benchdnn: improve memory estimation

tests/benchdnn/binary/bench_binary.cpp

@@ -1,5 +1,5 @@
 /*******************************************************************************
-* Copyright 2019-2024 Intel Corporation
+* Copyright 2019-2025 Intel Corporation
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -25,17 +25,7 @@
 
 namespace binary {
 
-using create_func_t = std::function<int(
-        std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &, const prb_t *,
-        res_t *)>;
-using check_cache_func_t = std::function<int(
-        std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &, const prb_t *,
-        res_t *)>;
-using do_func_t = std::function<int(
-        const std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &,
-        const prb_t *, res_t *)>;
-using driver_task_executor_t = task_executor_t<prb_t, perf_report_t,
-        create_func_t, check_cache_func_t, do_func_t>;
+TASK_EXECUTOR_DECL_TYPES;
 
 void check_correctness(
         const settings_t &s, driver_task_executor_t &task_executor) {
@@ -52,8 +42,7 @@ void check_correctness(
                 i_inplace, i_attr, i_ctx_init, i_ctx_exe, s.impl_filter);
         if (s.pattern && !match_regex(prb.str(), s.pattern)) return;
 
-        task_executor.submit(
-                prb, s.perf_template, createit, check_cacheit, doit);
+        task_executor.submit(prb, s.perf_template, createit, checkit, doit);
     }
 }
 

tests/benchdnn/binary/binary.cpp

@@ -295,10 +295,15 @@ int createit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
     return OK;
 }
 
-int check_cacheit(
-        std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
+int checkit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
         const prb_t *prb, res_t *res) {
-    return check_caches(v_prim[0], prb, res);
+    if (has_bench_mode_bit(mode_bit_t::exec)) {
+        SAFE(check_total_size(res), WARN);
+    }
+    if (has_bench_mode_bit(mode_bit_t::corr)) {
+        SAFE(check_caches(v_prim[0], prb, res), WARN);
+    }
+    return OK;
 }
 
 int doit(const std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,

tests/benchdnn/binary/binary.hpp

@@ -1,5 +1,5 @@
 /*******************************************************************************
-* Copyright 2019-2024 Intel Corporation
+* Copyright 2019-2025 Intel Corporation
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -167,8 +167,7 @@ void compute_ref(const prb_t *prb, const args_t &args,
 
 int createit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
         const prb_t *prb, res_t *res);
-int check_cacheit(
-        std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
+int checkit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
         const prb_t *prb, res_t *res);
 int doit(const std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
         const prb_t *prb, res_t *res);

tests/benchdnn/bnorm/bench_bnorm.cpp

@@ -1,5 +1,5 @@
 /*******************************************************************************
-* Copyright 2017-2024 Intel Corporation
+* Copyright 2017-2025 Intel Corporation
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -26,17 +26,7 @@
 
 namespace bnorm {
 
-using create_func_t = std::function<int(
-        std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &, const prb_t *,
-        res_t *)>;
-using check_cache_func_t = std::function<int(
-        std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &, const prb_t *,
-        res_t *)>;
-using do_func_t = std::function<int(
-        const std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &,
-        const prb_t *, res_t *)>;
-using driver_task_executor_t = task_executor_t<prb_t, perf_report_t,
-        create_func_t, check_cache_func_t, do_func_t>;
+TASK_EXECUTOR_DECL_TYPES;
 
 void check_correctness(
         const settings_t &s, driver_task_executor_t &task_executor) {
@@ -55,8 +45,7 @@ void check_correctness(
                 i_ctx_init, i_ctx_exe, s.impl_filter);
         if (s.pattern && !match_regex(prb.str(), s.pattern)) return;
 
-        task_executor.submit(
-                prb, s.perf_template, createit, check_cacheit, doit);
+        task_executor.submit(prb, s.perf_template, createit, checkit, doit);
     }
 }
 

tests/benchdnn/bnorm/bnorm.cpp

@@ -1,5 +1,5 @@
 /*******************************************************************************
-* Copyright 2017-2024 Intel Corporation
+* Copyright 2017-2025 Intel Corporation
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -718,11 +718,16 @@ int createit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
     return OK;
 }
 
