[rocprofiler-systems] AMD SMI collector using policy-based design (#3703)

Author: adjordje-amd

projects/rocprofiler-systems/cmake/Packages.cmake

@@ -266,6 +266,55 @@ endif()
 
 target_link_libraries(rocprofiler-systems-rocm INTERFACE amd_smi)
 
+# Detect AMD SMI library version from header
+set(_AMDSMI_HEADER "${ROCM_PATH}/include/amd_smi/amdsmi.h")
+if(EXISTS "${_AMDSMI_HEADER}")
+    file(READ "${_AMDSMI_HEADER}" _AMDSMI_HEADER_CONTENTS)
+
+    string(
+        REGEX MATCH
+        "#define AMDSMI_LIB_VERSION_MAJOR ([0-9]+)"
+        _
+        "${_AMDSMI_HEADER_CONTENTS}"
+    )
+    set(ROCPROFSYS_AMDSMI_VERSION_MAJOR "${CMAKE_MATCH_1}")
+
+    string(
+        REGEX MATCH
+        "#define AMDSMI_LIB_VERSION_MINOR ([0-9]+)"
+        _
+        "${_AMDSMI_HEADER_CONTENTS}"
+    )
+    set(ROCPROFSYS_AMDSMI_VERSION_MINOR "${CMAKE_MATCH_1}")
+
+    message(
+        STATUS
+        "AMD SMI version detected: ${ROCPROFSYS_AMDSMI_VERSION_MAJOR}.${ROCPROFSYS_AMDSMI_VERSION_MINOR}"
+    )
+endif()
+
+# AINIC requires AMD SMI >= 26.3 AND ROCPROFSYS_USE_AINIC option
+set(ROCPROFSYS_BUILD_AINIC OFF CACHE INTERNAL "Build AINIC support")
+if(ROCPROFSYS_USE_AINIC)
+    if(
+        ROCPROFSYS_AMDSMI_VERSION_MAJOR GREATER 26
+        OR (
+            ROCPROFSYS_AMDSMI_VERSION_MAJOR EQUAL 26
+            AND ROCPROFSYS_AMDSMI_VERSION_MINOR GREATER 2
+        )
+    )
+        set(ROCPROFSYS_BUILD_AINIC ON CACHE INTERNAL "Build AINIC support" FORCE)
+        message(STATUS "AINIC support enabled (AMD SMI >= 26.3)")
+    else()
+        message(
+            STATUS
+            "AINIC disabled: AMD SMI ${ROCPROFSYS_AMDSMI_VERSION_MAJOR}.${ROCPROFSYS_AMDSMI_VERSION_MINOR} < 26.3"
+        )
+    endif()
+else()
+    message(STATUS "AINIC disabled: ROCPROFSYS_USE_AINIC is OFF")
+endif()
+
 # ----------------------------------------------------------------------------------------#
 #
 # ROCpd

projects/rocprofiler-systems/source/lib/common/CMakeLists.txt

@@ -26,6 +26,7 @@ target_sources(
         ${CMAKE_CURRENT_SOURCE_DIR}/invoke.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/join.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/setup.hpp
+        ${CMAKE_CURRENT_SOURCE_DIR}/span.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/traits.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/md5sum.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/static_object.hpp

projects/rocprofiler-systems/source/lib/core/perfetto.hpp

@@ -53,9 +53,10 @@ std::unique_ptr<::perfetto::TracingSession>& get_perfetto_session(
 template <typename Tp>
 struct perfetto_counter_track
 {
-    using track_map_t = std::map<uint32_t, std::vector<::perfetto::CounterTrack>>;
-    using name_map_t  = std::map<uint32_t, std::vector<std::unique_ptr<std::string>>>;
-    using data_t      = std::pair<name_map_t, track_map_t>;
+    using category_type = Tp;
+    using track_map_t   = std::map<uint32_t, std::vector<::perfetto::CounterTrack>>;
+    using name_map_t    = std::map<uint32_t, std::vector<std::unique_ptr<std::string>>>;
+    using data_t        = std::pair<name_map_t, track_map_t>;
 
     static auto   init() { (void) get_data(); }
     static auto   exists(size_t _idx, int64_t _n = -1);

projects/rocprofiler-systems/source/lib/core/trace_cache/cache_manager.hpp

@@ -26,9 +26,9 @@
 #include "core/trace_cache/metadata_registry.hpp"
 #include "core/trace_cache/sample_type.hpp"
 #include "core/trace_cache/storage_parser.hpp"
-
+#include "library/pmc/collectors/gpu/sample.hpp"
+#include "library/pmc/collectors/nic/sample.hpp"
 #include "library/runtime.hpp"
-
 #include <memory>
 #include <unistd.h>
 
