Add threading

Makefile

@@ -1,6 +1,6 @@
 CC            = clang
 CFLAGS        = -g -O3 -Iraylib/include -Iwapp/src
-LDFLAGS       = '-Wl,-rpath,$$ORIGIN/raylib/lib' -Lraylib/lib -lraylib -lm
+LDFLAGS       = '-Wl,-rpath,$$ORIGIN/raylib/lib' -Lraylib/lib -lraylib -lm -lpthread
 BASEDIR       = $(shell realpath ./)
 RL_SRCDIR     = ${BASEDIR}/raylib-src/src
 RL_BUILDDIR   = ${BASEDIR}/raylib-build

dod.c

@@ -6,6 +6,9 @@
 #include <stdbool.h>
 #include <assert.h>
 #include <string.h>
+#include <pthread.h>
+
+#define THREAD_COUNT 4
 
 #define MOVABLE_TAG_SHIFT     0
 #define RENDERABLE_TAG_SHIFT  1
@@ -57,6 +60,17 @@ struct Manager {
   u64      count;
 };
 
+typedef struct PositionThreadArgs PositionThreadArgs;
+struct PositionThreadArgs {
+  const Rect *zones;
+  Rect       *rects;
+  Velocity   *velocities;
+  u8         *tags;
+  u64        count;
+  f32        delta;
+};
+
+typedef void *(*PThreadRoutine)(void *);
 typedef void (*ScaleInitialiser)(Scale *scale, XOR256State *state);
 typedef void (*VelocityInitialiser)(Velocity *velocity, XOR256State *state);
 typedef void (*RaylibDrawRectFunc)(int posX, int posY, int width, int height, Color color);
@@ -68,6 +82,7 @@ void init_scale_zone(Scale *scale, XOR256State *state);
 void init_velocity(Velocity *velocity, XOR256State *state);
 void zero_velocity(Velocity *velocity, XOR256State *state);
 void update_positions(u8 *tags, Rect *rects, Velocity *velocities, u64 count, f32 delta);
+void *update_position_thread(PositionThreadArgs *args);
 void render_entities(const u8 *tags, const Rect *rects, u64 count);
 u8 collides(const Rect *rect, const Rect *collider);
 i16 get_random_velocity(XOR256State *state);
@@ -103,7 +118,6 @@ int main(void) {
 
   f32 last_time = GetFrameTime();
   f32 delta, cur_time;
-  u64 count = 0;
 
   while (!WindowShouldClose()) {
     f64 time = GetTime();
@@ -111,8 +125,6 @@ int main(void) {
       break;
     }
 
-    ++count;
-
     cur_time = GetFrameTime();
     delta    = cur_time - last_time;
 
@@ -133,8 +145,6 @@ int main(void) {
 
   CloseWindow();
 
-  printf("%lu\n", count);
-
   return 0;
 }
 
@@ -189,6 +199,46 @@ void zero_velocity(Velocity *velocity, XOR256State *state) {
 }
 
 void update_positions(u8 *tags, Rect *rects, Velocity *velocities, u64 count, f32 delta) {
+  persistent pthread_t threads[THREAD_COUNT]       = {0};
+  persistent PositionThreadArgs args[THREAD_COUNT] = {0};
+
+  u64 total_entities_count  = count - ZONE_COUNT;
+  u64 thread_entities_count = (u64)(ceil((f64)total_entities_count / THREAD_COUNT));
+
+  i32 result;
+  u64 start = ZONE_COUNT;
+  u64 end, args_count;
+
+  for (u64 i = 0; i < THREAD_COUNT; ++i) {
+    if (total_entities_count > thread_entities_count) {
+      end = start + thread_entities_count;
+      total_entities_count -= thread_entities_count;
+    } else {
+      end = start + total_entities_count;
+      total_entities_count = 0;
+    }
+
+    args_count = end - start;
+
+    args[i].zones      = &(rects[0]);
+    args[i].tags       = &(tags[start]);
+    args[i].rects      = &(rects[start]);
+    args[i].velocities = &(velocities[start]);
+    args[i].count      = args_count;
+    args[i].delta      = delta;
+
+    start += args_count;
+
+    result = pthread_create(&(threads[i]), NULL, (PThreadRoutine)update_position_thread, (void *)&(args[i]));
+    assert(result == 0);
+  }
+
+  for (u64 i = 0; i < THREAD_COUNT; ++i) {
+    pthread_join(threads[i], NULL);
+  }
+}
+
+void *update_position_thread(PositionThreadArgs *args) {
   persistent f32 multipliers[2] = {1.0f, 0.5f};
   persistent u64 inside_zone_mask = 0x7;
 
@@ -196,33 +246,35 @@ void update_positions(u8 *tags, Rect *rects, Velocity *velocities, u64 count, f3
   f32 pos_x, pos_y;
   f32 max_x, max_y;
 
-  for (u64 i = ZONE_COUNT; i < count; ++i) {
-    tags[i] &= inside_zone_mask;
+  for (u64 i = 0; i < args->count; ++i) {
+    args->tags[i] &= inside_zone_mask;
     for (u64 j = 0; j < ZONE_COUNT; ++j) {
-      tags[i] |= collides(&rects[i], &rects[j]) << INSIDE_ZONE_TAG_SHIFT;
+      args->tags[i] |= collides(&args->rects[i], &args->zones[j]) << INSIDE_ZONE_TAG_SHIFT;
     }
 
-    index = (tags[i] & ENTITY_TAG_INSIDE_ZONE) >> INSIDE_ZONE_TAG_SHIFT;
+    index = (args->tags[i] & ENTITY_TAG_INSIDE_ZONE) >> INSIDE_ZONE_TAG_SHIFT;
 
-    max_x = WIDTH - rects[i].scale.width;
-    max_y = HEIGHT - rects[i].scale.height;
+    max_x = WIDTH - args->rects[i].scale.width;
+    max_y = HEIGHT - args->rects[i].scale.height;
 
-    pos_x = rects[i].position.x + velocities[i].x * multipliers[index] * delta;
-    pos_y = rects[i].position.y + velocities[i].y * multipliers[index] * delta;
+    pos_x = args->rects[i].position.x + args->velocities[i].x * multipliers[index] * args->delta;
+    pos_y = args->rects[i].position.y + args->velocities[i].y * multipliers[index] * args->delta;
 
     if (pos_x < 0 || pos_x >= max_x) {
       pos_x = min(max(pos_x, 0), max_x);
-      velocities[i].x *= -1;
+      args->velocities[i].x *= -1;
     }
 
     if (pos_y < 0 || pos_y >= max_y) {
       pos_y = min(max(pos_y, 0), max_y);
-      velocities[i].y *= -1;
+      args->velocities[i].y *= -1;
     }
 
-    rects[i].position.x = roundf(pos_x);
-    rects[i].position.y = roundf(pos_y);
+    args->rects[i].position.x = roundf(pos_x);
+    args->rects[i].position.y = roundf(pos_y);
   }
+
+  return args;
 }
 
 void render_entities(const u8 *tags, const Rect *rects, u64 count) {

run_perf

@@ -25,5 +25,5 @@ major-faults,\
 minor-faults,\
 task-clock"
 
-# perf stat -e $EVENTS -r $REPEATS -o no_dod_stats ./no_dod
+perf stat -e $EVENTS -r $REPEATS -o no_dod_stats ./no_dod
 perf stat -e $EVENTS -r $REPEATS -o dod_stats ./dod