benchdnn: supply info about prim_ref in case of failure

tests/benchdnn/common.hpp

@@ -168,6 +168,7 @@ struct res_t {
     size_t errors, total;
     timer::timer_map_t timer_map;
     std::string impl_name;
+    std::string prim_ref_repro;
     skip_reason_t reason;
     size_t ibytes, obytes;
     bool mem_check_done;

tests/benchdnn/conv/conv.cpp

@@ -284,9 +284,11 @@ dnnl_status_t init_pd(init_pd_args_t<prb_t> &init_pd_args) {
     return dnnl_success;
 }
 
-int init_prim_ref(
-        benchdnn_dnnl_wrapper_t<dnnl_primitive_t> &prim_ref, const prb_t *prb) {
+int init_prim_ref(benchdnn_dnnl_wrapper_t<dnnl_primitive_t> &prim_ref,
+        const prb_t *prb, res_t *res) {
     if (!(has_bench_mode_bit(mode_bit_t::corr) && fast_ref)) return OK;
+    // Create prim_ref if only original prim was successfully created.
+    if (res->state != INITIALIZED) return OK;
 
     // f32 cases should go through reference no matter what.
     if (is_cpu() && (prb->src_dt() == dnnl_f32 && prb->wei_dt() == dnnl_f32))
@@ -328,6 +330,7 @@ int init_prim_ref(
 
             BENCHDNN_PRINT(5, "CPU reference oneDNN implementation: %s\n",
                     query_impl_info(pdw).c_str());
+            res->prim_ref_repro = prb_cpu.str();
             break;
         }
     }
@@ -534,7 +537,7 @@ int createit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
     v_prim.resize(2); // regular + cpu_ref
     SAFE(init_prim(prb->ctx_init, v_prim[0], init_pd, prb, res), WARN);
     // Use CPU prim as the reference in GPU testing to reduce testing time.
-    SAFE(init_prim_ref(v_prim[1], prb), WARN);
+    SAFE(init_prim_ref(v_prim[1], prb, res), WARN);
     return OK;
 }
 

tests/benchdnn/deconv/deconv.cpp

@@ -285,9 +285,11 @@ dnnl_status_t init_pd(init_pd_args_t<prb_t> &init_pd_args) {
     return dnnl_success;
 }
 
-int init_prim_ref(
-        benchdnn_dnnl_wrapper_t<dnnl_primitive_t> &prim_ref, const prb_t *prb) {
+int init_prim_ref(benchdnn_dnnl_wrapper_t<dnnl_primitive_t> &prim_ref,
+        const prb_t *prb, res_t *res) {
     if (!(has_bench_mode_bit(mode_bit_t::corr) && fast_ref)) return OK;
+    // Create prim_ref if only original prim was successfully created.
+    if (res->state != INITIALIZED) return OK;
 
     // f32 cases should go through reference no matter what.
     if (is_cpu() && (prb->src_dt() == dnnl_f32 && prb->wei_dt() == dnnl_f32))
@@ -326,6 +328,7 @@ int init_prim_ref(
 
             BENCHDNN_PRINT(5, "CPU reference oneDNN implementation: %s\n",
                     query_impl_info(pdw).c_str());
+            res->prim_ref_repro = prb_cpu.str();
             break;
         }
     }
@@ -527,7 +530,7 @@ int createit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
     v_prim.resize(2); // regular + cpu_ref
     SAFE(init_prim(prb->ctx_init, v_prim[0], init_pd, prb, res), WARN);
     // Use CPU prim as the reference in GPU testing to reduce testing time.
-    SAFE(init_prim_ref(v_prim[1], prb), WARN);
+    SAFE(init_prim_ref(v_prim[1], prb, res), WARN);
     return OK;
 }
 

tests/benchdnn/dnnl_common.hpp

@@ -32,6 +32,7 @@
 #include "utils/dims.hpp"
 #include "utils/dnnl_query.hpp"
 #include "utils/numeric.hpp"
+#include "utils/parallel.hpp"
 
 #include "tests/test_thread.hpp"
 
@@ -580,6 +581,25 @@ void check_correctness(const prb_t *prb, const std::vector<data_kind_t> &kinds,
 
