Merge branch 'master' into migration-successful

.gitmodules

@@ -1,3 +0,0 @@
-[submodule "src/main/c/lib/stb"]
-	path = src/main/c/lib/stb
-	url = https://github.com/nothings/stb.git

README.md

@@ -1,5 +0,0 @@
-
-
-
-dependencies:
- - pthread

src/main/c/chunkmask.c

@@ -1,8 +1,7 @@
 #include <jni.h>
 #include <stdint.h>
-#include "../includes/libfluidsim.h"
-#include "../includes/utilities.h"
-#include "../includes/chunkmask.h"
+#include "includes/utilities.h"
+#include "includes/chunkmask.h"
 
 uint32_t matrix_transform(JNIEnv * env, jobjectArray jrx);
 

src/main/c/compile.sh

@@ -41,29 +41,24 @@ rm -f ./*.dll
 
 
 #compile object files
-COMPILE_FLAGS="-c -fPIC -m64 -mavx -mavx2 -O1"
-INPUT_FILES="./src/densitystep.c"
-OUTPUT_FILE="./densitystep.o"
-gcc $COMPILE_FLAGS -I"$BASE_INCLUDE_DIR" -I"$OS_INCLUDE_DIR" $INPUT_FILES -o $OUTPUT_FILE
+# COMPILE_FLAGS="-c -fPIC -m64 -mavx -mavx2 -march=native -Ofast -msse -msse2 -msse3 -mmmx -m3dnow"
+# INPUT_FILES="./densitystep.c"
+# OUTPUT_FILE="./densitystep.o"
+# gcc $COMPILE_FLAGS -I"$BASE_INCLUDE_DIR" -I"$OS_INCLUDE_DIR" $INPUT_FILES -o $OUTPUT_FILE
 
-COMPILE_FLAGS="-c -fPIC -m64 -mavx -mavx2 -O1"
-INPUT_FILES="./src/velocitystep.c"
-OUTPUT_FILE="./velocitystep.o"
-gcc $COMPILE_FLAGS -I"$BASE_INCLUDE_DIR" -I"$OS_INCLUDE_DIR" $INPUT_FILES -o $OUTPUT_FILE
+# COMPILE_FLAGS="-c -fPIC -m64 -mavx -mavx2 -march=native -Ofast -msse -msse2 -msse3 -mmmx -m3dnow"
+# INPUT_FILES="./velocitystep.c"
+# OUTPUT_FILE="./velocitystep.o"
+# gcc $COMPILE_FLAGS -I"$BASE_INCLUDE_DIR" -I"$OS_INCLUDE_DIR" $INPUT_FILES -o $OUTPUT_FILE
 
-COMPILE_FLAGS="-c -fPIC -m64 -mavx -mavx2 -O1"
-INPUT_FILES="./src/chunkmask.c"
-OUTPUT_FILE="./chunkmask.o"
-gcc $COMPILE_FLAGS -I"$BASE_INCLUDE_DIR" -I"$OS_INCLUDE_DIR" $INPUT_FILES -o $OUTPUT_FILE
+# COMPILE_FLAGS="-c -fPIC -m64 -mavx -mavx2 -march=native -Ofast -msse -msse2 -msse3 -mmmx -m3dnow"
+# INPUT_FILES="./chunkmask.c"
+# OUTPUT_FILE="./chunkmask.o"
+# gcc $COMPILE_FLAGS -I"$BASE_INCLUDE_DIR" -I"$OS_INCLUDE_DIR" $INPUT_FILES -o $OUTPUT_FILE
 
-COMPILE_FLAGS="-c -fPIC -m64 -mavx -mavx2 -O1"
-INPUT_FILES="./src/threadpool.c"
-OUTPUT_FILE="./threadpool.o"
-gcc $COMPILE_FLAGS -I"$BASE_INCLUDE_DIR" -I"$OS_INCLUDE_DIR" $INPUT_FILES -o $OUTPUT_FILE
-
-COMPILE_FLAGS="-c -fPIC -m64 -mavx -mavx2 -O1"
-INPUT_FILES="./src/libfluidsim.c"
-OUTPUT_FILE="./libfluidsim.o"
+COMPILE_FLAGS="-c -fPIC -m64 -mavx -mavx2 -march=native -Ofast -msse -msse2 -msse3 -mmmx -m3dnow"
+INPUT_FILES="./fluidsim.c"
+OUTPUT_FILE="./fluidsim.o"
 gcc $COMPILE_FLAGS -I"$BASE_INCLUDE_DIR" -I"$OS_INCLUDE_DIR" $INPUT_FILES -o $OUTPUT_FILE
 
 
@@ -71,7 +66,7 @@ gcc $COMPILE_FLAGS -I"$BASE_INCLUDE_DIR" -I"$OS_INCLUDE_DIR" $INPUT_FILES -o $OU
 #compile shared object file
 OUTPUT_FILE="libfluidsim$LIB_ENDING"
 COMPILE_FLAGS="-shared"
-INPUT_FILES="densitystep.o velocitystep.o chunkmask.o threadpool.o libfluidsim.o"
+INPUT_FILES="fluidsim.o"
 gcc $COMPILE_FLAGS $INPUT_FILES -o $OUTPUT_FILE
 
 #move to resources

src/main/c/densitystep.c

@@ -2,39 +2,54 @@
 #include <immintrin.h>
 #include <stdint.h>
 
-#include "../includes/libfluidsim.h"
-#include "../includes/utilities.h"
-#include "../includes/chunkmask.h"
+#include "includes/utilities.h"
+#include "includes/chunkmask.h"
 
+void advectDensity(uint32_t chunk_mask, int N, int b, float ** d, float ** d0, float * u, float * v, float * w, float dt);
 
-void addDensity(
+/*
+ * Class:     electrosphere_FluidSim
+ * Method:    addDensity
+ * Signature: (II[Ljava/nio/ByteBuffer;Ljava/nio/ByteBuffer;F)V
+ */
+void Java_electrosphere_FluidSim_addDensity
+  (
   int N,
   int chunk_mask,
-  float * x,
-  float * s,
+  float ** d,
+  float ** d0,
   float dt){
     int i;
     int size=N*N*N;
+    float * x = GET_ARR_RAW(env,d,CENTER_LOC);
+    float * s = GET_ARR_RAW(env,d0,CENTER_LOC);
     for(i=0; i<size; i++){
         x[i] += dt*s[i];
     }
 }
 
-
-void diffuseDensity(
+/*
+ * Class:     electrosphere_FluidSim
+ * Method:    solveDiffuseDensity
+ * Signature: (II[Ljava/nio/ByteBuffer;Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;FFF)V
+ */
+void Java_electrosphere_FluidSim_solveDiffuseDensity
+  (
   int N,
   int chunk_mask,
-  float * jrx,
-  float * jrx0,
+  float ** d,
+  float ** d0,
+  float ** jru,
+  float ** jrv,
+  float ** jrw,
   float DIFFUSION_CONST,
   float VISCOSITY_CONST,
   float dt){
     float a=dt*DIFFUSION_CONST*N*N*N;
     float c=1+6*a;
     int i, j, k, l, m;
-    float * x = jrx;
-    float * x0 = jrx0;
-
+    float * x = GET_ARR_RAW(env,d,CENTER_LOC);
+    float * x0 = GET_ARR_RAW(env,d0,CENTER_LOC);
     
     __m256 aScalar = _mm256_set1_ps(a);
     __m256 cScalar = _mm256_set1_ps(c);
@@ -66,21 +81,41 @@ void diffuseDensity(
     }
 }
 
-void advectDensity(uint32_t chunk_mask, int N, float * jrd, float * jrd0, float * u, float * v, float * w, float dt){
+/*
+ * Class:     electrosphere_FluidSim
+ * Method:    advectDensity
+ * Signature: (II[Ljava/nio/ByteBuffer;Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;FFF)V
+ */
+void Java_electrosphere_FluidSim_advectDensity
+  (
+  int N,
+  int chunk_mask,
+  float ** d,
+  float ** d0,
+  float ** jru,
+  float ** jrv,
+  float ** jrw,
+  float DIFFUSION_CONST,
+  float VISCOSITY_CONST,
+  float dt){
+    advectDensity(chunk_mask,N,3,d,d0,GET_ARR_RAW(env,jru,CENTER_LOC),GET_ARR_RAW(env,jrv,CENTER_LOC),GET_ARR_RAW(env,jrw,CENTER_LOC),dt);
+}
+
+void advectDensity(uint32_t chunk_mask, int N, int b, float ** d, float ** d0, float * u, float * v, float * w, float dt){
     int i, j, k, i0, j0, k0, i1, j1, k1;
     int m,n,o;
     float x, y, z, s0, t0, s1, t1, u1, u0, dtx,dty,dtz;
     
     dtx=dty=dtz=dt*N;
 
-    float * d = jrd;
+    float * center_d = GET_ARR_RAW(env,d,CENTER_LOC);
 
-    float * d0 = jrd0;
+    float * center_d0 = GET_ARR_RAW(env,d0,CENTER_LOC);
 
     for(k=1; k<N-1; k++){
         for(j=1; j<N-1; j++){
             for(i=1; i<N-1; i++){
-                d0 = jrd0;
+                center_d0 = GET_ARR_RAW(env,d0,CENTER_LOC);
                 //calculate location to pull from
                 x = i-dtx*u[IX(i,j,k)];
                 y = j-dty*v[IX(i,j,k)];
@@ -88,12 +123,12 @@ void advectDensity(uint32_t chunk_mask, int N, float * jrd, float * jrd0, float
 
                 m = n = o = 1;
 
-                // if(x < 1){ m -= 1; }
-                // if(x >= N-1){ m += 1; }
-                // if(y < 1){ n -= 1; }
-                // if(y >= N-1){ n += 1; }
-                // if(z < 1){ o -= 1; }
-                // if(z >= N-1){ o += 1; }
+                if(x < 1){ m -= 1; }
+                if(x >= N-1){ m += 1; }
+                if(y < 1){ n -= 1; }
+                if(y >= N-1){ n += 1; }
+                if(z < 1){ o -= 1; }
+                if(z >= N-1){ o += 1; }
 
