#ifndef CORE_PROFILE_H #define CORE_PROFILE_H // NOTE: measuring without paying for it. // // always on, a handful of nanoseconds each: // - arena allocation counts / bytes / peak (fields in Arena, plain increments, one thread each) // - process wide malloc / free counts and live heap bytes: the host interposes malloc and bumps // relaxed atomics that each sit on their own cache line, so threads never fight over one // - per frame and per audio callback timers (vdso clock_gettime, no syscall) // // sampled only while a Profiler view is on screen, twice a second, on the main thread: // - /proc/self/status (rss, peak rss, anon / file / stack), mallinfo2, getrusage // - per registered thread: cpu time (thread clock), page faults and context switches // (/proc/self/task/), touched stack (mincore over the thread's stack) #define PROFILE_MAX_THREADS 24 #define PROFILE_MAX_ARENAS 32 #define PROFILE_FRAMES 256 enum ProfileUIPhase { PROFILE_UI_KEYBOARD, PROFILE_UI_PANELS, PROFILE_UI_STATUS, PROFILE_UI_OVERLAY, PROFILE_UI_PHASES, }; #define PROFILE_SITES 4096 // allocation site table, pow2 typedef struct ProfileThread ProfileThread; struct ProfileThread { char name[16]; s32 tid; b32 main; clockid_t clock; u8 *stack_low; // lowest address of the stack mapping u64 stack_size; // sampled u64 stack_touched; // resident stack pages: the high water mark u64 cpu_ns; f32 cpu; // fraction of one core u64 minflt, majflt; u64 vcsw, ivcsw; f32 minflt_per_s; f32 csw_per_s; }; typedef struct ProfileArena ProfileArena; struct ProfileArena { char name[24]; Arena *arena; u64 last_count; f32 allocs_per_s; }; typedef struct Profiler Profiler; struct Profiler { //- heap, written from any thread by the malloc interposer CacheAligned u64 malloc_count; CacheAligned u64 free_count; CacheAligned s64 heap_live; // bytes, from malloc_usable_size CacheAligned u64 heap_peak; CacheAligned u64 heap_total; // bytes ever handed out //- allocation sites (MP3_PLAYER_ALLOC_SITES=1): caller address -> count, lock free b32 track_sites; u64 site_address[PROFILE_SITES]; u64 site_count[PROFILE_SITES]; u64 site_bytes[PROFILE_SITES]; //- registry CacheAligned SpinLock lock; ProfileThread threads[PROFILE_MAX_THREADS]; u32 thread_count; ProfileArena arenas[PROFILE_MAX_ARENAS]; u32 arena_count; //- frames, main thread // NOTE: rings over the last 256 rendered frames, every statistic comes from the same window f32 frame_ms[PROFILE_FRAMES]; // cpu ms per rendered frame f32 ui_ring[PROFILE_FRAMES]; // building the ui f32 render_ring[PROFILE_FRAMES]; // vk_render (wait + record + submit) and friends f32 idle_ring[PROFILE_FRAMES]; // blocked in poll before the frame u32 frame_head; u32 frames_total; f32 ui_ms, ui_p99; // over the ring, at the last sample f32 render_ms, render_p99; f32 idle_ms; f32 ui_phase_us[4]; // ProfileUIPhase, smoothed, written by the ui each frame //- samples, main thread, 2 hz while visible u64 sample_ns; b32 sampling; // a Profiler view is on screen u64 page_size; u64 rss, rss_peak, rss_anon, rss_file, rss_shmem, vm_size, vm_data, vm_stack, vm_lib; u64 heap_arena, heap_in_use, heap_free, heap_mmapped; u64 minflt, majflt, vcsw, ivcsw; f32 minflt_per_s, majflt_per_s, csw_per_s; f32 mallocs_per_s, frees_per_s; u64 last_malloc_count, last_free_count; f32 cpu_process; u64 cpu_process_ns; f32 p50, p99, worst, average; // frame ms over the ring }; global_variable Profiler profiler_instance; global_variable Profiler *global_profiler = &profiler_instance; internal void profile_thread_register(const char *name); // call from the thread itself internal void profile_arena_register(const char *name, Arena *arena); internal void profile_frame(Profiler *profiler, f32 frame_ms, f32 ui_ms, f32 render_ms, f32 idle_ms); internal void profile_sample(Profiler *profiler); // main thread #endif // CORE_PROFILE_H