[Dynamic Selection] Adding sycl profiling for auto tune policy (#1464)

* Added static assert for keyArgs==Args * WIP for jit compilation * Fixing Autotune bug * Added sycl profiling without host task * Corrections in dynamic traits * Remove comments in dynamic_selection_traits * Adding back jit compolation restrictions * Adressing comments to add thread safety, better traits * Added backend trait to check if profiling is enabled * Backend traits, renaming and adding a lock to lazy report * Fixed memory leaks * No nullptr for selection handle in async waiter constructor * Added an erase and remove_if * Adressing comments for profiling report, Adding profiling to default sycl backend constructor * Changes to std::chrono::duration * Addressed comments for sycl backend * Addressed comments to sycl backend * Fix USM memory leaks and create unique task names (#1537) * test/parallel_api/dynamic_selection/sycl/test_auto_tune_policy_sycl.pass.cpp - fix USM shared memory leaks Signed-off-by: Sergey Kopienko <[email protected]> * test/parallel_api/dynamic_selection/sycl/test_auto_tune_policy_sycl.pass.cpp - create unique task names for h.single_task([](){}); Signed-off-by: Sergey Kopienko <[email protected]> --------- Signed-off-by: Sergey Kopienko <[email protected]> * Fix the format of report method - using `report_duration` in second param (#1539) * include/oneapi/dpl/internal/dynamic_selection_traits.h - fix traits to specify report method second parameter type Signed-off-by: Sergey Kopienko <[email protected]> * Change second parameter type of auto_tune_selection_type::report method - to report_duration (std::chrono::milliseconds) Signed-off-by: Sergey Kopienko <[email protected]> --------- Signed-off-by: Sergey Kopienko <[email protected]> * Fixing clang format --------- Signed-off-by: Sergey Kopienko <[email protected]> Co-authored-by: Sergey Kopienko <[email protected]>
oneapi-src · Apr 30, 2024 · 5ea755c · 5ea755c
1 parent f6a8c40
commit 5ea755c
Show file tree

Hide file tree

Showing 7 changed files with 347 additions and 103 deletions.
diff --git a/include/oneapi/dpl/internal/dynamic_selection_impl/auto_tune_policy.h b/include/oneapi/dpl/internal/dynamic_selection_impl/auto_tune_policy.h
@@ -16,12 +16,14 @@
 #include <mutex>
 #include <utility>
 #include <chrono>
+#include <ratio>
 #include <limits>
 #include <vector>
 #include <type_traits>
 #include <tuple>
 #include <unordered_map>
 #include "oneapi/dpl/internal/dynamic_selection_traits.h"
+#include "oneapi/dpl/internal/dynamic_selection_impl/backend_traits.h"
 #if _DS_BACKEND_SYCL != 0
 #    include "oneapi/dpl/internal/dynamic_selection_impl/sycl_backend.h"
 #endif
@@ -47,13 +49,16 @@ class auto_tune_policy
     using size_type = typename std::vector<typename Backend::resource_type>::size_type;
     using timing_t = uint64_t;
 
+    using report_clock_type = std::chrono::steady_clock;
+    using report_duration = std::chrono::milliseconds;
+
     static constexpr timing_t never_resample = 0;
     static constexpr size_type use_best_resource = ~size_type(0);
 
     struct resource_with_index_t
     {
         wrapped_resource_t r_;
-        size_type index_;
+        size_type index_ = 0;
     };
 
     struct time_data_t
@@ -66,7 +71,7 @@ class auto_tune_policy
     {
         std::mutex m_;
 