                 //If the out of bounds coordinate is in bounds for a neighbor chunk, use that chunk as source instead
                 // if(CK(m,n,o) != CENTER_LOC){
@@ -195,18 +230,18 @@ void advectDensity(uint32_t chunk_mask, int N, float * jrd, float * jrd0, float
                 // if(k1 < 0){
                 //     k1 = 0;
                 // }
-                d[IX(i,j,k)] = 
+                center_d[IX(i,j,k)] = 
                 s0*(
-                    t0*u0*d0[IX(i0,j0,k0)]+
-                    t1*u0*d0[IX(i0,j1,k0)]+
-                    t0*u1*d0[IX(i0,j0,k1)]+
-                    t1*u1*d0[IX(i0,j1,k1)]
+                    t0*u0*center_d0[IX(i0,j0,k0)]+
+                    t1*u0*center_d0[IX(i0,j1,k0)]+
+                    t0*u1*center_d0[IX(i0,j0,k1)]+
+                    t1*u1*center_d0[IX(i0,j1,k1)]
                 )+
                 s1*(
-                    t0*u0*d0[IX(i1,j0,k0)]+
-                    t1*u0*d0[IX(i1,j1,k0)]+
-                    t0*u1*d0[IX(i1,j0,k1)]+
-                    t1*u1*d0[IX(i1,j1,k1)]
+                    t0*u0*center_d0[IX(i1,j0,k0)]+
+                    t1*u0*center_d0[IX(i1,j1,k0)]+
+                    t0*u1*center_d0[IX(i1,j0,k1)]+
+                    t1*u1*center_d0[IX(i1,j1,k1)]
                 );
             }
         }