@@ -40,8 +40,9 @@ namespace trace_cache
 using storage_parser_t =
     storage_parser<type_identifier_t, kernel_dispatch_sample, memory_copy_sample,
                    memory_allocate_sample, region_sample, in_time_sample,
-                   pmc_event_with_sample, amd_smi_sample, cpu_freq_sample,
-                   backtrace_region_sample, scratch_memory_sample, ainic_sample>;
+                   pmc_event_with_sample, pmc::collectors::gpu::sample,
+                   pmc::collectors::nic::sample, cpu_freq_sample, backtrace_region_sample,
+                   scratch_memory_sample>;
 
 using buffer_storage_t = buffer_storage<flush_worker_factory_t, type_identifier_t>;
 

projects/rocprofiler-systems/source/lib/core/trace_cache/cache_type_traits.hpp

@@ -22,6 +22,7 @@
 
 #pragma once
 #include "common/span.hpp"
+#include <array>
 #include <cstdint>
 #include <optional>
 #include <string_view>
@@ -104,6 +105,17 @@ struct is_vector<std::vector<T>> : std::true_type
 template <typename T>
 inline constexpr bool is_vector_v = is_vector<T>::value;
 
+template <typename T>
+struct is_array : std::false_type
+{};
+
+template <typename T, size_t N>
+struct is_array<std::array<T, N>> : std::true_type
+{};
+
+template <typename T>
+inline constexpr bool is_array_v = is_array<T>::value;
+
 template <typename T>
 static constexpr bool is_string_view_v =
     std::is_same_v<std::decay_t<T>, std::string_view>;
@@ -122,7 +134,7 @@ inline constexpr bool is_optional_v = is_optional<T>::value;
 template <typename T>
 inline constexpr bool is_supported_type_v =
     is_span_v<T> || std::is_integral_v<T> || std::is_floating_point_v<T> ||
-    is_string_view_v<T> || is_vector_v<T> || is_optional_v<T>;
+    is_string_view_v<T> || is_vector_v<T> || is_optional_v<T> || is_array_v<T>;
 
 template <typename T>
 struct is_enum_class

projects/rocprofiler-systems/source/lib/core/trace_cache/metadata_registry.hpp

@@ -37,7 +37,7 @@
 #include <rocprofiler-sdk/callback_tracing.h>
 #include <rocprofiler-sdk/cxx/name_info.hpp>
 #include <set>
-#include <sstream>
+#include <spdlog/fmt/ranges.h>
 #include <stdint.h>
 #include <string.h>
 #include <string>
@@ -62,18 +62,6 @@ struct process
     uint32_t    end;
 };
 
-template <typename Category>
-inline std::string
-annotate_category(std::optional<int> first_section  = std::nullopt,
-                  std::optional<int> second_section = std::nullopt)
-{
-    std::stringstream ss;
-    ss << std::string(tim::trait::name<Category>::value);
-    if(first_section) ss << "_" << std::to_string(*first_section);
-    if(second_section) ss << "_" << std::to_string(*second_section);
-    return ss.str();
-}
-
 struct pmc
 {
     agent_type  type;
@@ -131,16 +119,12 @@ struct thread
 
 template <typename Category>
 inline std::string
-annotate_with_device_id(uint32_t           device_id,
-                        std::optional<int> first_section  = std::nullopt,
-                        std::optional<int> second_section = std::nullopt)
+format_track_name(std::optional<int> first_section  = std::nullopt,
+                  std::optional<int> second_section = std::nullopt)
 {
-    std::stringstream ss;
-    ss << std::string(tim::trait::name<Category>::value) + " [" +
-              std::to_string(device_id) + "]";
-    if(first_section) ss << "_" << std::to_string(*first_section);
-    if(second_section) ss << "_" << std::to_string(*second_section);
-    return ss.str();
+    return fmt::format("{}{}{}", tim::trait::name<Category>::value,
+                       first_section ? fmt::format("_{}", *first_section) : "",
+                       second_section ? fmt::format("_{}", *second_section) : "");
 }
 
 template <typename Category>