tune boid to boid buffer size and threshold

Makefile

@@ -14,10 +14,11 @@ WEB_HEAP_SIZE := 335544320
 WEB_STACK_SIZE := 196608
 WEB_SHELL := src/shell.html
 
-COMPILER_FLAGS := -std=c++17 -Wno-enum-compare -O3 -g -flto -msse4.2 -mavx
+DEBUG_MODE := no
+
+COMPILER_FLAGS := -std=c++17 -Wno-enum-compare -O3 -flto -msse4.2 -mavx
 COMPILER_FLAGS += -DRPROF_IMPLEMENTATION
 COMPILER_FLAGS += -DRAYGUI_IMPLEMENTATION
-# COMPILER_FLAGS += -DRLGL_IMPLEMENTATION
 LINKER_FLAGS   := -lraylib
 
 # SOURCES := $(wildcard src/*.cpp)
@@ -25,6 +26,12 @@ SOURCES := src/main.cpp
 
 COMPILER_FLAGS += -I$(SUBMODULES_PATH)/raygui/src/
 
+ifeq ($(DEBUG_MODE), yes)
+	COMPILER_FLAGS += -DDEBUG -g
+else
+	COMPILER_FLAGS += -DRPROF_STUB_OUT
+endif
+
 # ----------------- Prepare variables for targets ------------------
 
 EXT :=

compile_flags.txt

@@ -4,3 +4,5 @@
 -DRLGL_IMPLEMENTATION
 -DRPROF_IMPLEMENTATION
 -DRAYGUI_IMPLEMENTATION
+-DDEBUG
+-DSIMD256

src/boid-playground.hpp

@@ -9,7 +9,12 @@
 
 #define ARRAY_LEN(arr) (sizeof(arr)/sizeof(arr[0]))
 #define LogTrace(...) TraceLog(LOG_TRACE, __VA_ARGS__)
+
+#ifdef DEBUG
 #define DEBUG_ASSERT(...) assert(__VA_ARGS__)
+#else
+#define DEBUG_ASSERT(...)
+#endif
 
 typedef uint16_t uboid_t;
 #define MAX_BOIDS (1 << (sizeof(uboid_t)*8))

src/main.cpp

@@ -7,7 +7,6 @@
     #include <emscripten/emscripten.h>
 #endif
 
-// #define RPROF_STUB_OUT
 // #define RPROF_ONLY_TOTAL_TIME
 #include "rprof.h"
 
@@ -33,8 +32,8 @@ void UpdateDrawFrame();
 static void profiling_test();
 
 int main() {
-	profiling_test();
-	return 0;
+	// profiling_test();
+	// return 0;
 
 	SetTraceLogLevel(LOG_TRACE);
 

src/world.cpp

@@ -170,6 +170,11 @@ static void chunkgrid_init(MemoryArena *arena, ChunkGrid *grid, int width, int h
 }
 
 static int g_prof_interactions = 0;
+#ifdef DEBUG
+#define INCREMENT_INTERACTIONS() g_prof_interactions++;
+#else
+#define INCREMENT_INTERACTIONS()
+#endif
 
 static int nearest_multiple(int num, int divisor) {
 	return (num / divisor + (num % divisor > 0 ? 1 : 0)) * divisor;
@@ -181,7 +186,7 @@ struct boid_pair {
 };
 
 #ifdef SIMD256
-#define GET_F32_CHUNK_FROM_BOIDS(i, SIDE, FIELD)                                                         \
+#define GET_F32_CHUNK_FROM_BOIDS(i, SIDE, FIELD)                                                             \
 		_mm256_set_ps(                                                                                       \
 			boids[b2b_cmps[8*i+7].SIDE].FIELD,                                                               \
 			boids[b2b_cmps[8*i+6].SIDE].FIELD,                                                               \
@@ -193,7 +198,7 @@ struct boid_pair {
 			boids[b2b_cmps[8*i+0].SIDE].FIELD                                                                \
 		)
 #else
-#define GET_F32_CHUNK_FROM_BOIDS(i, SIDE, FIELD)                                                         \
+#define GET_F32_CHUNK_FROM_BOIDS(i, SIDE, FIELD)                                                             \
 		_mm_set_ps(                                                                                          \
 			boids[b2b_cmps[4*i+3].SIDE].FIELD,                                                               \
 			boids[b2b_cmps[4*i+2].SIDE].FIELD,                                                               \
@@ -220,6 +225,10 @@ static void world_calc_distances_and_angles(World *world, BoidList *local_boids,
 	int32_t *do_append_mask2_f32 = (int32_t*)arena_malloc(&world->frame_arena, sizeof(int32_t)*SIMD_32B_LANES, 32);
 