src/main/c/fluidsim.c

@@ -0,0 +1,778 @@
+#include <jni.h>
+#include <stdint.h>
+#include "includes/utilities.h"
+#include "includes/chunkmask.h"
+#include "includes/electrosphere_FluidSim.h"
+#include "includes/mainFunctions.h"
+
+#include "./chunkmask.c"
+#include "./velocitystep.c"
+#include "./densitystep.c"
+
+//include stb ds
+#define STB_DS_IMPLEMENTATION
+#include "./lib/stb/stb_ds.h"
+
+#define DIM 18
+#define LINEARSOLVERTIMES 20
+
+#define DIFFUSION_CONSTANT 0.000001
+#define VISCOSITY_CONSTANT 0.000001
+
+/**
+ * A chunk
+*/
+typedef struct {
+    jobject jchunk;
+    float * d[27];
+    float * d0[27];
+    float * u[27];
+    float * v[27];
+    float * w[27];
+    float * u0[27];
+    float * v0[27];
+    float * w0[27];
+    jobjectArray jd;
+    jobjectArray jd0;
+    jobjectArray ju;
+    jobjectArray jv;
+    jobjectArray jw;
+    jobjectArray ju0;
+    jobjectArray jv0;
+    jobjectArray jw0;
+    int chunkMask;
+} Chunk;
+
+//all chunks
+Chunk ** chunks = NULL;
+
+#define getChunk(i) (*env)->CallObjectMethod(env,chunkList,jListGet,i)
+#define getBuffArr(buffId) (*env)->GetObjectField(env,chunkJRaw,buffId)
+#define setBuffArr(buffId,value) (*env)->SetObjectField(env,chunkJRaw,buffId,value)
+
+//jni help:
+//https://stackoverflow.com/questions/39823375/clarification-about-getfieldid
+
+JNIEXPORT void JNICALL Java_electrosphere_FluidSim_simulate(
+    JNIEnv * env,
+    jclass class,
+    jobject chunkList,
+    jfloat timestep
+){
+    jclass listClass = (*env)->FindClass(env,"java/util/List");
+    jclass fluidSimClass = (*env)->FindClass(env,"electrosphere/FluidSim");
+    //JNIEnv *env, jclass clazz, const char *name, const char *sig
+    jmethodID jListSize = (*env)->GetMethodID(env, listClass, "size", "()I");
+    jmethodID jListGet  = (*env)->GetMethodID(env, listClass, "get", "(I)Ljava/lang/Object;");
+    jmethodID jListAdd  = (*env)->GetMethodID(env, listClass, "add", "(Ljava/lang/Object;)Z");
+
+    //ByteBuffer[]
+    jfieldID dJId = (*env)->GetFieldID(env,fluidSimClass,"density","[Ljava/nio/ByteBuffer;");
+    jfieldID d0JId = (*env)->GetFieldID(env,fluidSimClass,"densityAddition","[Ljava/nio/ByteBuffer;");
+    jfieldID uJId = (*env)->GetFieldID(env,fluidSimClass,"uVector","[Ljava/nio/ByteBuffer;");
+    jfieldID vJId = (*env)->GetFieldID(env,fluidSimClass,"vVector","[Ljava/nio/ByteBuffer;");
+    jfieldID wJId = (*env)->GetFieldID(env,fluidSimClass,"wVector","[Ljava/nio/ByteBuffer;");
+    jfieldID u0JId = (*env)->GetFieldID(env,fluidSimClass,"uAdditionVector","[Ljava/nio/ByteBuffer;");
+    jfieldID v0JId = (*env)->GetFieldID(env,fluidSimClass,"vAdditionVector","[Ljava/nio/ByteBuffer;");
+    jfieldID w0JId = (*env)->GetFieldID(env,fluidSimClass,"wAdditionVector","[Ljava/nio/ByteBuffer;");
+    jfieldID chunkmaskJId = (*env)->GetFieldID(env,fluidSimClass,"chunkMask","I");
+
+    //the number of chunks
+    int numChunks = (*env)->CallIntMethod(env,chunkList,jListSize);
+
+    //current chunk (this)
+    jobject chunkJRaw;
+    //current chunk fields
+    jobjectArray jd;
+    jobjectArray jd0;
+    jobjectArray u;
+    jobjectArray v;
+    jobjectArray w;
+    jobjectArray u0;
+    jobjectArray v0;
+    jobjectArray w0;
+    int chunkMask;
+
+    //solve chunk mask
+    for(int i = 0; i < numChunks; i++){
+        chunkJRaw = getChunk(i);
+        chunkMask = Java_electrosphere_FluidSim_calculateChunkMask(env,chunkJRaw,getBuffArr(dJId));
+        (*env)->SetIntField(env,chunkJRaw,chunkmaskJId,chunkMask);
+
+        Chunk * newChunk;
+        if(i >= stbds_arrlen(chunks)){
+            // printf("allocate chunk %d\n",i);
+            // fflush(stdout);
+            newChunk = (Chunk *)malloc(sizeof(Chunk));
+            // printf("new chunk %p\n",newChunk);
+            // fflush(stdout);
+            stbds_arrput(chunks,newChunk);
+            // printf("new chunk %p\n",chunks[i]);
+            // fflush(stdout);
+        } else {
+            newChunk = chunks[i];
+            // printf("get chunk %d: %p\n",i,newChunk);
+            // fflush(stdout);
+        }
+        jd = (*env)->GetObjectField(env,chunkJRaw,dJId);
+        jd0 = (*env)->GetObjectField(env,chunkJRaw,d0JId);
+        u = (*env)->GetObjectField(env,chunkJRaw,uJId);
+        v = (*env)->GetObjectField(env,chunkJRaw,vJId);
+        w = (*env)->GetObjectField(env,chunkJRaw,wJId);
+        u0 = (*env)->GetObjectField(env,chunkJRaw,u0JId);
+        v0 = (*env)->GetObjectField(env,chunkJRaw,v0JId);
+        w0 = (*env)->GetObjectField(env,chunkJRaw,w0JId);
+        newChunk->jd = jd;
+        newChunk->jd0 = jd0;
+        newChunk->ju = u;
+        newChunk->jv = v;
+        newChunk->jw = w;
+        newChunk->ju0 = u0;
+        newChunk->jv0 = v0;
+        newChunk->jw0 = w0;
+        newChunk->jchunk = chunkJRaw;
+        newChunk->chunkMask = chunkMask;
+        for(int j = 0; j < 27; j++){
+            if((chunkMask & CHUNK_INDEX_ARR[j]) > 0){
+                newChunk->d[j] = GET_ARR(env,jd,j);
+                newChunk->d0[j] = GET_ARR(env,jd0,j);
+                newChunk->u[j] = GET_ARR(env,u,j);
+                newChunk->v[j] = GET_ARR(env,v,j);
+                newChunk->w[j] = GET_ARR(env,w,j);
+                newChunk->u0[j] = GET_ARR(env,u0,j);
+                newChunk->v0[j] = GET_ARR(env,v0,j);
+                newChunk->w0[j] = GET_ARR(env,w0,j);
+            }
+        }
+        // for(int j = 0; j < 27; j++){
+        //     newChunk.d[j] = GET_ARR(env,d,j);
+        // }
+        // chunks[i].d0 = getBuffArr(d0JId);
+        // chunks[i].u = getBuffArr(uJId);
+        // chunks[i].v = getBuffArr(vJId);
+        // chunks[i].w = getBuffArr(wJId);
+        // chunks[i].u0 = getBuffArr(u0JId);
+        // chunks[i].v0 = getBuffArr(v0JId);
+        // chunks[i].w0 = getBuffArr(w0JId);
+    }
+
+    // printf("%p\n",chunks[0].d);
+
+    //solve chunk mask
+    for(int i = 0; i < numChunks; i++){
+        Chunk * currentChunk = chunks[i];
+        chunkJRaw = currentChunk->jchunk;
+        chunkMask = currentChunk->chunkMask;
+        jd = currentChunk->jd;
+        jd0 = currentChunk->jd0;
+        u = currentChunk->ju;
+        v = currentChunk->jv;
+        w = currentChunk->jw;
+        u0 = currentChunk->ju0;
+        v0 = currentChunk->jv0;
+        w0 = currentChunk->jw0;
+        //old assignments
+        // chunkJRaw = getChunk(i);
+        // chunkMask = (*env)->GetIntField(env,chunkJRaw,chunkmaskJId);
+        // d = getBuffArr(dJId);
+        // d0 = getBuffArr(d0JId);
+        // u = getBuffArr(uJId);
+        // v = getBuffArr(vJId);
+        // w = getBuffArr(wJId);
+        // u0 = getBuffArr(u0JId);
+        // v0 = getBuffArr(v0JId);
+        // w0 = getBuffArr(w0JId);
+        Java_electrosphere_FluidSim_addSourceToVectors(
+            DIM,
+            chunkMask,
+            currentChunk->u,
+            currentChunk->v,
+            currentChunk->w,
+            currentChunk->u0,
+            currentChunk->v0,
+            currentChunk->w0,
+            DIFFUSION_CONSTANT,
+            VISCOSITY_CONSTANT,
+            timestep
+        );
+    }
+    //swap all vector fields
+    {
+        //swap vector fields
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+
+            float * tmpArr;
+            for(int j = 0; j < 27; j++){
+                tmpArr = currentChunk->u[j];
+                currentChunk->u[j] = currentChunk->u0[j];
+                currentChunk->u0[j] = tmpArr;
+            }
+            for(int j = 0; j < 27; j++){
+                tmpArr = currentChunk->v[j];
+                currentChunk->v[j] = currentChunk->v0[j];
+                currentChunk->v0[j] = tmpArr;
+            }
+            for(int j = 0; j < 27; j++){
+                tmpArr = currentChunk->w[j];
+                currentChunk->w[j] = currentChunk->w0[j];
+                currentChunk->w0[j] = tmpArr;
+            }
+        }
+        //copy neighbors
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            jd = currentChunk->jd;
+            jd0 = currentChunk->jd0;
+            u = currentChunk->ju;
+            v = currentChunk->jv;
+            w = currentChunk->jw;
+            u0 = currentChunk->ju0;
+            v0 = currentChunk->jv0;
+            w0 = currentChunk->jw0;
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v);
+            copyNeighborsRaw(DIM,chunkMask,0,3,currentChunk->w);
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u0);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v0);
+            copyNeighborsRaw(DIM,chunkMask,0,3,currentChunk->w0);
+        }
+    }
+    //solve vector diffusion
+    {
+        for(int l = 0; l < LINEARSOLVERTIMES; l++){
+            //solve vector diffusion
+            for(int i = 0; i < numChunks; i++){
+                Chunk * currentChunk = chunks[i];
+                chunkJRaw = currentChunk->jchunk;
+                chunkMask = currentChunk->chunkMask;
+                jd = currentChunk->jd;
+                jd0 = currentChunk->jd0;
+                u = currentChunk->ju;
+                v = currentChunk->jv;
+                w = currentChunk->jw;
+                u0 = currentChunk->ju0;
+                v0 = currentChunk->jv0;
+                w0 = currentChunk->jw0;
+                Java_electrosphere_FluidSim_solveVectorDiffuse(DIM,chunkMask,currentChunk->u,currentChunk->v,currentChunk->w,currentChunk->u0,currentChunk->v0,currentChunk->w0,DIFFUSION_CONSTANT,VISCOSITY_CONSTANT,timestep);
+            }
+            //update array for vectors
+            for(int i = 0; i < numChunks; i++){
+                Chunk * currentChunk = chunks[i];
+                chunkJRaw = currentChunk->jchunk;
+                chunkMask = currentChunk->chunkMask;
+                jd = currentChunk->jd;
+                jd0 = currentChunk->jd0;
+                u = currentChunk->ju;
+                v = currentChunk->jv;
+                w = currentChunk->jw;
+                u0 = currentChunk->ju0;
+                v0 = currentChunk->jv0;
+                w0 = currentChunk->jw0;
+                setBoundsToNeighborsRaw(DIM,chunkMask,1,currentChunk->u);
+                setBoundsToNeighborsRaw(DIM,chunkMask,2,currentChunk->v);
+                setBoundsToNeighborsRaw(DIM,chunkMask,3,currentChunk->w);
+                copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u);
+                copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v);
+                copyNeighborsRaw(DIM,chunkMask,0,3,currentChunk->w);
+            }
+        }
+    }
+    //solve projection
+    {
+        //update array for vectors
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            jd = currentChunk->jd;
+            jd0 = currentChunk->jd0;
+            u = currentChunk->ju;
+            v = currentChunk->jv;
+            w = currentChunk->jw;
+            u0 = currentChunk->ju0;
+            v0 = currentChunk->jv0;
+            w0 = currentChunk->jw0;
+            setBoundsToNeighborsRaw(DIM,chunkMask,1,currentChunk->u);
+            setBoundsToNeighborsRaw(DIM,chunkMask,2,currentChunk->v);
+            setBoundsToNeighborsRaw(DIM,chunkMask,3,currentChunk->w);
+            setBoundsToNeighborsRaw(DIM,chunkMask,1,currentChunk->u0);
+            setBoundsToNeighborsRaw(DIM,chunkMask,2,currentChunk->v0);
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v);
+            copyNeighborsRaw(DIM,chunkMask,0,3,currentChunk->w);
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u0);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v0);
+        }
+        //setup projection
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            jd = currentChunk->jd;
+            jd0 = currentChunk->jd0;
+            u = currentChunk->ju;
+            v = currentChunk->jv;
+            w = currentChunk->jw;
+            u0 = currentChunk->ju0;
+            v0 = currentChunk->jv0;
+            w0 = currentChunk->jw0;
+            Java_electrosphere_FluidSim_setupProjection(DIM,chunkMask,currentChunk->u,currentChunk->v,currentChunk->w,currentChunk->u0,currentChunk->v0,currentChunk->w0,DIFFUSION_CONSTANT,VISCOSITY_CONSTANT,timestep);
+        }
+        //update array for vectors
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            jd = currentChunk->jd;
+            jd0 = currentChunk->jd0;
+            u = currentChunk->ju;
+            v = currentChunk->jv;
+            w = currentChunk->jw;
+            u0 = currentChunk->ju0;
+            v0 = currentChunk->jv0;
+            w0 = currentChunk->jw0;
+            setBoundsToNeighborsRaw(DIM,chunkMask,1,currentChunk->u0);
+            setBoundsToNeighborsRaw(DIM,chunkMask,2,currentChunk->v0);
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u0);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v0);
+        }
+        //samples u0, v0
+        //sets u0
+        //these should have just been mirrored in the above
+        //
+        //Perform main projection solver
+        for(int l = 0; l < LINEARSOLVERTIMES; l++){
+            for(int i = 0; i < numChunks; i++){
+                Chunk * currentChunk = chunks[i];
+                chunkJRaw = currentChunk->jchunk;
+                chunkMask = currentChunk->chunkMask;
+                jd = currentChunk->jd;
+                jd0 = currentChunk->jd0;
+                u = currentChunk->ju;
+                v = currentChunk->jv;
+                w = currentChunk->jw;
+                u0 = currentChunk->ju0;
+                v0 = currentChunk->jv0;
+                w0 = currentChunk->jw0;
+                Java_electrosphere_FluidSim_solveProjection(DIM,chunkMask,currentChunk->u,currentChunk->v,currentChunk->w,currentChunk->u0,currentChunk->v0,currentChunk->w0,DIFFUSION_CONSTANT,VISCOSITY_CONSTANT,timestep);
+            }
+            for(int i = 0; i < numChunks; i++){
+                Chunk * currentChunk = chunks[i];
+                chunkJRaw = currentChunk->jchunk;
+                chunkMask = currentChunk->chunkMask;
+                jd = currentChunk->jd;
+                jd0 = currentChunk->jd0;
+                u = currentChunk->ju;
+                v = currentChunk->jv;
+                w = currentChunk->jw;
+                u0 = currentChunk->ju0;
+                v0 = currentChunk->jv0;
+                w0 = currentChunk->jw0;
+                setBoundsToNeighborsRaw(DIM,chunkMask,1,currentChunk->u0);
+                copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u0);
+            }
+        }
+        //samples u,v,w,u0
+        //sets u,v,w
+        //Finalize projection
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            jd = currentChunk->jd;
+            jd0 = currentChunk->jd0;
+            u = currentChunk->ju;
+            v = currentChunk->jv;
+            w = currentChunk->jw;
+            u0 = currentChunk->ju0;
+            v0 = currentChunk->jv0;
+            w0 = currentChunk->jw0;
+            Java_electrosphere_FluidSim_finalizeProjection(DIM,chunkMask,currentChunk->u,currentChunk->v,currentChunk->w,currentChunk->u0,currentChunk->v0,currentChunk->w0,DIFFUSION_CONSTANT,VISCOSITY_CONSTANT,timestep);
+        }
+        //set boundaries a final time for u,v,w
+        //...
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            jd = currentChunk->jd;
+            jd0 = currentChunk->jd0;
+            u = currentChunk->ju;
+            v = currentChunk->jv;
+            w = currentChunk->jw;
+            u0 = currentChunk->ju0;
+            v0 = currentChunk->jv0;
+            w0 = currentChunk->jw0;
+            setBoundsToNeighborsRaw(DIM,chunkMask,1,currentChunk->u);
+            setBoundsToNeighborsRaw(DIM,chunkMask,2,currentChunk->v);
+            setBoundsToNeighborsRaw(DIM,chunkMask,3,currentChunk->w);
+            setBoundsToNeighborsRaw(DIM,chunkMask,1,currentChunk->u0);
+            setBoundsToNeighborsRaw(DIM,chunkMask,2,currentChunk->v0);
+            setBoundsToNeighborsRaw(DIM,chunkMask,3,currentChunk->w0);
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v);
+            copyNeighborsRaw(DIM,chunkMask,0,3,currentChunk->w);
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u0);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v0);
+            copyNeighborsRaw(DIM,chunkMask,0,3,currentChunk->w0);
+        }
+    }
+    //swap all vector fields
+    {
+        //swap vector fields
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+
+            float * tmpArr;
+            for(int j = 0; j < 27; j++){
+                tmpArr = currentChunk->u[j];
+                currentChunk->u[j] = currentChunk->u0[j];
+                currentChunk->u0[j] = tmpArr;
+            }
+            for(int j = 0; j < 27; j++){
+                tmpArr = currentChunk->v[j];
+                currentChunk->v[j] = currentChunk->v0[j];
+                currentChunk->v0[j] = tmpArr;
+            }
+            for(int j = 0; j < 27; j++){
+                tmpArr = currentChunk->w[j];
+                currentChunk->w[j] = currentChunk->w0[j];
+                currentChunk->w0[j] = tmpArr;
+            }
+        }
+        //copy neighbors
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            jd = currentChunk->jd;
+            jd0 = currentChunk->jd0;
+            u = currentChunk->ju;
+            v = currentChunk->jv;
+            w = currentChunk->jw;
+            u0 = currentChunk->ju0;
+            v0 = currentChunk->jv0;
+            w0 = currentChunk->jw0;
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v);
+            copyNeighborsRaw(DIM,chunkMask,0,3,currentChunk->w);
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u0);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v0);
+            copyNeighborsRaw(DIM,chunkMask,0,3,currentChunk->w0);
+        }
+    }
+    //advect vectors across boundaries
+    {
+        //update border arrs
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            jd = currentChunk->jd;
+            jd0 = currentChunk->jd0;
+            u = currentChunk->ju;
+            v = currentChunk->jv;
+            w = currentChunk->jw;
+            u0 = currentChunk->ju0;
+            v0 = currentChunk->jv0;
+            w0 = currentChunk->jw0;
+            setBoundsToNeighborsRaw(DIM,chunkMask,1,currentChunk->u);
+            setBoundsToNeighborsRaw(DIM,chunkMask,2,currentChunk->v);
+            setBoundsToNeighborsRaw(DIM,chunkMask,3,currentChunk->w);
+            setBoundsToNeighborsRaw(DIM,chunkMask,1,currentChunk->u0);
+            setBoundsToNeighborsRaw(DIM,chunkMask,2,currentChunk->v0);
+            setBoundsToNeighborsRaw(DIM,chunkMask,3,currentChunk->w0);
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v);
+            copyNeighborsRaw(DIM,chunkMask,0,3,currentChunk->w);
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u0);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v0);
+            copyNeighborsRaw(DIM,chunkMask,0,3,currentChunk->w0);
+        }
+        //advect
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            jd = currentChunk->jd;
+            jd0 = currentChunk->jd0;
+            u = currentChunk->ju;
+            v = currentChunk->jv;
+            w = currentChunk->jw;
+            u0 = currentChunk->ju0;
+            v0 = currentChunk->jv0;
+            w0 = currentChunk->jw0;
+            Java_electrosphere_FluidSim_advectVectors(DIM,chunkMask,currentChunk->u,currentChunk->v,currentChunk->w,currentChunk->u0,currentChunk->v0,currentChunk->w0,DIFFUSION_CONSTANT,VISCOSITY_CONSTANT,timestep);
+        }
+        //update neighbor arr
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            jd = currentChunk->jd;
+            jd0 = currentChunk->jd0;
+            u = currentChunk->ju;
+            v = currentChunk->jv;
+            w = currentChunk->jw;
+            u0 = currentChunk->ju0;
+            v0 = currentChunk->jv0;
+            w0 = currentChunk->jw0;
+            setBoundsToNeighborsRaw(DIM,chunkMask,1,currentChunk->u);
+            setBoundsToNeighborsRaw(DIM,chunkMask,2,currentChunk->v);
+            setBoundsToNeighborsRaw(DIM,chunkMask,3,currentChunk->w);
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v);
+            copyNeighborsRaw(DIM,chunkMask,0,3,currentChunk->w);
+        }
+    }
+    //solve projection
+    {
+        //update array for vectors
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            jd = currentChunk->jd;
+            jd0 = currentChunk->jd0;
+            u = currentChunk->ju;
+            v = currentChunk->jv;
+            w = currentChunk->jw;
+            u0 = currentChunk->ju0;
+            v0 = currentChunk->jv0;
+            w0 = currentChunk->jw0;
+            setBoundsToNeighborsRaw(DIM,chunkMask,1,currentChunk->u);
+            setBoundsToNeighborsRaw(DIM,chunkMask,2,currentChunk->v);
+            setBoundsToNeighborsRaw(DIM,chunkMask,3,currentChunk->w);
+            setBoundsToNeighborsRaw(DIM,chunkMask,1,currentChunk->u0);
+            setBoundsToNeighborsRaw(DIM,chunkMask,2,currentChunk->v0);
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v);
+            copyNeighborsRaw(DIM,chunkMask,0,3,currentChunk->w);
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u0);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v0);
+        }
+        //setup projection
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            jd = currentChunk->jd;
+            jd0 = currentChunk->jd0;
+            u = currentChunk->ju;
+            v = currentChunk->jv;
+            w = currentChunk->jw;
+            u0 = currentChunk->ju0;
+            v0 = currentChunk->jv0;
+            w0 = currentChunk->jw0;
+            Java_electrosphere_FluidSim_setupProjection(DIM,chunkMask,currentChunk->u,currentChunk->v,currentChunk->w,currentChunk->u0,currentChunk->v0,currentChunk->w0,DIFFUSION_CONSTANT,VISCOSITY_CONSTANT,timestep);
+        }
+        //update array for vectors
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            jd = currentChunk->jd;
+            jd0 = currentChunk->jd0;
+            u = currentChunk->ju;
+            v = currentChunk->jv;
+            w = currentChunk->jw;
+            u0 = currentChunk->ju0;
+            v0 = currentChunk->jv0;
+            w0 = currentChunk->jw0;
+            setBoundsToNeighborsRaw(DIM,chunkMask,1,currentChunk->u0);
+            setBoundsToNeighborsRaw(DIM,chunkMask,2,currentChunk->v0);
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u0);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v0);
+        }
+        //samples u0, v0
+        //sets u0
+        //these should have just been mirrored in the above
+        //
+        //Perform main projection solver
+        for(int l = 0; l < LINEARSOLVERTIMES; l++){
+            for(int i = 0; i < numChunks; i++){
+                Chunk * currentChunk = chunks[i];
+                chunkJRaw = currentChunk->jchunk;
+                chunkMask = currentChunk->chunkMask;
+                jd = currentChunk->jd;
+                jd0 = currentChunk->jd0;
+                u = currentChunk->ju;
+                v = currentChunk->jv;
+                w = currentChunk->jw;
+                u0 = currentChunk->ju0;
+                v0 = currentChunk->jv0;
+                w0 = currentChunk->jw0;
+                Java_electrosphere_FluidSim_solveProjection(DIM,chunkMask,currentChunk->u,currentChunk->v,currentChunk->w,currentChunk->u0,currentChunk->v0,currentChunk->w0,DIFFUSION_CONSTANT,VISCOSITY_CONSTANT,timestep);
+            }
+            for(int i = 0; i < numChunks; i++){
+                Chunk * currentChunk = chunks[i];
+                chunkJRaw = currentChunk->jchunk;
+                chunkMask = currentChunk->chunkMask;
+                jd = currentChunk->jd;
+                jd0 = currentChunk->jd0;
+                u = currentChunk->ju;
+                v = currentChunk->jv;
+                w = currentChunk->jw;
+                u0 = currentChunk->ju0;
+                v0 = currentChunk->jv0;
+                w0 = currentChunk->jw0;
+                setBoundsToNeighborsRaw(DIM,chunkMask,1,currentChunk->u0);
+                copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u0);
+            }
+        }
+        //samples u,v,w,u0
+        //sets u,v,w
+        //Finalize projection
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            jd = currentChunk->jd;
+            jd0 = currentChunk->jd0;
+            u = currentChunk->ju;
+            v = currentChunk->jv;
+            w = currentChunk->jw;
+            u0 = currentChunk->ju0;
+            v0 = currentChunk->jv0;
+            w0 = currentChunk->jw0;
+            Java_electrosphere_FluidSim_finalizeProjection(DIM,chunkMask,currentChunk->u,currentChunk->v,currentChunk->w,currentChunk->u0,currentChunk->v0,currentChunk->w0,DIFFUSION_CONSTANT,VISCOSITY_CONSTANT,timestep);
+        }
+        //set boundaries a final time for u,v,w
+        //...
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            jd = currentChunk->jd;
+            jd0 = currentChunk->jd0;
+            u = currentChunk->ju;
+            v = currentChunk->jv;
+            w = currentChunk->jw;
+            u0 = currentChunk->ju0;
+            v0 = currentChunk->jv0;
+            w0 = currentChunk->jw0;
+            setBoundsToNeighborsRaw(DIM,chunkMask,1,currentChunk->u);
+            setBoundsToNeighborsRaw(DIM,chunkMask,2,currentChunk->v);
+            setBoundsToNeighborsRaw(DIM,chunkMask,3,currentChunk->w);
+            setBoundsToNeighborsRaw(DIM,chunkMask,1,currentChunk->u0);
+            setBoundsToNeighborsRaw(DIM,chunkMask,2,currentChunk->v0);
+            setBoundsToNeighborsRaw(DIM,chunkMask,3,currentChunk->w0);
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v);
+            copyNeighborsRaw(DIM,chunkMask,0,3,currentChunk->w);
+            copyNeighborsRaw(DIM,chunkMask,0,1,currentChunk->u0);
+            copyNeighborsRaw(DIM,chunkMask,0,2,currentChunk->v0);
+            copyNeighborsRaw(DIM,chunkMask,0,3,currentChunk->w0);
+        }
+    }
+
+
+
+
+    ///------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+    ///------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+    ///------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+    ///------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+    ///------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+
+
+
+
+    //add density
+    {
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            Java_electrosphere_FluidSim_addDensity(DIM,chunkMask,currentChunk->d,currentChunk->d0,timestep);
+        }
+    }
+    //swap all density arrays
+    {
+        //swap vector fields
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+
+            float * tmpArr;
+            for(int j = 0; j < 27; j++){
+                tmpArr = currentChunk->d[j];
+                currentChunk->d[j] = currentChunk->d0[j];
+                currentChunk->d0[j] = tmpArr;
+            }
+        }
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            jd = currentChunk->jd;
+            jd0 = currentChunk->jd0;
+            copyNeighborsRaw(DIM,chunkMask,0,0,currentChunk->d);
+            copyNeighborsRaw(DIM,chunkMask,0,0,currentChunk->d0);
+        }
+    }
+    //diffuse density
+    {
+        for(int l = 0; l < LINEARSOLVERTIMES; l++){
+            for(int i = 0; i < numChunks; i++){
+                Chunk * currentChunk = chunks[i];
+                chunkJRaw = currentChunk->jchunk;
+                chunkMask = currentChunk->chunkMask;
+                u = currentChunk->ju;
+                v = currentChunk->jv;
+                w = currentChunk->jw;
+                Java_electrosphere_FluidSim_solveDiffuseDensity(DIM,chunkMask,currentChunk->d,currentChunk->d0,currentChunk->u,currentChunk->v,currentChunk->w,DIFFUSION_CONSTANT,VISCOSITY_CONSTANT,timestep);
+            }
+            for(int i = 0; i < numChunks; i++){
+                Chunk * currentChunk = chunks[i];
+                chunkJRaw = currentChunk->jchunk;
+                chunkMask = currentChunk->chunkMask;
+                setBoundsToNeighborsRaw(DIM,chunkMask,0,currentChunk->d);
+            }
+        }
+    }
+    //swap all density arrays
+    {
+        //swap vector fields
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            float * tmpArr;
+            for(int j = 0; j < 27; j++){
+                tmpArr = currentChunk->d[j];
+                currentChunk->d[j] = currentChunk->d0[j];
+                currentChunk->d0[j] = tmpArr;
+            }
+        }
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            copyNeighborsRaw(DIM,chunkMask,0,0,currentChunk->d);
+            copyNeighborsRaw(DIM,chunkMask,0,0,currentChunk->d0);
+        }
+    }
+    //advect density
+    {
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            Java_electrosphere_FluidSim_advectDensity(DIM,chunkMask,currentChunk->d,currentChunk->d0,currentChunk->u,currentChunk->v,currentChunk->w,DIFFUSION_CONSTANT,VISCOSITY_CONSTANT,timestep);
+        }
+    }
+    //mirror densities
+    {
+        for(int i = 0; i < numChunks; i++){
+            Chunk * currentChunk = chunks[i];
+            chunkJRaw = currentChunk->jchunk;
+            chunkMask = currentChunk->chunkMask;
+            setBoundsToNeighborsRaw(DIM,chunkMask,0,currentChunk->d);
+        }
+    }
+}

