Use fixed size array for serialization (#906)

Co-authored-by: Benjamin Welton <ben@amd.com>

source/lib/rocprofiler-sdk-tool/generateCSV.cpp

@@ -30,6 +30,7 @@
 #include <rocprofiler-sdk/marker/api_id.h>
 #include <unistd.h>
 
+#include <cstdint>
 #include <iomanip>
 #include <string_view>
 #include <utility>
@@ -553,8 +554,9 @@ generate_csv(tool_table*                                                     too
     {
         auto kernel_id          = record.dispatch_data.dispatch_info.kernel_id;
         auto counter_name_value = std::map<std::string, uint64_t>{};
-        for(const auto& count : record.records)
+        for(uint64_t i = 0; i < record.counter_count; i++)
         {
+            const auto& count        = record.records.at(i);
             auto        rec          = count.record_counter;
             std::string counter_name = tool_functions->tool_get_counter_info_name_fn(rec.id);
             auto        search       = counter_name_value.find(counter_name);

source/lib/rocprofiler-sdk-tool/helper.hpp

@@ -35,6 +35,7 @@
 #include <rocprofiler-sdk/fwd.h>
 #include <rocprofiler-sdk/registration.h>
 #include <rocprofiler-sdk/rocprofiler.h>
+#include <cstdint>
 #include <rocprofiler-sdk/cxx/name_info.hpp>
 #include <rocprofiler-sdk/cxx/serialization.hpp>
 
@@ -270,18 +271,22 @@ struct rocprofiler_tool_record_counter_t
 
 struct rocprofiler_tool_counter_collection_record_t
 {
-    rocprofiler_profile_counting_dispatch_data_t   dispatch_data    = {};
-    std::vector<rocprofiler_tool_record_counter_t> records          = {};
-    uint64_t                                       thread_id        = 0;
-    uint64_t                                       arch_vgpr_count  = 0;
-    uint64_t                                       sgpr_count       = 0;
-    uint64_t                                       lds_block_size_v = 0;
+    rocprofiler_profile_counting_dispatch_data_t       dispatch_data    = {};
+    std::array<rocprofiler_tool_record_counter_t, 256> records          = {};
+    uint64_t                                           thread_id        = 0;
+    uint64_t                                           arch_vgpr_count  = 0;
+    uint64_t                                           sgpr_count       = 0;
+    uint64_t                                           lds_block_size_v = 0;
+    uint64_t                                           counter_count    = 0;
 
     template <typename ArchiveT>
     void save(ArchiveT& ar) const
     {
         ar(cereal::make_nvp("dispatch_data", dispatch_data));
-        ar(cereal::make_nvp("records", records));
+        // should be removed when moving to buffered tracing
+        std::vector<rocprofiler_tool_record_counter_t> tmp{records.begin(),
+                                                           records.begin() + counter_count};
+        ar(cereal::make_nvp("records", tmp));
         ar(cereal::make_nvp("thread_id", thread_id));
         ar(cereal::make_nvp("arch_vgpr_count", arch_vgpr_count));
         ar(cereal::make_nvp("sgpr_count", sgpr_count));

source/lib/rocprofiler-sdk-tool/tool.cpp

@@ -941,11 +941,19 @@ counter_record_callback(rocprofiler_profile_counting_dispatch_data_t dispatch_da
 
     for(size_t count = 0; count < record_count; count++)
     {
+        // Unlikely to trigger, temporary until we move to buffered callbacks
+        if(count >= counter_record.records.size())
+        {
+            ROCP_WARNING << "Exceeded maximum counter capacity, skipping remaining";
+            break;
+        }
+
         auto _counter_id = rocprofiler_counter_id_t{};
         ROCPROFILER_CALL(rocprofiler_query_record_counter_id(record_data[count].id, &_counter_id),
                          "query record counter id");
-        counter_record.records.emplace_back(
-            rocprofiler_tool_record_counter_t{_counter_id, record_data[count]});
+        counter_record.records[count] =
+            rocprofiler_tool_record_counter_t{_counter_id, record_data[count]};
+        counter_record.counter_count++;
     }
 
     write_ring_buffer(counter_record, domain_type::COUNTER_COLLECTION);

tests/counter-collection/validate.py

@@ -28,7 +28,7 @@ def test_counter_values(input_data):
     counter_data = data["buffer_records"]["counter_collection"]
 
     for itr in counter_info:
-        if itr["is_constant"] == 1 and itr["name"] == "size":
+        if itr["is_constant"] == 1:
             continue
         assert itr["id"]["handle"] > 0, f"{itr}"
         assert itr["is_constant"] in (0, 1), f"{itr}"