-int check_cacheit(
-        std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
+int checkit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
         const prb_t *prb, res_t *res) {
-    SAFE(check_caches(v_prim[0], prb, res), WARN);
-    if (v_prim[1]) { SAFE(check_caches(v_prim[1], prb, res), WARN); }
+    if (has_bench_mode_bit(mode_bit_t::exec)) {
+        SAFE(check_total_size(res), WARN);
+        if (v_prim[1]) SAFE(check_total_size(res), WARN);
+    }
+    if (has_bench_mode_bit(mode_bit_t::corr)) {
+        SAFE(check_caches(v_prim[0], prb, res), WARN);
+        if (v_prim[1]) { SAFE(check_caches(v_prim[1], prb, res), WARN); }
+    }
     return OK;
 }
 

tests/benchdnn/bnorm/bnorm.hpp

@@ -1,5 +1,5 @@
 /*******************************************************************************
-* Copyright 2017-2024 Intel Corporation
+* Copyright 2017-2025 Intel Corporation
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -278,8 +278,7 @@ void compute_ref(const prb_t *prb, const args_t &args,
 
 int createit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
         const prb_t *prb, res_t *res);
-int check_cacheit(
-        std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
+int checkit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
         const prb_t *prb, res_t *res);
 int doit(const std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
         const prb_t *prb, res_t *res);

tests/benchdnn/common.cpp

@@ -239,7 +239,7 @@ static void zfree_protect(void *ptr) {
 
 void *zmalloc(size_t size, size_t align) {
 #ifdef BENCHDNN_MEMORY_CHECK
-    if (has_bench_mode_bit(mode_bit_t::corr)) { return zmalloc_protect(size); }
+    if (has_bench_mode_bit(mode_bit_t::exec)) { return zmalloc_protect(size); }
 #endif
 
     void *ptr;
@@ -264,7 +264,7 @@ void *zmalloc(size_t size, size_t align) {
 void zfree(void *ptr) {
     if (!ptr) return;
 #ifdef BENCHDNN_MEMORY_CHECK
-    if (has_bench_mode_bit(mode_bit_t::corr)) {
+    if (has_bench_mode_bit(mode_bit_t::exec)) {
         zfree_protect(ptr);
         return;
     }

tests/benchdnn/concat/bench_concat.cpp

@@ -1,5 +1,5 @@
 /*******************************************************************************
-* Copyright 2019-2024 Intel Corporation
+* Copyright 2019-2025 Intel Corporation
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -25,17 +25,7 @@
 
 namespace concat {
 
-using create_func_t = std::function<int(
-        std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &, const prb_t *,
-        res_t *)>;
-using check_cache_func_t = std::function<int(
-        std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &, const prb_t *,
-        res_t *)>;
-using do_func_t = std::function<int(
-        const std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &,
-        const prb_t *, res_t *)>;
-using driver_task_executor_t = task_executor_t<prb_t, perf_report_t,
-        create_func_t, check_cache_func_t, do_func_t>;
+TASK_EXECUTOR_DECL_TYPES;
 
 void check_correctness(
         const settings_t &s, driver_task_executor_t &task_executor) {
@@ -51,8 +41,7 @@ void check_correctness(
                 i_attr, i_ctx_init, i_ctx_exe, s.impl_filter);
         if (s.pattern && !match_regex(prb.str(), s.pattern)) return;
 
-        task_executor.submit(
-                prb, s.perf_template, createit, check_cacheit, doit);
+        task_executor.submit(prb, s.perf_template, createit, checkit, doit);
     }
 }
 

tests/benchdnn/concat/concat.cpp

@@ -1,5 +1,5 @@
 /*******************************************************************************
-* Copyright 2019-2024 Intel Corporation
+* Copyright 2019-2025 Intel Corporation
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -196,21 +196,28 @@ int createit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
     return OK;
 }
 