src/main/c/includes/electrosphere_FluidSim.h

@@ -16,46 +16,14 @@ extern "C" {
 #undef electrosphere_FluidSim_LINEARSOLVERTIMES
 #define electrosphere_FluidSim_LINEARSOLVERTIMES 20L
 #undef electrosphere_FluidSim_GRAVITY
-#define electrosphere_FluidSim_GRAVITY 0.0f
+#define electrosphere_FluidSim_GRAVITY -100.0f
 /*
  * Class:     electrosphere_FluidSim
- * Method:    createThreadpool
- * Signature: (I)J
+ * Method:    simulate
+ * Signature: (Ljava/util/List;F)V
  */
-JNIEXPORT jlong JNICALL Java_electrosphere_FluidSim_createThreadpool
-  (JNIEnv *, jclass, jint);
-
-/*
- * Class:     electrosphere_FluidSim
- * Method:    unlockThreads
- * Signature: (J)J
- */
-JNIEXPORT jlong JNICALL Java_electrosphere_FluidSim_unlockThreads
-  (JNIEnv *, jclass, jlong);
-
-/*
- * Class:     electrosphere_FluidSim
- * Method:    queueChunk
- * Signature: (JII[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;FFF)V
- */
-JNIEXPORT void JNICALL Java_electrosphere_FluidSim_queueChunk
-  (JNIEnv *, jclass, jlong, jint, jint, jobjectArray, jobjectArray, jobjectArray, jobjectArray, jobjectArray, jobjectArray, jobjectArray, jobjectArray, jfloat, jfloat, jfloat);
-
-/*
- * Class:     electrosphere_FluidSim
- * Method:    submitWork
- * Signature: (J)V
- */
-JNIEXPORT void JNICALL Java_electrosphere_FluidSim_submitWork
-  (JNIEnv *, jclass, jlong);
-
-/*
- * Class:     electrosphere_FluidSim
- * Method:    fetchWork
- * Signature: (J)V
- */
-JNIEXPORT void JNICALL Java_electrosphere_FluidSim_fetchWork
-  (JNIEnv *, jclass, jlong);
+JNIEXPORT void JNICALL Java_electrosphere_FluidSim_simulate
+  (JNIEnv *, jclass, jobject, jfloat);
 
