Gbaraldi/threadtrace2 (#724)

* Added first ATT API * Finalizing thread trace API * Fixing more rebase conflicts * Added codeobj disassembly sample * Fixing merge issues with rebase [2] * Adding ATT packets * Implemented thread trace intercept * Moved codeobj parser to same repo as rocprofiler * Moved thread trace to new API * Fixing merge conflicts * Fixing more merge conflicts * Adding thread trace packet reuse * Merged aql_profile_v2 headers * Linked ATT sample to aqlprofile * Updated decoder to include non-loaded codeobjs * Implemented ISA decoder into ATT sample * Added marker_id to vaddr * Updating aql_profile_v2 API to memcpy * Updating thread trace API to include 64bit markers. Using the result of ISA matching. * Added instruction type and cycles summary * Updated sample with selection of kernel by kernel_object * Added option to copy from memory kernels * Moved tool_data in thread_trace to dynamic alloc * Restoring hsa.cpp * Fixed ATT sample crash. General improvements. * Moved codeobj library to outside src/ * Updated license header * Moved codeobj_capture to camelcase * Solving some more merge conflicts * Update samples/advanced_thread_trace/CMakeLists.txt Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update samples/advanced_thread_trace/CMakeLists.txt Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update samples/code_object_isa_decode/CMakeLists.txt Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update source/lib/rocprofiler-sdk/thread_trace/CMakeLists.txt * Removing unused parameter check * Adding const to isEmpty * Removing unused warning * Adding libdw-dev to requirements * Running clang-format * Commenting out new aql calls * Clang format * Unused variable fix * Adding codeobj-decoder coverage * Commenting out threadtrace * Update samples/CMakeLists.txt Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * P * WOverloaded * Addressing clang-tidy * Virtual destructor on ttracer class * Corr id * Fixing code source format * Update CMakeLists.txt * Build fixes * Update source/lib/rocprofiler-sdk-codeobj/code_object_track.cpp Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Fix shadowing * Update CMakeLists.txt * Update samples/CMakeLists.txt Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Ammar ELWazir <ammar.elwazir@amd.com> Co-authored-by: Ammar ELWazir <aelwazir@amd.com> Co-authored-by: Benjamin Welton <bewelton@amd.com>
2024-04-08 16:43:02 -03:00
@@ -90,6 +90,9 @@ constexpr rocprofiler_agent_t default_agent =
 void
 QueueController::add_queue(hsa_queue_t* id, std::unique_ptr<Queue> queue)
 {
+    for(auto& pre_initialize_fn : pre_initialize)
+        pre_initialize_fn(queue->get_agent(), get_core_table(), get_ext_table());
+
    CHECK(queue);
    _callback_cache.wlock([&](auto& callbacks) {
        _queues.wlock([&](auto& map) {
@@ -111,6 +114,11 @@ void
 QueueController::destroy_queue(hsa_queue_t* id)
 {
    if(!id) return;
+    _queues.wlock([&](auto& map) {
+        for(auto& deinitialize_fn : pre_deinitialize)
+            if(map.find(id) != map.end())
+                deinitialize_fn(map.at(id)->get_agent(), get_core_table(), get_ext_table());
+    });

    const auto* queue = get_queue(*id);

@@ -195,7 +203,8 @@ QueueController::init(CoreApiTable& core_table, AmdExtTable& ext_table)
    {
        constexpr auto expected_context_size = 160UL;
        static_assert(
-            sizeof(context::context) == expected_context_size,
+            sizeof(context::context) ==
+                expected_context_size + sizeof(std::shared_ptr<rocprofiler::ThreadTracer>),
            "If you added a new field to context struct, make sure there is a check here if it "
            "requires queue interception. Once you have done so, increment expected_context_size");

@@ -212,6 +221,20 @@ QueueController::init(CoreApiTable& core_table, AmdExtTable& ext_table)
                break;
            }
        }
+        else if(itr->thread_trace)
+        {
+            enable_intercepter                             = true;
+            std::weak_ptr<rocprofiler::ThreadTracer> trace = itr->thread_trace;
+            pre_initialize.emplace_back(
+                [trace](const AgentCache& cache, const CoreApiTable& core, const AmdExtTable& ext) {
+                    if(auto locked = trace.lock()) locked->resource_init(cache, core, ext);
+                });
+            pre_deinitialize.emplace_back(
+                [trace](const AgentCache& cache, const CoreApiTable&, const AmdExtTable&) {
+                    if(auto locked = trace.lock()) locked->resource_deinit(cache);
+                });
+            break;
+        }
    }

    if(enable_intercepter)