Simple profiling

1 files changed, 166 insertions, 4 deletions

diff --git a/reduced_system_layer.c b/reduced_system_layer.c
index 6473039..acd8cc6 100755
--- a/reduced_system_layer.c
+++ b/reduced_system_layer.c
@@ -53,7 +53,6 @@
 #/      + [x] Requests cache
 #/      + [x] Global anti-aliasing
 #/      + [x] Fill triangles
-#/    - Profiling
 #/    - Examples
 #/      - Conway's Game of Life
 #/      - Julia Set
@@ -139,6 +138,7 @@
 #/    - Testing
 #/    - Stackless coroutines
 #/    - Allocator
+#/    - Profiling
 #/  - Graphics
 #/    - Font
 #/    - Adaptive resolution
@@ -716,6 +716,11 @@ i64 network_send(u16 slot, Network_Address address, i64 size, u8 *data, u16 *loc
 void  dynamic_library_open            (u16 slot, c8 *name);
 void *dynamic_library_get_proc_address(u16 slot, c8 *proc);
 
+//  Profiling
+void PROFILER_init(u32 slot, c8 *name);
+void PROFILER_begin(u32 slot);
+void PROFILER_end(u32 slot);
+
 //  NOTE: This will shutdown main window, networking, audio,
 //        and close all opened dynamic libraries.
 //        If g_platform.graceful_shutdown is 0, this does nothing.
@@ -748,6 +753,13 @@ extern Platform g_platform;
 #ifndef REDUCED_SYSTEM_LAYER_IMPL_GUARD_
 #define REDUCED_SYSTEM_LAYER_IMPL_GUARD_
 
+enum {
+  PROFILE_DOWNSCALE,
+  PROFILE_RESIZE,
+  PROFILE_WAITING,
+  PROFILE_FRAME,
+};
+
 //  ----------------------------------------------------------------
 
 static void mem_set_(void *dst, u8 x, i64 size) {
@@ -1253,8 +1265,11 @@ static i64 _frame_duration [NUM_FRAMES_AVERAGED]          = {0};
 void run_main_window_event_loop(void) {
   init_main_window();
 #if !defined(__wasm__)
-  while (!g_platform.done)
+  while (!g_platform.done) {
+    PROFILER_begin(PROFILE_FRAME);
     update_and_render_frame();
+    PROFILER_end(PROFILE_FRAME);
+  }
   shutdown_all_systems();
 #endif
 }
@@ -1461,6 +1476,117 @@ void resize_dynamic_array_capacity(i64 *num, i64 *capacity, void **data, i64 ele
 
 //  ================================================================
 //
+//  Profiling
+//
+//  ================================================================
+
+typedef struct {
+  c8 *name;
+  i64 count;
+  i64 time_sec;
+  i64 time_nsec;
+  i64 begin_sec;
+  i64 begin_nsec;
+} Profiler_Slot_;
+
+i64             _profiler_num_slots = 0;
+Profiler_Slot_ *_profiler_slots     = NULL;
+
+void PROFILER_init(u32 slot, c8 *name) {
+  if ((i64) (u64) slot >= _profiler_num_slots)
+    resize_dynamic_array_exact(&_profiler_num_slots, (void **) &_profiler_slots, sizeof *_profiler_slots, (i64) (u64) slot + 1);
+
+  if ((i64) (u64) slot >= _profiler_num_slots) {
+    LOG_ERROR("Invalid slot: %d", (i32) slot);
+    return;
+  }
+
+  mem_set_(_profiler_slots + slot, 0, sizeof *_profiler_slots);
+  _profiler_slots[slot].name = name;
+}
+
+void PROFILER_begin(u32 slot) {
+  if ((i64) (u64) slot >= _profiler_num_slots) {
+    LOG_ERROR("Invalid slot: %d", (i32) slot);
+    return;
+  }
+
+  if (_profiler_slots[slot].count < 0) {
+    LOG_ERROR("Sanity");
+    return;
+  }
+
+  ++_profiler_slots[slot].count;
+  if (_profiler_slots[slot].count != 1)
+    return;
+
+  current_utc_time_in_sec_and_nsec(
+    &_profiler_slots[slot].begin_sec,
+    &_profiler_slots[slot].begin_nsec
+  );
+}
+
+static void normalize_time_(i64 *sec, i64 *nsec) {
+  if (sec == NULL || nsec == NULL) {
+    LOG_ERROR("Sanity");
+    return;
+  }
+
+  if (*nsec < 0) {
+    *sec  -= 1;
+    *nsec += 1000000000;
+  }
+
+  if (*nsec >= 1000000000) {
+    *sec += *nsec / 1000000000;
+    *nsec = *nsec % 1000000000;
+  }
+}
+
+void PROFILER_end(u32 slot) {
+  if ((i64) (u64) slot >= _profiler_num_slots) {
+    LOG_ERROR("Invalid slot: %d", (i32) slot);
+    return;
+  }
+
+  if (_profiler_slots[slot].count <= 0) {
+    LOG_ERROR("Sanity");
+    return;
+  }
+
+  --_profiler_slots[slot].count;
+  if (_profiler_slots[slot].count != 0)
+    return;
+
+  i64 sec, nsec;
+  current_utc_time_in_sec_and_nsec(&sec, &nsec);
+
+  sec  -= _profiler_slots[slot].begin_sec;
+  nsec -= _profiler_slots[slot].begin_nsec;
+
+  normalize_time_(&sec, &nsec);
+
+  for (i64 i = 0; i < _profiler_num_slots; ++i) {
+    if (_profiler_slots[i].count == 0)
+      continue;
+
+    _profiler_slots[i].time_sec  -= sec;
+    _profiler_slots[i].time_nsec -= nsec;
+
+    normalize_time_(&_profiler_slots[i].time_sec, &_profiler_slots[i].time_nsec);
+  }
+
+  _profiler_slots[slot].begin_sec  = 0;
+  _profiler_slots[slot].begin_nsec = 0;
+
+  _profiler_slots[slot].time_sec  += sec;
+  _profiler_slots[slot].time_nsec += nsec;
+
+  normalize_time_(&_profiler_slots[slot].time_sec, &_profiler_slots[slot].time_nsec);
+}
+
+//  ================================================================
+//
 //  Platform
 //
 //  ================================================================
@@ -1643,6 +1769,8 @@ static void convert_pixels_for_window_(void) {
       return;
     }
 