 #ifdef __cplusplus
 }

src/main/c/includes/libfluidsim.h

@@ -1,16 +0,0 @@
-//include guard
-#ifndef LIB_FLUID_SIM
-#define LIB_FLUID_SIM
-
-#include "./threadpool.h"
-
-void simulate(LibraryContext * context, int threadIndex);
-
-void addDensity(int N, int chunk_mask, float * jrx, float * x0, float dt);
-
-void diffuseDensity(int N, int chunk_mask, float * jrx, float * jrx0, float DIFFUSION_CONST, float VISCOSITY_CONST, float dt);
-
-void advectDensity(uint32_t chunk_mask, int N, float * jrd, float * jrd0, float * u, float * v, float * w, float dt);
-
-//close include guard
-#endif

src/main/c/includes/mainFunctions.h

@@ -0,0 +1,101 @@
+#ifndef MAINFUNC
+#define MAINFUNC
+
+#include <jni.h>
+
+/*
+ * Class:     electrosphere_FluidSim
+ * Method:    calculateChunkMask
+ * Signature: ([Ljava/nio/ByteBuffer;)I
+ */
+JNIEXPORT jint JNICALL Java_electrosphere_FluidSim_calculateChunkMask
+  (JNIEnv *, jobject, jobjectArray);
+
+/*
+ * Class:     electrosphere_FluidSim
+ * Method:    addSourceToVectors
+ * Signature: (II[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;FFF)V
+ */
+void Java_electrosphere_FluidSim_addSourceToVectors
+  (int, int, float **, float **, float **, float **, float **, float **, float, float, float);
+
+/*
+ * Class:     electrosphere_FluidSim
+ * Method:    solveVectorDiffuse
+ * Signature: (II[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;FFF)V
+ */
+void Java_electrosphere_FluidSim_solveVectorDiffuse
+  (int, int, float **, float **, float **, float **, float **, float **, float, float, float);
+
+/*
+ * Class:     electrosphere_FluidSim
+ * Method:    setupProjection
+ * Signature: (II[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;FFF)V
+ */
+void Java_electrosphere_FluidSim_setupProjection
+  (int, int, float **, float **, float **, float **, float **, float **, float, float, float);
+
+/*
+ * Class:     electrosphere_FluidSim
+ * Method:    solveProjection
+ * Signature: (II[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;FFF)V
+ */
+void Java_electrosphere_FluidSim_solveProjection
+  (int, int, float **, float **, float **, float **, float **, float **, float, float, float);
+
+/*
+ * Class:     electrosphere_FluidSim
+ * Method:    finalizeProjection
+ * Signature: (II[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;FFF)V
+ */
+void Java_electrosphere_FluidSim_finalizeProjection
+  (int, int,  float **, float **, float **, float **, float **, float **, float, float, float);
+
+/*
+ * Class:     electrosphere_FluidSim
+ * Method:    advectVectors
+ * Signature: (II[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;FFF)V
+ */
+void Java_electrosphere_FluidSim_advectVectors
+  (int, int, float **, float **, float **, float **, float **, float **, float, float, float);
+
+/*
+ * Class:     electrosphere_FluidSim
+ * Method:    addDensity
+ * Signature: (II[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;F)V
+ */
+void Java_electrosphere_FluidSim_addDensity
+  (int, int, float **, float **, float);
+
+/*
+ * Class:     electrosphere_FluidSim
+ * Method:    solveDiffuseDensity
+ * Signature: (II[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;FFF)V
+ */
+void Java_electrosphere_FluidSim_solveDiffuseDensity
+  (int, int, float **, float **, float **, float **, float **, float, float, float);
+
+/*
+ * Class:     electrosphere_FluidSim
+ * Method:    advectDensity
+ * Signature: (II[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;[Ljava/nio/ByteBuffer;FFF)V
+ */
+void Java_electrosphere_FluidSim_advectDensity
+  (int, int, float **, float **, float **, float **, float **, float, float, float);
+
+void setBoundsToNeighborsRaw
+  (
+  int N,
+  int chunk_mask,
+  int vector_dir,
+  float ** neighborArray);
+
+void copyNeighborsRaw
+  (
+  int N,
+  int chunk_mask,
+  int cx,
+  int vector_dir,
+  float ** neighborArray);
+
+#endif

src/main/c/includes/threadpool.h

@@ -1,48 +0,0 @@
-#ifndef THREADPOOL
-#define THREADPOOL
-
-#include <pthread.h>
-
-
-/**
- * A threadpool used to distribute work
-*/
-typedef struct {
-    int numThreads;
-    pthread_t * threads;
-    pthread_barrier_t * barrierMain;
-    pthread_barrier_t * barrierWithParentThread;
-} ThreadPool;
-
-
-/**
- * A single chunk to have fluid simulated
-*/
-typedef struct {
-    int chunkMask;
-    float timestep;
-    float * d;
-    float * u;
-    float * v;
-    float * w;
-    float * d0;
-    float * u0;
-    float * v0;
-    float * w0;
-    float diffuseConst;
-    float viscosityConst;
-} Chunk;
-
-
-
-/**
- * Overall state of the whole application
-*/
-typedef struct {
-    ThreadPool * threadpool;
-    Chunk * chunks;
-    int numChunks;
-} LibraryContext;
-
-
-#endif

src/main/c/includes/utilities.h

@@ -4,19 +4,11 @@
 #ifndef UTILITIES_H
 #define UTILITIES_H
 
-//swaps where two pointers are pointing
 #define SWAP(x0,x) {float *tmp=x0;x0=x;x=tmp;}
-
-//gets the index of a 3d point in an N^3 array
 #define IX(i,j,k) ((i)+(N)*(j)+(N*N)*(k))
-
-//gets the index of a 3d point in a 3^3 array
 #define CK(m,n,o) ((m)+(n)*(3)+(o)*(3)*(3))
-
-//gets the raw array of a java buffer
 #define GET_ARR(env,src,i) (*env)->GetDirectBufferAddress(env,(*env)->GetObjectArrayElement(env,src,i))
-
-//returns true if the array exists, false otherwise
+#define GET_ARR_RAW(env,src,i) src[i]
 #define ARR_EXISTS(chunk_mask,m,n,o) (chunk_mask & CHUNK_INDEX_ARR[CK(m,n,o)]) > 0
 
 #endif

src/main/c/lib/stb

@@ -1 +0,0 @@
-Subproject commit ae721c50eaf761660b4f90cc590453cdb0c2acd0

src/main/c/src/libfluidsim.c

@@ -1,77 +0,0 @@
-#include <jni.h>
-#include <stdio.h>
-#include <immintrin.h>
-#include <stdint.h>
-#include <unistd.h>
-#include <pthread.h>
-#include <semaphore.h>
-
-//git libs
-#include "../lib/stb/stb_ds.h"
-
-//all other headers
-#include "../includes/libfluidsim.h"
-#include "../includes/threadpool.h"
-#include "../includes/chunkmask.h"
-#include "../includes/utilities.h"
-
-#define DIM 18
-#define LINEARSOLVERTIMES 20
-
-
-//call this if the child threads need to wait on one another
-void sync(LibraryContext * context){
-    pthread_barrier_wait(context->threadpool->barrierMain);
-}
-
-/**
- * Main simulation function
-*/
-void simulate(LibraryContext * context, int threadIndex){
-    int numThreads = context->threadpool->numThreads;
-    int numChunks = context->numChunks;
-
-    //add densities to chunks
-    for(int i = threadIndex; i < numChunks; i = i + numThreads){
-        Chunk chunk = context->chunks[i];
-        addDensity(DIM,chunk.chunkMask,chunk.d,chunk.d0,chunk.timestep);
-    }
-    sync(context);
-
-    //diffuse density
-    for(int l = 0; l < LINEARSOLVERTIMES; l++){
-        for(int i = threadIndex; i < numChunks; i = i + numThreads){
-            Chunk chunk = context->chunks[i];
-            SWAP(chunk.d,chunk.d0);
-            diffuseDensity(
-                DIM,
-                chunk.chunkMask,
-                chunk.d,
-                chunk.d0,
-                chunk.diffuseConst,
-                chunk.viscosityConst,
-                chunk.timestep
-            );
-            SWAP(chunk.d,chunk.d0);
-        }
-        sync(context);
-    }
-
-    //advect density
-    for(int i = threadIndex; i < numChunks; i = i + numThreads){
-        Chunk chunk = context->chunks[i];
-        advectDensity(
-            chunk.chunkMask,
-            DIM,
-            chunk.d,
-            chunk.d0,
-            chunk.u,
-            chunk.v,
-            chunk.w,
-            chunk.timestep
-        );
-    }
-
-    //call this if the child threads need to wait on one another
-    sync(context);
-}