-int check_cacheit(
-        std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
+int checkit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
         const prb_t *prb, res_t *res) {
-    // Assume it doesn't change through the execution.
-    static int capacity = 0;
-    static auto st = dnnl_get_primitive_cache_capacity(&capacity);
-    if (st != dnnl_success) return FAIL;
-    if (capacity > 0 && prb->n_inputs() + 1 > capacity) {
-        BENCHDNN_PRINT(2, "%s\n",
-                "[INFO] The number of potential internal reorder pds plus "
-                "concat itself exceeds the cache capacity which will lead to a "
-                "test case false-positive failure.");
-        return OK;
+    if (has_bench_mode_bit(mode_bit_t::exec)) {
+        SAFE(check_total_size(res), WARN);
     }
-    return check_caches(v_prim[0], prb, res);
+    if (has_bench_mode_bit(mode_bit_t::corr)) {
+        // The assumtion is the capacity doesn't change through the execution.
+        static int capacity = 0;
+        static auto st = dnnl_get_primitive_cache_capacity(&capacity);
+        if (st != dnnl_success) return FAIL;
+
+        if (capacity > 0 && prb->n_inputs() + 1 > capacity) {
+            BENCHDNN_PRINT(2, "%s\n",
+                    "[INFO] The number of potential internal reorder pds plus "
+                    "concat itself exceeds the cache capacity which will lead "
+                    "to a test case false-positive failure.");
+            return OK;
+        }
+
+        SAFE(check_caches(v_prim[0], prb, res), WARN);
+    }
+    return OK;
 }
 
 int doit(const std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,

tests/benchdnn/concat/concat.hpp

@@ -1,5 +1,5 @@
 /*******************************************************************************
-* Copyright 2019-2024 Intel Corporation
+* Copyright 2019-2025 Intel Corporation
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -176,8 +176,7 @@ void compute_ref(const prb_t *prb, const args_t &args,
 
 int createit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
         const prb_t *prb, res_t *res);
-int check_cacheit(
-        std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
+int checkit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
         const prb_t *prb, res_t *res);
 int doit(const std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
         const prb_t *prb, res_t *res);

tests/benchdnn/conv/bench_conv.cpp

@@ -27,17 +27,7 @@
 
 namespace conv {
 
-using create_func_t = std::function<int(
-        std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &, const prb_t *,
-        res_t *)>;
-using check_cache_func_t = std::function<int(
-        std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &, const prb_t *,
-        res_t *)>;
-using do_func_t = std::function<int(
-        const std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &,
-        const prb_t *, res_t *)>;
-using driver_task_executor_t = task_executor_t<prb_t, perf_report_t,
-        create_func_t, check_cache_func_t, do_func_t>;
+TASK_EXECUTOR_DECL_TYPES;
 
 void check_correctness(
         const settings_t &s, driver_task_executor_t &task_executor) {
@@ -76,11 +66,11 @@ void check_correctness(
         bool has_dw_po = i_attr.post_ops.convolution_index() >= 0;
         auto &conv_createit
                 = has_dw_po ? conv_dw_fusion::createit : conv::createit;
-        auto &conv_check_cacheit = has_dw_po ? conv_dw_fusion::check_cacheit
-                                             : conv::check_cacheit;
+        auto &conv_checkit
+                = has_dw_po ? conv_dw_fusion::checkit : conv::checkit;
         auto &conv_doit = has_dw_po ? conv_dw_fusion::doit : conv::doit;
-        task_executor.submit(prb, s.perf_template, conv_createit,
-                conv_check_cacheit, conv_doit);
+        task_executor.submit(
+                prb, s.perf_template, conv_createit, conv_checkit, conv_doit);
     }
 }
 