+    PROFILER_begin(PROFILE_DOWNSCALE);
+
     f32       k = 1.f / (aa_scale * aa_scale);
     vec4_f32 *s = g_platform.pixels;
     u32      *d = _internal_pixels;
@@ -1671,6 +1799,8 @@ static void convert_pixels_for_window_(void) {
       for (; d < d_end; ++d, ++im)
         *d = rgb_u32_from_f32_((vec3_f32) { .x = im->x * k, .y = im->y * k, .z = im->z * k, }) | 0xff000000;
     }
+
+    PROFILER_end(PROFILE_DOWNSCALE);
   }
 
   //  Resize internal pixel buffer in-place.
@@ -1692,6 +1822,8 @@ static void convert_pixels_for_window_(void) {
       return;
     }
 
+    PROFILER_begin(PROFILE_RESIZE);
+
     f64 di = ((f64) src_width)  / dst_width;
     f64 dj = ((f64) src_height) / dst_height;
 
@@ -1720,6 +1852,8 @@ static void convert_pixels_for_window_(void) {
           *d = s[(i64) ii];
       }
     }
+
+    PROFILER_end(PROFILE_RESIZE);
   }
 }
 
@@ -1781,9 +1915,14 @@ void dynamic_library_open(u16 slot, c8 *name) {
     return;
   }
 
-  if (slot >= _dynamic_libraries.num_slots)
+  if ((i64) (u64) slot >= _dynamic_libraries.num_slots)
     resize_dynamic_array_exact(&_dynamic_libraries.num_slots, (void **) &_dynamic_libraries.slots, sizeof *_dynamic_libraries.slots, (i64) slot + 1);
 
+  if ((i64) (u64) slot >= _dynamic_libraries.num_slots) {
+    LOG_ERROR("Invalid slot: %d", (i32) slot);
+    return;
+  }
+
   if (_dynamic_libraries.slots[slot].handle != NULL) {
     dlclose(_dynamic_libraries.slots[slot].handle);
     _dynamic_libraries.slots[slot].handle = NULL;
@@ -1801,7 +1940,7 @@ void dynamic_library_open(u16 slot, c8 *name) {
 }
 
 void *dynamic_library_get_proc_address(u16 slot, c8 *proc) {
-  if (slot >= _dynamic_libraries.num_slots) {
+  if ((i64) (u64) slot >= _dynamic_libraries.num_slots) {
     LOG_ERROR("Invalid slot: %d", (i32) slot);
     return NULL;
   }
@@ -1948,6 +2087,11 @@ static void network_cleanup_(void) {
 b8 network_open(u16 slot, Network_Address address, u16 *local_port) {
   network_init_(slot);
 
+  if ((i64) (u64) slot >= _num_sockets) {
+    LOG_ERROR("Invalid slot: %d", (i32) slot);
+    return 0;
+  }
+
   b8 change_address =
        !_sockets[slot].ready
     ||  _sockets[slot].address.protocol != address.protocol
@@ -2940,6 +3084,11 @@ void put_image_to_main_window_(void) {
 }
 
 void init_main_window(void) {
+  PROFILER_init(PROFILE_DOWNSCALE, "Downscale anti-aliasing");
+  PROFILER_init(PROFILE_RESIZE,    "Resize in-place");
+  PROFILER_init(PROFILE_WAITING,   "Waiting for events");
+  PROFILER_init(PROFILE_FRAME,     "Update frame");
+
   for (i64 i = 0; i < NUM_FRAMES_AVERAGED; ++i)
     _frame_duration[i] = (MIN_FRAME_DURATION + MAX_FRAME_DURATION) / 2;
 
@@ -3165,6 +3314,16 @@ void init_main_window(void) {
 }
 
 void shutdown_all_systems(void) {
+  if (_profiler_num_slots > 0) {
+    printf("\nPROFILER REPORT\n\n");
+    for (i64 i = 0; i < _profiler_num_slots; ++i) {
+      if (_profiler_slots[i].name == NULL)
+       continue;
+      printf("%-31s %lld.%09lld\n", _profiler_slots[i].name, _profiler_slots[i].time_sec, _profiler_slots[i].time_nsec);
+    }
+    printf("\n");
+  }
+
   g_platform.done = 1;
 
   if (!g_platform.graceful_shutdown)
@@ -3725,6 +3884,8 @@ i32 handle_main_window_events(void) {
 }
 
 i32 wait_main_window_events(void) {
+  PROFILER_begin(PROFILE_WAITING);
+
   i32 num_events = 0;
 
   for (;;) {
@@ -3734,6 +3895,7 @@ i32 wait_main_window_events(void) {
     usleep(0);
   }
 
+  PROFILER_end(PROFILE_WAITING);
   return num_events;
 }