src/main/c/src/threadpool.c

@@ -1,222 +0,0 @@
-#include <jni.h>
-#include <stdio.h>
-#include <immintrin.h>
-#include <stdint.h>
-#include <unistd.h>
-#include <pthread.h>
-#include <semaphore.h>
-
-//include stb ds
-#define STB_DS_IMPLEMENTATION
-#include "../lib/stb/stb_ds.h"
-
-//all other headers
-#include "../includes/libfluidsim.h"
-#include "../includes/threadpool.h"
-#include "../includes/chunkmask.h"
-#include "../includes/utilities.h"
-
-typedef struct {
-  LibraryContext * libraryContext;
-  int threadIndex; //the index for this thread in all threadpool objects
-} ThreadData;
-
-
-/**
- * Main loops for the thread
-*/
-void * mainThreadLoop(void * data);
-
-/*
- * Class:     electrosphere_FluidSim
- * Method:    createThreadpool
- * Signature: (I)J
- */
-JNIEXPORT jlong JNICALL Java_electrosphere_FluidSim_createThreadpool
-  (JNIEnv * env,
-  jclass class,
-  jint numThreads){
-
-  LibraryContext * libraryContext = (LibraryContext *)malloc(sizeof(LibraryContext));
-  libraryContext->numChunks = 0;
-  libraryContext->chunks = NULL; //must be tied to NULL at the start
-
-  //init threadpool
-  libraryContext->threadpool = (ThreadPool *)malloc(sizeof(ThreadPool));
-  ThreadPool * pool = libraryContext->threadpool;
-  pool->threads = (pthread_t *)malloc(sizeof(pthread_t) * numThreads);
-  pool->numThreads = numThreads;
-
-  //used for storing return codes from pthread calls
-  int retCode = 0;
-
-  //
-  //create thread barriers
-  printf("Creating barrier for %d threads\n",pool->numThreads);
-  pool->barrierMain = (pthread_barrier_t *)malloc(sizeof(pthread_barrier_t));
-  retCode = pthread_barrier_init(pool->barrierMain,NULL,pool->numThreads);
-  if(retCode != 0){
-    printf("Failed to create main barrier! %d\n",retCode);
-  }
-
-  //+1 is for the main thread
-  printf("Creating barrier for %d threads\n",pool->numThreads+1);
-  pool->barrierWithParentThread = (pthread_barrier_t *)malloc(sizeof(pthread_barrier_t));
-  retCode = pthread_barrier_init(pool->barrierWithParentThread,NULL,(pool->numThreads)+1);
-  if(retCode != 0){
-    printf("Failed to create barrier with main thread! %d\n",retCode);
-  }
-
-  //start threads
-  for(int i = 0; i < pool->numThreads; i++){
-
-    //create data to be sent to thread
-    ThreadData * threadData = (ThreadData *)malloc(sizeof(ThreadData));
-    threadData->libraryContext = libraryContext;
-    threadData->threadIndex = i;
-
-    //create thread
-    retCode = pthread_create(&pool->threads[i], NULL, mainThreadLoop, threadData);
-    if(retCode != 0){
-      printf("Failed to create thread!~ %d\n",retCode);
-    }
-  }
-
-  printf("Finished creating threads\n");
-
-  return (jlong)libraryContext;
-}
-
-
-/*
- * Debug function
- */
-JNIEXPORT jlong JNICALL Java_electrosphere_FluidSim_unlockThreads(
-  JNIEnv * env,
-  jclass class,
-  jlong threadPoolPtrRaw){
-  LibraryContext * libraryContext = (LibraryContext *)threadPoolPtrRaw;
-  pthread_barrier_wait(libraryContext->threadpool->barrierWithParentThread);
-}
-
-
-/*
- * Queues a single chunk to do work
- */
-JNIEXPORT void JNICALL Java_electrosphere_FluidSim_queueChunk(
-  JNIEnv * env,
-  jclass class,
-  jlong libraryContextPtr,
-  jint DIM,
-  jint chunkMask,
-  jobjectArray dr,
-  jobjectArray d0r,
-  jobjectArray ur,
-  jobjectArray vr,
-  jobjectArray wr,
-  jobjectArray u0r,
-  jobjectArray v0r,
-  jobjectArray w0r,
-  jfloat DIFFUSION_CONSTANT,
-  jfloat VISCOSITY_CONSTANT,
-  jfloat timestep){
-  LibraryContext * libraryContext = (LibraryContext *)libraryContextPtr;
-
-  //if the array is full, malloc a new chunk, otherwise overwrite an existing one
-  int arrayCapacity = stbds_arrlen(libraryContext->chunks);
-  if(libraryContext->numChunks >= arrayCapacity){
-    Chunk * newChunk = (Chunk *)malloc(sizeof(Chunk));
-    newChunk->chunkMask = chunkMask;
-    newChunk->u = GET_ARR(env,ur,CENTER_LOC);
-    newChunk->v = GET_ARR(env,vr,CENTER_LOC);
-    newChunk->w = GET_ARR(env,wr,CENTER_LOC);
-    newChunk->u0 = GET_ARR(env,u0r,CENTER_LOC);
-    newChunk->v0 = GET_ARR(env,v0r,CENTER_LOC);
-    newChunk->w0 = GET_ARR(env,w0r,CENTER_LOC);
-    newChunk->d = GET_ARR(env,dr,CENTER_LOC);
-    newChunk->d0 = GET_ARR(env,d0r,CENTER_LOC);
-    newChunk->timestep = timestep;
-    newChunk->diffuseConst = DIFFUSION_CONSTANT;
-    newChunk->viscosityConst = VISCOSITY_CONSTANT;
-    stbds_arrput(libraryContext->chunks,newChunk[0]);
-  } else {
-    Chunk * currentChunk = &libraryContext->chunks[libraryContext->numChunks];
-    currentChunk->chunkMask = chunkMask;
-    currentChunk->u = GET_ARR(env,ur,CENTER_LOC);
-    currentChunk->v = GET_ARR(env,vr,CENTER_LOC);
-    currentChunk->w = GET_ARR(env,wr,CENTER_LOC);
-    currentChunk->u0 = GET_ARR(env,u0r,CENTER_LOC);
-    currentChunk->v0 = GET_ARR(env,v0r,CENTER_LOC);
-    currentChunk->w0 = GET_ARR(env,w0r,CENTER_LOC);
-    currentChunk->d = GET_ARR(env,dr,CENTER_LOC);
-    currentChunk->d0 = GET_ARR(env,d0r,CENTER_LOC);
-    currentChunk->timestep = timestep;
-    currentChunk->diffuseConst = DIFFUSION_CONSTANT;
-    currentChunk->viscosityConst = VISCOSITY_CONSTANT;
-  }
-
-  libraryContext->numChunks++;
-}
-
-/*
- * Submits a request to the threadpool to do all the simulation
- */
-JNIEXPORT void JNICALL Java_electrosphere_FluidSim_submitWork(
-  JNIEnv * env,
-  jclass class,
-  jlong contextPtr
-  ){
-  LibraryContext * libraryContext = (LibraryContext *)contextPtr;
-  //begin work
-  //wait to begin work until parent signals its ready
-  pthread_barrier_wait(libraryContext->threadpool->barrierWithParentThread);
-}
-
-
-/*
- * blocks until the simulation finishes, then grabs the results of the sim
- */
-JNIEXPORT void JNICALL Java_electrosphere_FluidSim_fetchWork(
-  JNIEnv * env,
-  jclass class,
-  jlong contextPtr
-  ){
-  LibraryContext * libraryContext = (LibraryContext *)contextPtr;
-  
-  //Roll threads over to the beginning of the workflow again, meanwhile do main thread cleanup work
-  pthread_barrier_wait(libraryContext->threadpool->barrierWithParentThread);
-
-  //main thread does work to setup child chunks for receiving data here
-  libraryContext->numChunks = 0;
-}
-
-
-/**
- * Main loop for threads
-*/
-void * mainThreadLoop(void * dataRaw){
-  ThreadData * threadData = (ThreadData *)dataRaw;
-  LibraryContext * libraryContext = threadData->libraryContext;
-  int threadIndex = threadData->threadIndex;
-
-
-  //main thread loop
-  int running = 1;
-  while(running==1){
-
-    //begin work
-    //wait to begin work until parent signals its ready
-    pthread_barrier_wait(libraryContext->threadpool->barrierWithParentThread);
-
-    //do main work
-    simulate(libraryContext,threadData->threadIndex);
-
-
-
-    //finalize work
-    //the parent thread needs to call the barrier as well with whatever method it uses to grab data from the sim back to java
-    pthread_barrier_wait(libraryContext->threadpool->barrierWithParentThread);
-
-
-  }
-}

src/main/c/velocitystep.c

@@ -1,11 +1,9 @@
-#include <jni.h>
 #include <stdio.h>
 #include <immintrin.h>
 #include <stdint.h>
 
-#include "../includes/libfluidsim.h"
-#include "../includes/utilities.h"
-#include "../includes/chunkmask.h"
+#include "includes/utilities.h"
+#include "includes/chunkmask.h"
 
 
 #define BOUND_NO_DIR 0
@@ -17,29 +15,28 @@
 #define SET_BOUND_USE_NEIGHBOR 1
 
 void add_source(int N, float * x, float * s, float dt);
-void advect(JNIEnv * env, uint32_t chunk_mask, int N, int b, jobjectArray jrd, jobjectArray jrd0, float * u, float * v, float * w, float dt);
+void advect(uint32_t chunk_mask, int N, int b, float ** jrd, float ** jrd0, float * u, float * v, float * w, float dt);
 
 
 /*
  * Adds force to all vectors
  */
-void addSourceToVectors
-  (JNIEnv * env,
-  jobject this,
-  jint N,
-  jint chunk_mask,
-  jobjectArray jru,
-  jobjectArray jrv,
-  jobjectArray jrw,
-  jobjectArray jru0,
-  jobjectArray jrv0,
-  jobjectArray jrw0,
-  jfloat DIFFUSION_CONST,
-  jfloat VISCOSITY_CONST,
-  jfloat dt){
-    add_source(N,GET_ARR(env,jru,CENTER_LOC),GET_ARR(env,jru0,CENTER_LOC),dt);
-    add_source(N,GET_ARR(env,jrv,CENTER_LOC),GET_ARR(env,jrv0,CENTER_LOC),dt);
-    add_source(N,GET_ARR(env,jrw,CENTER_LOC),GET_ARR(env,jrw0,CENTER_LOC),dt);
+void Java_electrosphere_FluidSim_addSourceToVectors
+  (
+  int N,
+  int chunk_mask,
+  float ** jru,
+  float ** jrv,
+  float ** jrw,
+  float ** jru0,
+  float ** jrv0,
+  float ** jrw0,
+  float DIFFUSION_CONST,
+  float VISCOSITY_CONST,
+  float dt){
+    add_source(N,GET_ARR_RAW(env,jru,CENTER_LOC),GET_ARR_RAW(env,jru0,CENTER_LOC),dt);
+    add_source(N,GET_ARR_RAW(env,jrv,CENTER_LOC),GET_ARR_RAW(env,jrv0,CENTER_LOC),dt);
+    add_source(N,GET_ARR_RAW(env,jrw,CENTER_LOC),GET_ARR_RAW(env,jrw0,CENTER_LOC),dt);
 }
 
 void add_source(int N, float * x, float * s, float dt){
@@ -53,29 +50,28 @@ void add_source(int N, float * x, float * s, float dt){
 /*
  * Solves vector diffusion along all axis
  */
-JNIEXPORT void JNICALL Java_electrosphere_FluidSim_solveVectorDiffuse
-  (JNIEnv * env,
-  jobject this,
-  jint N,
-  jint chunk_mask,
-  jobjectArray jru,
-  jobjectArray jrv,
-  jobjectArray jrw,
-  jobjectArray jru0,
-  jobjectArray jrv0,
-  jobjectArray jrw0,
-  jfloat DIFFUSION_CONST,
-  jfloat VISCOSITY_CONST,
-  jfloat dt){
+void Java_electrosphere_FluidSim_solveVectorDiffuse
+  (
+  int N,
+  int chunk_mask,
+  float ** jru,
+  float ** jrv,
+  float ** jrw,
+  float ** jru0,
+  float ** jrv0,
+  float ** jrw0,
+  float DIFFUSION_CONST,
+  float VISCOSITY_CONST,
+  float dt){
     float a=dt*VISCOSITY_CONST*N*N*N;
     float c=1+6*a;
     int i, j, k, l, m;
-    float * u = GET_ARR(env,jru,CENTER_LOC);
-    float * v = GET_ARR(env,jrv,CENTER_LOC);
-    float * w = GET_ARR(env,jrw,CENTER_LOC);
-    float * u0 = GET_ARR(env,jru0,CENTER_LOC);
-    float * v0 = GET_ARR(env,jrv0,CENTER_LOC);
-    float * w0 = GET_ARR(env,jrw0,CENTER_LOC);
+    float * u = GET_ARR_RAW(env,jru,CENTER_LOC);
+    float * v = GET_ARR_RAW(env,jrv,CENTER_LOC);
+    float * w = GET_ARR_RAW(env,jrw,CENTER_LOC);
+    float * u0 = GET_ARR_RAW(env,jru0,CENTER_LOC);
+    float * v0 = GET_ARR_RAW(env,jrv0,CENTER_LOC);
+    float * w0 = GET_ARR_RAW(env,jrw0,CENTER_LOC);
     
