16-byte alignment

src/k3.c

@@ -170,7 +170,7 @@ struct k3Mdl *k3MdlCreate(size_t verts, size_t indices, size_t boneCount, vec3 *
 	ret->anims = NULL;
 	
 	if(k3IsSoftSkinning && pos && boneids && boneweights) {
-		ret->cpuSkinning.pos = malloc(sizeof(*ret->cpuSkinning.pos) * verts);
+		ret->cpuSkinning.pos = _mm_malloc(sizeof(*ret->cpuSkinning.pos) * verts, 16);
 		memcpy(ret->cpuSkinning.pos, pos, sizeof(*ret->cpuSkinning.pos) * verts);
 		
 		ret->cpuSkinning.boneids = malloc(sizeof(*ret->cpuSkinning.boneids) * verts * 4);
@@ -342,7 +342,7 @@ struct k3Mdl *k3MdlCopySubs(struct k3Mdl *src) {
 	memcpy(dst->meshes, src->meshes, sizeof(*dst->meshes) * dst->meshCount);
 	
 	dst->boneCount = src->boneCount;
-	dst->invBind = malloc(sizeof(*dst->invBind) * dst->boneCount);
+	dst->invBind = _mm_malloc(sizeof(*dst->invBind) * dst->boneCount, 16);
 	memcpy(dst->invBind, src->invBind, sizeof(*dst->invBind) * dst->boneCount);
 	dst->boneParents = malloc(sizeof(*dst->boneParents) * dst->boneCount);
 	memcpy(dst->boneParents, src->boneParents, sizeof(*dst->boneParents) * dst->boneCount);

src/k3particles.c

@@ -67,10 +67,10 @@ static void random_cone_vector(float coneAngle, vec3 output) {
 }
 
 static void regenerate_model(struct k3CPUQuadParticles *this, vec3 cameraRight, vec3 cameraUp, vec3 cameraFront) {
-	vec3 *vpos = calloc(sizeof(*vpos), 4 * this->capacity);
+	vec3 *vpos = _mm_malloc(sizeof(*vpos) * 4 * this->capacity, 16);
-	uint8_t *vcols = calloc(4, 4 * this->capacity);
+	uint8_t *vcols = _mm_malloc(4 * 4 * this->capacity, 16);
-	vec2 *vuvs = calloc(sizeof(*vuvs), 4 * this->capacity);
+	vec2 *vuvs = _mm_malloc(sizeof(*vuvs) * 4 * this->capacity, 16);
-	uint8_t *vnrms = calloc(3, 4 * this->capacity);
+	uint8_t *vnrms = _mm_malloc(3 * 4 * this->capacity, 16);
 	
 	vec3 halfRight, halfUp;
 	glm_vec3_scale(cameraRight, 0.5, halfRight);
@@ -128,10 +128,10 @@ static void regenerate_model(struct k3CPUQuadParticles *this, vec3 cameraRight,
 	
 	this->mdl->meshes[0].idxNumber = this->count * 6;
 	
-	free(vpos);
+	_mm_free(vpos);
-	free(vcols);
+	_mm_free(vcols);
-	free(vuvs);
+	_mm_free(vuvs);
-	free(vnrms);
+	_mm_free(vnrms);
 }
 
 static void copy_particle(struct k3CPUQuadParticles *this, size_t from, size_t to) {

src/k3particles.h

@@ -5,6 +5,19 @@
 #include"k3.h"
 
 struct k3CPUQuadParticles {
+	vec4 colorStart;
+	vec4 colorEnd;
+	
+	vec3 origin;
+	vec3 gravity;
+	vec3 emissionConeDirection;
+	float emissionRate;
+	float emissionConeAngle;
+	float particleLifetime;
+	
+	bool emissionEnabled;
+	float emissionLifetime;
+	
 	size_t capacity;
 	size_t count;
 	vec3 *positions;
@@ -13,19 +26,6 @@ struct k3CPUQuadParticles {
 	float *lifetimes;
 	
 	struct k3Mdl *mdl;
-	
-	vec4 colorStart;
-	vec4 colorEnd;
-	
-	vec3 origin;
-	vec3 gravity;
-	float emissionRate;
-	vec3 emissionConeDirection;
-	float emissionConeAngle;
-	float particleLifetime;
-	
-	bool emissionEnabled;
-	float emissionLifetime;
 };
 
 void k3CPUQuadParticlesInit(struct k3CPUQuadParticles*, struct k3Mat*);