 	int simd_iteration_count = nearest_multiple(*b2b_cmps_count, SIMD_32B_LANES)/SIMD_32B_LANES;
+	for (int i = *b2b_cmps_count; i < simd_iteration_count*8; i++) {
+		b2b_cmps[i] = { 0 };
+	}
+
 	for (int i = 0; i < simd_iteration_count; i++) {
 		__simd from_pos_x = GET_F32_CHUNK_FROM_BOIDS(i, from, pos.x);
 		__simd from_pos_y = GET_F32_CHUNK_FROM_BOIDS(i, from, pos.y);
@@ -267,11 +276,11 @@ static void world_calc_distances_and_angles(World *world, BoidList *local_boids,
 			uboid_t to_boid   = b2b_cmps[cmp_idx].to;
 			if (do_append_mask1_f32[j]) {
 				boid_list_append(&world->frame_arena, &local_boids[from_boid], to_boid);
-				g_prof_interactions++;
+				INCREMENT_INTERACTIONS();
 			}
 			if (do_append_mask2_f32[j]) {
 				boid_list_append(&world->frame_arena, &local_boids[to_boid], from_boid);
-				g_prof_interactions++;
+				INCREMENT_INTERACTIONS();
 			}
 		}
 	}
@@ -280,18 +289,24 @@ static void world_calc_distances_and_angles(World *world, BoidList *local_boids,
 	RPROF_STOP();
 }
 
+#define B2B_CAPACITY 1024*10
+#define B2B_THRESHOLD B2B_CAPACITY*0.25
+
+static inline void append_b2b_cmp(World *world, BoidList *local_boids, boid_pair *b2b_cmps, int *b2b_cmps_count, boid_pair b2b_cmp) {
+	if ((*b2b_cmps_count) == B2B_CAPACITY) {
+		world_calc_distances_and_angles(world, local_boids, b2b_cmps, b2b_cmps_count);
+	}
+
+	b2b_cmps[(*b2b_cmps_count)++] = b2b_cmp;
+}
+
 static void world_compute_local_boids(BoidList *local_boids, World *world, ChunkGrid *chunks) {
 	Boid *boids = world->boids.data();
 	int boid_count = world->boids.size();
 	MemoryArena *arena = &world->frame_arena;
 
-	int b2b_padding = 8;
-	int b2b_capacity = 2048*10;
-	boid_pair b2b_cmps[b2b_capacity + b2b_padding];
+	boid_pair b2b_cmps[B2B_CAPACITY + SIMD_32B_LANES];
 	int b2b_cmps_count = 0;
-	for (int i = 0; i < b2b_padding; i++) {
-		memset(&b2b_cmps[b2b_capacity + i], 0, sizeof(boid_pair));
-	}
 
 	RPROF_START("Move chunk data to static arrays");
 	uboid_t *static_chunks[chunks->width * chunks->height];
@@ -319,12 +334,11 @@ static void world_compute_local_boids(BoidList *local_boids, World *world, Chunk
 				uboid_t to_boids_count = chunk->count-i-1;
 
 				for (int j = 0; j < to_boids_count; j++) {
-					// TODO:
-					// DEBUG_ASSERT(b2b_cmps_count < b2b_capacity-1);
-					b2b_cmps[b2b_cmps_count++] = {
+					boid_pair b2b_cmp = {
 						.from = from_boid,
 						.to = to_boids[j]
 					};
+					append_b2b_cmp(world, local_boids, b2b_cmps, &b2b_cmps_count, b2b_cmp);
 				}
 			}
 
@@ -342,17 +356,16 @@ static void world_compute_local_boids(BoidList *local_boids, World *world, Chunk
 				uboid_t *neighbour_boids = static_chunks[neighbour_idx];
 				for (int i = 0; i < chunk->count; i++) {
 					for (int j = 0; j < neighbour_chunk->count; j++) {
-						// TODO:
-						// DEBUG_ASSERT(b2b_cmps_count < b2b_capacity-1);
-						b2b_cmps[b2b_cmps_count++] = {
+						boid_pair b2b_cmp = {
 							.from = chunk_boids[i],
 							.to = neighbour_boids[j]
 						};
+						append_b2b_cmp(world, local_boids, b2b_cmps, &b2b_cmps_count, b2b_cmp);
 					}
 				}
 			}
 
-			if (b2b_cmps_count > 2048*3) {
+			if (b2b_cmps_count > B2B_THRESHOLD) {
 				world_calc_distances_and_angles(world, local_boids, b2b_cmps, &b2b_cmps_count);
 			}
 		}