     __m256 aScalar = _mm256_set1_ps(a);
     __m256 cScalar = _mm256_set1_ps(c);
@@ -162,20 +158,19 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_solveVectorDiffuse
 /*
  * Sets up a projection system of equations
  */
-JNIEXPORT void JNICALL Java_electrosphere_FluidSim_setupProjection
-  (JNIEnv * env,
-  jobject this,
-  jint N,
-  jint chunk_mask,
-  jobjectArray jru,
-  jobjectArray jrv,
-  jobjectArray jrw,
-  jobjectArray jru0,
-  jobjectArray jrv0,
-  jobjectArray jrw0,
-  jfloat DIFFUSION_CONST,
-  jfloat VISCOSITY_CONST,
-  jfloat dt){
+void Java_electrosphere_FluidSim_setupProjection
+  (
+  int N,
+  int chunk_mask,
+  float ** jru,
+  float ** jrv,
+  float ** jrw,
+  float ** jru0,
+  float ** jrv0,
+  float ** jrw0,
+  float DIFFUSION_CONST,
+  float VISCOSITY_CONST,
+  float dt){
     int i, j, k;
 
     __m256 xVector = _mm256_set1_ps(N);
@@ -185,12 +180,12 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_setupProjection
     __m256 zeroVec = _mm256_set1_ps(0);
     __m256 vector, vector2, vector3;
 
-    float * u = GET_ARR(env,jru,CENTER_LOC);
-    float * v = GET_ARR(env,jrv,CENTER_LOC);
-    float * w = GET_ARR(env,jrw,CENTER_LOC);
+    float * u = GET_ARR_RAW(env,jru,CENTER_LOC);
+    float * v = GET_ARR_RAW(env,jrv,CENTER_LOC);
+    float * w = GET_ARR_RAW(env,jrw,CENTER_LOC);
 
-    float * p = GET_ARR(env,jru0,CENTER_LOC);
-    float * div = GET_ARR(env,jrv0,CENTER_LOC);
+    float * p = GET_ARR_RAW(env,jru0,CENTER_LOC);
+    float * div = GET_ARR_RAW(env,jrv0,CENTER_LOC);
 
     float scalar = 1.0/3.0;
     float h = 1.0/N;
@@ -257,28 +252,27 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_setupProjection
 /*
  * Solves a projection system of equations
  */
-JNIEXPORT void JNICALL Java_electrosphere_FluidSim_solveProjection
-  (JNIEnv * env,
-  jobject this,
-  jint N,
-  jint chunk_mask,
-  jobjectArray jru,
-  jobjectArray jrv,
-  jobjectArray jrw,
-  jobjectArray jru0,
-  jobjectArray jrv0,
-  jobjectArray jrw0,
-  jfloat DIFFUSION_CONST,
-  jfloat VISCOSITY_CONST,
-  jfloat dt){
+void Java_electrosphere_FluidSim_solveProjection
+  (
+  int N,
+  int chunk_mask,
+  float ** jru,
+  float ** jrv,
+  float ** jrw,
+  float ** jru0,
+  float ** jrv0,
+  float ** jrw0,
+  float DIFFUSION_CONST,
+  float VISCOSITY_CONST,
+  float dt){
     int a = 1;
     int c = 6;
     int i, j, k, l, m;
     __m256 aScalar = _mm256_set1_ps(a);
     __m256 cScalar = _mm256_set1_ps(c);
 
-    float * p = GET_ARR(env,jru0,CENTER_LOC);
-    float * div = GET_ARR(env,jrv0,CENTER_LOC);
+    float * p = GET_ARR_RAW(env,jru0,CENTER_LOC);
+    float * div = GET_ARR_RAW(env,jrv0,CENTER_LOC);
     // update for each cell
     for(k=1; k<N-1; k++){
         for(j=1; j<N-1; j++){
@@ -312,20 +306,19 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_solveProjection
 /*
  * Finalizes a projection (subtract curl, set bounds, etc)
  */
-JNIEXPORT void JNICALL Java_electrosphere_FluidSim_finalizeProjection
-  (JNIEnv * env,
-  jobject this,
-  jint N,
-  jint chunk_mask,
-  jobjectArray jru,
-  jobjectArray jrv,
-  jobjectArray jrw,
-  jobjectArray jru0,
-  jobjectArray jrv0,
-  jobjectArray jrw0,
-  jfloat DIFFUSION_CONST,
-  jfloat VISCOSITY_CONST,
-  jfloat dt){
+void Java_electrosphere_FluidSim_finalizeProjection
+  (
+  int N,
+  int chunk_mask,
+  float ** jru,
+  float ** jrv,
+  float ** jrw,
+  float ** jru0,
+  float ** jrv0,
+  float ** jrw0,
+  float DIFFUSION_CONST,
+  float VISCOSITY_CONST,
+  float dt){
     int i, j, k;
     // __m256 constScalar = _mm256_set1_ps(0.5f*N);
     __m256 xScalar = _mm256_set1_ps(0.5*N);
@@ -333,12 +326,12 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_finalizeProjection
     __m256 zScalar = _mm256_set1_ps(0.5*N);
     __m256 vector, vector2, vector3;
 
-    float * u = GET_ARR(env,jru,CENTER_LOC);
-    float * v = GET_ARR(env,jrv,CENTER_LOC);
-    float * w = GET_ARR(env,jrw,CENTER_LOC);
+    float * u = GET_ARR_RAW(env,jru,CENTER_LOC);
+    float * v = GET_ARR_RAW(env,jrv,CENTER_LOC);
+    float * w = GET_ARR_RAW(env,jrw,CENTER_LOC);
 
-    float * p = GET_ARR(env,jru0,CENTER_LOC);
-    float * div = GET_ARR(env,jrv0,CENTER_LOC);
+    float * p = GET_ARR_RAW(env,jru0,CENTER_LOC);
+    float * div = GET_ARR_RAW(env,jrv0,CENTER_LOC);
 
     float h = 1.0 / N;
 
@@ -407,41 +400,40 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_finalizeProjection
 /*
  * Advects u, v, and w
  */
-JNIEXPORT void JNICALL Java_electrosphere_FluidSim_advectVectors
-(JNIEnv * env,
-  jobject this,
-  jint N,
-  jint chunk_mask,
-  jobjectArray jru,
-  jobjectArray jrv,
-  jobjectArray jrw,
-  jobjectArray jru0,
-  jobjectArray jrv0,
-  jobjectArray jrw0,
-  jfloat DIFFUSION_CONST,
-  jfloat VISCOSITY_CONST,
-  jfloat dt){
-    advect(env,chunk_mask,N,1,jru,jru0,GET_ARR(env,jru0,CENTER_LOC),GET_ARR(env,jrv0,CENTER_LOC),GET_ARR(env,jrw0,CENTER_LOC),dt);
-    advect(env,chunk_mask,N,2,jrv,jrv0,GET_ARR(env,jru0,CENTER_LOC),GET_ARR(env,jrv0,CENTER_LOC),GET_ARR(env,jrw0,CENTER_LOC),dt);
-    advect(env,chunk_mask,N,3,jrw,jrw0,GET_ARR(env,jru0,CENTER_LOC),GET_ARR(env,jrv0,CENTER_LOC),GET_ARR(env,jrw0,CENTER_LOC),dt);
+void Java_electrosphere_FluidSim_advectVectors
+(
+  int N,
+  int chunk_mask,
+  float ** jru,
+  float ** jrv,
+  float ** jrw,
+  float ** jru0,
+  float ** jrv0,
+  float ** jrw0,
+  float DIFFUSION_CONST,
+  float VISCOSITY_CONST,
+  float dt){
+    advect(chunk_mask,N,1,jru,jru0,GET_ARR_RAW(env,jru0,CENTER_LOC),GET_ARR_RAW(env,jrv0,CENTER_LOC),GET_ARR_RAW(env,jrw0,CENTER_LOC),dt);
+    advect(chunk_mask,N,2,jrv,jrv0,GET_ARR_RAW(env,jru0,CENTER_LOC),GET_ARR_RAW(env,jrv0,CENTER_LOC),GET_ARR_RAW(env,jrw0,CENTER_LOC),dt);
+    advect(chunk_mask,N,3,jrw,jrw0,GET_ARR_RAW(env,jru0,CENTER_LOC),GET_ARR_RAW(env,jrv0,CENTER_LOC),GET_ARR_RAW(env,jrw0,CENTER_LOC),dt);
 }
 
 
-void advect(JNIEnv * env, uint32_t chunk_mask, int N, int b, jobjectArray jrd, jobjectArray jrd0, float * u, float * v, float * w, float dt){
+void advect(uint32_t chunk_mask, int N, int b, float ** jrd, float ** jrd0, float * u, float * v, float * w, float dt){
     int i, j, k, i0, j0, k0, i1, j1, k1;
     int m,n,o;
     float x, y, z, s0, t0, s1, t1, u1, u0, dtx,dty,dtz;
     
     dtx=dty=dtz=dt*N;
 
-    float * d = GET_ARR(env,jrd,CENTER_LOC);
+    float * d = GET_ARR_RAW(env,jrd,CENTER_LOC);
 
-    float * d0 = GET_ARR(env,jrd0,CENTER_LOC);
+    float * d0 = GET_ARR_RAW(env,jrd0,CENTER_LOC);
 
     for(k=1; k<N-1; k++){
         for(j=1; j<N-1; j++){
             for(i=1; i<N-1; i++){
-                d0 = GET_ARR(env,jrd0,CENTER_LOC);
+                d0 = GET_ARR_RAW(env,jrd0,CENTER_LOC);
                 //calculate location to pull from
                 x = i-dtx*u[IX(i,j,k)];
                 y = j-dty*v[IX(i,j,k)];
@@ -507,7 +499,7 @@ void advect(JNIEnv * env, uint32_t chunk_mask, int N, int b, jobjectArray jrd, j
 
 
                     // printf("Hit other chunk\n");
-                    d0 = GET_ARR(env,jrd0,CK(m,n,o));
+                    d0 = GET_ARR_RAW(env,jrd0,CK(m,n,o));
                     x = x + CHUNK_NORMALIZE_U[CK(m,n,o)] * (N-2);
                     // printf("%d => %f\n",m,x);
                     y = y + CHUNK_NORMALIZE_V[CK(m,n,o)] * (N-2);
@@ -669,15 +661,14 @@ void advect(JNIEnv * env, uint32_t chunk_mask, int N, int b, jobjectArray jrd, j
     }
 }
 
-JNIEXPORT void JNICALL Java_electrosphere_FluidSim_setBoundsToNeighbors
-  (JNIEnv * env,
-  jobject this,
-  jint N,
-  jint chunk_mask,
-  jint vector_dir,
-  jobjectArray neighborArray){
+void setBoundsToNeighborsRaw
+  (
+  int N,
+  int chunk_mask,
+  int vector_dir,
+  float ** neighborArray){
     int DIM = N;
-    float * target = GET_ARR(env,neighborArray,CENTER_LOC);
+    float * target = GET_ARR_RAW(env,neighborArray,CENTER_LOC);
     float * source;
     for(int x=1; x < DIM-1; x++){
         for(int y = 1; y < DIM-1; y++){
@@ -726,19 +717,19 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_setBoundsToNeighbors
     target[IX(DIM-1,DIM-1,DIM-1)] = (float)((target[IX(DIM-1,DIM-1,DIM-2)]+target[IX(DIM-1,DIM-2,DIM-1)]+target[IX(DIM-1,DIM-1,DIM-2)])/3.0);
 }
 