-        std::chrono::steady_clock::time_point t0_;
+        report_clock_type::time_point t0_;
 
         timing_t best_timing_ = std::numeric_limits<timing_t>::max();
         resource_with_index_t best_resource_;
@@ -80,7 +85,7 @@ class auto_tune_policy
         timing_t resample_time_ = 0.0;
 
         tuner_t(resource_with_index_t br, size_type resources_size, timing_t rt)
-            : t0_(std::chrono::steady_clock::now()), best_resource_(br), max_resource_to_profile_(resources_size),
+            : t0_(report_clock_type::now()), best_resource_(br), max_resource_to_profile_(resources_size),
               resample_time_(rt)
         {
         }
@@ -100,8 +105,8 @@ class auto_tune_policy
             }
             else
             {
-                auto now = std::chrono::steady_clock::now();
-                auto ms = std::chrono::duration_cast<std::chrono::milliseconds>(now - t0_).count();
+                const auto now = report_clock_type::now();
+                const auto ms = std::chrono::duration_cast<report_duration>(now - t0_).count();
                 if (ms < resample_time_)
                 {
                     return use_best_resource;
@@ -169,9 +174,9 @@ class auto_tune_policy
         };
 
         void
-        report(const execution_info::task_time_t&, const typename execution_info::task_time_t::value_type& v) const
+        report(const execution_info::task_time_t&, report_duration v) const
         {
-            tuner_->add_new_timing(resource_, v);
+            tuner_->add_new_timing(resource_, v.count());
         }
     };
 
@@ -217,6 +222,10 @@ class auto_tune_policy
     select(Function&& f, Args&&... args)
     {
         static_assert(sizeof...(KeyArgs) == sizeof...(Args));
+        if constexpr (backend_traits::lazy_report_v<Backend>)
+        {
+            backend_->lazy_report();
+        }
         if (state_)
         {
             std::lock_guard<std::mutex> l(state_->m_);

diff --git a/include/oneapi/dpl/internal/dynamic_selection_impl/backend_traits.h b/include/oneapi/dpl/internal/dynamic_selection_impl/backend_traits.h
@@ -0,0 +1,55 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Copyright (C) 2023 Intel Corporation
+//
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _ONEDPL_INTERNAL_BACKEND_TRAITS_H
+#define _ONEDPL_INTERNAL_BACKEND_TRAITS_H
+
+#include <utility>
+#include <type_traits>
+
+namespace oneapi
+{
+namespace dpl
+{
+namespace experimental
+{
+namespace internal
+{
+template <typename Backend>
+auto
+has_lazy_report_impl(...) -> std::false_type;
+
+template <typename Backend>
+auto
+has_lazy_report_impl(int) -> decltype(std::declval<Backend>().lazy_report(), std::true_type{});
+
+template <typename Backend>
+struct has_lazy_report : decltype(has_lazy_report_impl<Backend>(0))
+{
+};
+
+} //namespace internal
+
+namespace backend_traits
+{
+template <typename S>
+struct lazy_report_value
+{
+    static constexpr bool value = ::oneapi::dpl::experimental::internal::has_lazy_report<S>::value;
+};
+template <typename S>
+inline constexpr bool lazy_report_v = lazy_report_value<S>::value;
+
+} //namespace backend_traits
+
+} // namespace experimental
+} // namespace dpl
+} // namespace oneapi
+
+#endif /*_ONEDPL_INTERNAL_BACKEND_TRAITS_H*/
diff --git a/include/oneapi/dpl/internal/dynamic_selection_impl/dynamic_load_policy.h b/include/oneapi/dpl/internal/dynamic_selection_impl/dynamic_load_policy.h
@@ -19,6 +19,7 @@
 #include <type_traits>
 #include <utility>
 #include "oneapi/dpl/internal/dynamic_selection_traits.h"
+#include "oneapi/dpl/internal/dynamic_selection_impl/backend_traits.h"
 #if _DS_BACKEND_SYCL != 0
 #    include "oneapi/dpl/internal/dynamic_selection_impl/sycl_backend.h"
 #endif
@@ -150,6 +151,10 @@ struct dynamic_load_policy
     selection_type
     select(Args&&...)
     {
+        if constexpr (backend_traits::lazy_report_v<Backend>)
+        {
+            backend_->lazy_report();
+        }
         if (state_)
         {
             std::lock_guard<std::mutex> l(state_->m_);