN implementation: %s\n",
             query_impl_info(pdw).c_str());
+
     res->prim_ref_repro = prb_cpu->str();
+    // Replace engine kind for repro line from GPU to CPU.
+    const auto eng_pos = res->prim_ref_repro.find("engine=gpu");
+    if (eng_pos != std::string::npos) {
+        // Replace `g` in `gpu` with `c`
+        res->prim_ref_repro[eng_pos + 7] = 'c';
+    }
+
     prim_ref.reset(prim_ref_ptr);
     return OK;
 }

diff --git a/tests/benchdnn/utils/task.hpp b/tests/benchdnn/utils/task.hpp
@@ -38,8 +38,11 @@ struct task_t {
         , perf_template_(perf_template)
         , idx_(idx) {}
 
-    int create() {
-        BENCHDNN_PRINT(1, "create: %s\n", prb_.str());
+    int create(bool in_parallel) {
+        // Report creation status for problems only in sequential mode as in
+        // parallel it's still not clear which one failed.
+        if (!in_parallel) BENCHDNN_PRINT(1, "create: %s\n", prb_.str());
+
         if (skip_start(&res_, idx_)) return OK;
         if (bench_mode == bench_mode_t::list) return res_.state = LISTED, OK;
 

diff --git a/tests/benchdnn/utils/task_executor.hpp b/tests/benchdnn/utils/task_executor.hpp
@@ -60,10 +60,10 @@ struct task_executor_t {
         // Special case is needed for THREADPOOL RUNTIME. Both `Parallel_nd` and
         // `createit` calls activate threadpool which causes undesired behavior.
         if (tasks_.size() == 1)
-            tasks_[0].create();
+            tasks_[0].create(/* in_parallel = */ false);
         else
-            benchdnn_parallel_nd(
-                    tasks_.size(), [&](int i) { tasks_[i].create(); });
+            benchdnn_parallel_nd(tasks_.size(),
+                    [&](int i) { tasks_[i].create(/* in_parallel = */ true); });
 
         // Check caches first to avoid filling cache with service reorders.
         for (auto &t : tasks_) {