tests/benchdnn/conv/conv.cpp

@@ -333,7 +333,6 @@ int init_prim_ref(benchdnn_dnnl_wrapper_t<dnnl_primitive_t> &prim_ref,
         prim_ref_dt.erase(prim_ref_dt.begin());
         prim_ref_bia_dt.erase(prim_ref_bia_dt.begin());
     }
-    dnnl_primitive_t prim_ref_ {};
 
     for_(const auto &prim_ref_dt_i : prim_ref_dt)
     for (const auto &prim_ref_bia_dt_i : prim_ref_bia_dt) {
@@ -342,35 +341,11 @@ int init_prim_ref(benchdnn_dnnl_wrapper_t<dnnl_primitive_t> &prim_ref,
                 prb->mb, cpu_attr, prb->ctx_init, prb->ctx_exe,
                 prb->impl_filter};
 
-        init_pd_args_t<prb_t> init_pd_args(
-                /* res = */ nullptr, get_cpu_engine(), &prb_cpu, prb->dir,
-                /* hint = */ nullptr, /* src_md = */ nullptr);
-        init_pd(init_pd_args);
-
-        benchdnn_dnnl_wrapper_t<dnnl_primitive_desc_t> pdw;
-        // `is_service_prim=true` prevents from filtering the implementation
-        // by name which is intended through a `get_prim_ref_impl_filter()`.
-        // As `fetch_impl` doesn't have any further logic related to it, it's
-        // safe to set it to `false`.
-        fetch_impl(pdw, init_pd_args, get_prim_ref_impl_filter(),
-                /* res = */ nullptr,
-                /* is_service_prim = */ false);
-
-        // Prim desc wasn't created - try the next set...
-        if (!pdw) continue;
-
-        auto st = dnnl_primitive_create(&prim_ref_, pdw);
-        // Primitive wasn't created - try the next set...
-        if (st != dnnl_success) continue;
-
-        BENCHDNN_PRINT(5, "CPU reference oneDNN implementation: %s\n",
-                query_impl_info(pdw).c_str());
-        res->prim_ref_repro = prb_cpu.str();
-        prim_ref.reset(prim_ref_);
-        return OK;
+        auto st = init_prim_ref_common(prim_ref, &prb_cpu, res);
+        if (st == OK) return OK;
     }
 
-    prim_ref.reset(prim_ref_);
+    prim_ref.reset(nullptr);
     return OK;
 }
 
@@ -577,11 +552,26 @@ int createit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
     return OK;
 }
 
-int check_cacheit(
-        std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
+int checkit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
         const prb_t *prb, res_t *res) {
-    SAFE(check_caches(v_prim[0], prb, res), WARN);
-    // Don't check caches for CPU prim as the reference.
+    if (has_bench_mode_bit(mode_bit_t::exec)) {
+        const auto &prim_ref = v_prim[1];
+        if (prim_ref) {
+            // Copy res to avoid save/restore state and reason.
+            res_t res_copy = *res;
+            SAFE(check_total_size(&res_copy, prim_ref), WARN);
+            if (res_copy.state == SKIPPED) {
+                v_prim[1].reset(nullptr);
+                SAFE(check_total_size(res), WARN);
+            }
+        } else {
+            SAFE(check_total_size(res), WARN);
+        }
+    }
+    if (has_bench_mode_bit(mode_bit_t::corr)) {
+        SAFE(check_caches(v_prim[0], prb, res), WARN);
+        // Don't check caches for CPU prim as the reference.
+    }
     return OK;
 }
 

tests/benchdnn/conv/conv.hpp

@@ -312,8 +312,7 @@ void compute_ref(const prb_t *prb, const args_t &args,
 
 int createit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
         const prb_t *prb, res_t *res);
-int check_cacheit(
-        std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
+int checkit(std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
         const prb_t *prb, res_t *res);
 int doit(const std::vector<benchdnn_dnnl_wrapper_t<dnnl_primitive_t>> &v_prim,
         const prb_t *prb, res_t *res);