         TIME_COMPARE(cmp.compare(mem_fp, mem_dt, prb->attr, res));
     }
+
+    if (prim_ref && res->state == FAILED) {
+        static cpu_cache_args_t cpu_cache_args {};
+        SAFE_V(get_cpu_cache_size(cpu_cache_args));
+
+        BENCHDNN_PRINT(0,
+                "[PRIM_REF][INFO]: L2_size:%zu bytes; per_core_L3_size:%zu "
+                "bytes; nthr:%d; impl_name:%s\n",
+                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");
+        res->prim_ref_repro[eng_pos + 7] = 'c'; // Replace `g` in `gpu` with `c`
+
+        BENCHDNN_PRINT(
+                0, "[PRIM_REF][REPRO]: %s\n", res->prim_ref_repro.c_str());
+    }
 }
 
 typedef std::function<dnnl_status_t(

tests/benchdnn/ip/ip.cpp

@@ -88,9 +88,11 @@ dnnl_status_t init_pd(init_pd_args_t<prb_t> &init_pd_args) {
     return dnnl_success;
 }
 
-int init_prim_ref(
-        benchdnn_dnnl_wrapper_t<dnnl_primitive_t> &prim_ref, const prb_t *prb) {
+int init_prim_ref(benchdnn_dnnl_wrapper_t<dnnl_primitive_t> &prim_ref,
+        const prb_t *prb, res_t *res) {
     if (!(has_bench_mode_bit(mode_bit_t::corr) && fast_ref)) return OK;
+    // Create prim_ref if only original prim was successfully created.
+    if (res->state != INITIALIZED) return OK;
 
     // f32 cases should go through reference no matter what.
     if (is_cpu() && (prb->src_dt() == dnnl_f32 && prb->wei_dt() == dnnl_f32))
@@ -126,6 +128,7 @@ int init_prim_ref(
 
             BENCHDNN_PRINT(5, "CPU reference oneDNN implementation: %s\n",
                     query_impl_info(pdw).c_str());
+            res->prim_ref_repro = prb_cpu.str();
             break;
         }
     }
@@ -375,7 +378,7 @@ int createit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
     v_prim.resize(2); // regular + cpu_ref
     SAFE(init_prim(prb->ctx_init, v_prim[0], init_pd, prb, res), WARN);
     // Use CPU prim as the reference in GPU testing to reduce testing time.
-    SAFE(init_prim_ref(v_prim[1], prb), WARN);
+    SAFE(init_prim_ref(v_prim[1], prb, res), WARN);
     return OK;
 }
 

tests/benchdnn/matmul/matmul.cpp

@@ -147,9 +147,11 @@ dnnl_status_t init_pd(init_pd_args_t<prb_t> &init_pd_args) {
     return dnnl_success;
 }
 
-int init_prim_ref(
-        benchdnn_dnnl_wrapper_t<dnnl_primitive_t> &prim_ref, const prb_t *prb) {
+int init_prim_ref(benchdnn_dnnl_wrapper_t<dnnl_primitive_t> &prim_ref,
+        const prb_t *prb, res_t *res) {
     if (!(has_bench_mode_bit(mode_bit_t::corr) && fast_ref)) return OK;
+    // Create prim_ref if only original prim was successfully created.
+    if (res->state != INITIALIZED) return OK;
 
     // f32 cases should go through reference no matter what.
     if (is_cpu() && (prb->src_dt() == dnnl_f32 && prb->wei_dt() == dnnl_f32))
@@ -206,6 +208,7 @@ int init_prim_ref(
 
             BENCHDNN_PRINT(5, "CPU reference oneDNN implementation: %s\n",
                     query_impl_info(pdw).c_str());
+            res->prim_ref_repro = prb_cpu.str();
             break;
         }
     }
@@ -690,7 +693,7 @@ int createit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
     v_prim.resize(2); // regular + cpu_ref
     SAFE(init_prim(prb->ctx_init, v_prim[0], init_pd, prb, res), WARN);
     // Use CPU prim as the reference in GPU testing to reduce testing time.
-    SAFE(init_prim_ref(v_prim[1], prb), WARN);
+    SAFE(init_prim_ref(v_prim[1], prb, res), WARN);
     return OK;
 }