+
 /**
  * This exclusively copies neighbors to make sure zeroing out stuff doesn't break sim
 */
-JNIEXPORT void JNICALL Java_electrosphere_FluidSim_copyNeighbors
-  (JNIEnv * env,
-  jobject this,
-  jint N,
-  jint chunk_mask,
-  jint cx,
-  jint vector_dir,
-  jobjectArray neighborArray){
+void copyNeighborsRaw
+  (
+  int N,
+  int chunk_mask,
+  int cx,
+  int vector_dir,
+  float ** neighborArray){
     int DIM = N;
-    float * target = GET_ARR(env,neighborArray,CENTER_LOC);
+    float * target = GET_ARR_RAW(env,neighborArray,CENTER_LOC);
     float * source;
 
 
@@ -747,9 +738,20 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_copyNeighbors
     //   PLANES
     //
     //
+    // __m512 transferVector;// = _mm512_set1_ps(0.5*N);
+
+    //__m256 vector = _mm256_loadu_ps(&p[IX(i-1,j,k)]);
+    //vector = _mm256_add_ps(vector,_mm256_loadu_ps(&p[IX(i+1,j,k)]));
+    //vector = _mm256_add_ps(vector,_mm256_loadu_ps(&p[IX(i,j-1,k)]));
+    //_mm256_storeu_ps(&p[IX(i,j,k)],vector);
+    //__m256
+    //_mm256_loadu_ps
+    //_mm256_storeu_ps
     if(ARR_EXISTS(chunk_mask,0,1,1)){
-        source = GET_ARR(env,neighborArray,CK(0,1,1));
+        source = GET_ARR_RAW(env,neighborArray,CK(0,1,1));
         for(int x=1; x < DIM-1; x++){
+            // transferVector = _mm512_loadu_ps(&source[IX(DIM-2,x,1)]);
+            // _mm512_storeu_ps(&target[IX(0,x,1)],_mm512_loadu_ps(&source[IX(DIM-2,x,1)]));
             for(int y = 1; y < DIM-1; y++){
                 target[IX(0,x,y)] = source[IX(DIM-2,x,y)];
             }
@@ -757,8 +759,9 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_copyNeighbors
     }
 
     if(ARR_EXISTS(chunk_mask,2,1,1)){
-        source = GET_ARR(env,neighborArray,CK(2,1,1));
+        source = GET_ARR_RAW(env,neighborArray,CK(2,1,1));
         for(int x=1; x < DIM-1; x++){
+            // _mm512_storeu_ps(&target[IX(DIM-1,x,1)],_mm512_loadu_ps(&source[IX(1,x,1)]));
             for(int y = 1; y < DIM-1; y++){
                 target[IX(DIM-1,x,y)] = source[IX(1,x,y)];
             }
@@ -766,7 +769,7 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_copyNeighbors
     }
 
     if(ARR_EXISTS(chunk_mask,1,0,1)){
-        source = GET_ARR(env,neighborArray,CK(1,0,1));
+        source = GET_ARR_RAW(env,neighborArray,CK(1,0,1));
         for(int x=1; x < DIM-1; x++){
             for(int y = 1; y < DIM-1; y++){
                 target[IX(x,0,y)] = source[IX(x,DIM-2,y)];
@@ -775,7 +778,7 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_copyNeighbors
     }
 
     if(ARR_EXISTS(chunk_mask,1,2,1)){
-        source = GET_ARR(env,neighborArray,CK(1,2,1));
+        source = GET_ARR_RAW(env,neighborArray,CK(1,2,1));
         for(int x=1; x < DIM-1; x++){
             for(int y = 1; y < DIM-1; y++){
                 target[IX(x,DIM-1,y)] = source[IX(x,1,y)];
@@ -784,7 +787,7 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_copyNeighbors
     }
 
     if(ARR_EXISTS(chunk_mask,1,1,0)){
-        source = GET_ARR(env,neighborArray,CK(1,1,0));
+        source = GET_ARR_RAW(env,neighborArray,CK(1,1,0));
         for(int x=1; x < DIM-1; x++){
             for(int y = 1; y < DIM-1; y++){
                 target[IX(x,y,0)] = source[IX(x,y,DIM-2)];
@@ -793,7 +796,7 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_copyNeighbors
     }
 
     if(ARR_EXISTS(chunk_mask,1,1,2)){
-        source = GET_ARR(env,neighborArray,CK(1,1,2));
+        source = GET_ARR_RAW(env,neighborArray,CK(1,1,2));
         for(int x=1; x < DIM-1; x++){
             for(int y = 1; y < DIM-1; y++){
                 target[IX(x,y,DIM-1)] = source[IX(x,y,1)];
@@ -808,28 +811,28 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_copyNeighbors
     //
     //
     if(ARR_EXISTS(chunk_mask,0,0,1)){
-        source = GET_ARR(env,neighborArray,CK(0,0,1));
+        source = GET_ARR_RAW(env,neighborArray,CK(0,0,1));
         for(int x=1; x < DIM-1; x++){
                 target[IX(0,0,x)] = source[IX(DIM-2,DIM-2,x)];
         }
     }
 
     if(ARR_EXISTS(chunk_mask,2,0,1)){
-        source = GET_ARR(env,neighborArray,CK(2,0,1));
+        source = GET_ARR_RAW(env,neighborArray,CK(2,0,1));
         for(int x=1; x < DIM-1; x++){
                 target[IX(DIM-1,0,x)] = source[IX(1,DIM-2,x)];
         }
     }
 
     if(ARR_EXISTS(chunk_mask,0,2,1)){
-        source = GET_ARR(env,neighborArray,CK(0,2,1));
+        source = GET_ARR_RAW(env,neighborArray,CK(0,2,1));
         for(int x=1; x < DIM-1; x++){
                 target[IX(0,DIM-1,x)] = source[IX(DIM-2,1,x)];
         }
     }
 
     if(ARR_EXISTS(chunk_mask,2,2,1)){
-        source = GET_ARR(env,neighborArray,CK(2,2,1));
+        source = GET_ARR_RAW(env,neighborArray,CK(2,2,1));
         for(int x=1; x < DIM-1; x++){
                 target[IX(DIM-1,DIM-1,x)] = source[IX(1,1,x)];
         }
@@ -839,28 +842,28 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_copyNeighbors
     //
 
     if(ARR_EXISTS(chunk_mask,0,1,0)){
-        source = GET_ARR(env,neighborArray,CK(0,1,0));
+        source = GET_ARR_RAW(env,neighborArray,CK(0,1,0));
         for(int x=1; x < DIM-1; x++){
                 target[IX(0,x,0)] = source[IX(DIM-2,x,DIM-2)];
         }
     }
 
     if(ARR_EXISTS(chunk_mask,2,1,0)){
-        source = GET_ARR(env,neighborArray,CK(2,1,0));
+        source = GET_ARR_RAW(env,neighborArray,CK(2,1,0));
         for(int x=1; x < DIM-1; x++){
                 target[IX(DIM-1,x,0)] = source[IX(1,x,DIM-2)];
         }
     }
 
     if(ARR_EXISTS(chunk_mask,0,1,2)){
-        source = GET_ARR(env,neighborArray,CK(0,1,2));
+        source = GET_ARR_RAW(env,neighborArray,CK(0,1,2));
         for(int x=1; x < DIM-1; x++){
                 target[IX(0,x,DIM-1)] = source[IX(DIM-2,x,1)];
         }
     }
 
     if(ARR_EXISTS(chunk_mask,2,1,2)){
-        source = GET_ARR(env,neighborArray,CK(2,1,2));
+        source = GET_ARR_RAW(env,neighborArray,CK(2,1,2));
         for(int x=1; x < DIM-1; x++){
                 target[IX(DIM-1,x,DIM-1)] = source[IX(1,x,1)];
         }
@@ -870,28 +873,28 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_copyNeighbors
     //
 
     if(ARR_EXISTS(chunk_mask,1,0,0)){
-        source = GET_ARR(env,neighborArray,CK(1,0,0));
+        source = GET_ARR_RAW(env,neighborArray,CK(1,0,0));
         for(int x=1; x < DIM-1; x++){
                 target[IX(x,0,0)] = source[IX(x,DIM-2,DIM-2)];
         }
     }
 
     if(ARR_EXISTS(chunk_mask,1,2,0)){
-        source = GET_ARR(env,neighborArray,CK(1,2,0));
+        source = GET_ARR_RAW(env,neighborArray,CK(1,2,0));
         for(int x=1; x < DIM-1; x++){
                 target[IX(x,DIM-1,0)] = source[IX(x,1,DIM-2)];
         }
     }
 
     if(ARR_EXISTS(chunk_mask,1,0,2)){
-        source = GET_ARR(env,neighborArray,CK(1,0,2));
+        source = GET_ARR_RAW(env,neighborArray,CK(1,0,2));
         for(int x=1; x < DIM-1; x++){
                 target[IX(x,0,DIM-1)] = source[IX(x,DIM-2,1)];
         }
     }
 
     if(ARR_EXISTS(chunk_mask,1,2,2)){
-        source = GET_ARR(env,neighborArray,CK(1,2,2));
+        source = GET_ARR_RAW(env,neighborArray,CK(1,2,2));
         for(int x=1; x < DIM-1; x++){
                 target[IX(x,DIM-1,DIM-1)] = source[IX(x,1,1)];
         }
@@ -905,22 +908,22 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_copyNeighbors
     //
 
     if(ARR_EXISTS(chunk_mask,0,0,0)){
-        source = GET_ARR(env,neighborArray,CK(0,0,0));
+        source = GET_ARR_RAW(env,neighborArray,CK(0,0,0));
         target[IX(0,0,0)] = source[IX(DIM-2,DIM-2,DIM-2)];
     }
 
     if(ARR_EXISTS(chunk_mask,2,0,0)){
-        source = GET_ARR(env,neighborArray,CK(2,0,0));
+        source = GET_ARR_RAW(env,neighborArray,CK(2,0,0));
         target[IX(DIM-1,0,0)] = source[IX(1,DIM-2,DIM-2)];
     }
 
     if(ARR_EXISTS(chunk_mask,0,2,0)){
-        source = GET_ARR(env,neighborArray,CK(0,2,0));
+        source = GET_ARR_RAW(env,neighborArray,CK(0,2,0));
         target[IX(0,DIM-1,0)] = source[IX(DIM-2,1,DIM-2)];
     }
 
     if(ARR_EXISTS(chunk_mask,2,2,0)){
-        source = GET_ARR(env,neighborArray,CK(2,2,0));
+        source = GET_ARR_RAW(env,neighborArray,CK(2,2,0));
         target[IX(DIM-1,DIM-1,0)] = source[IX(1,1,DIM-2)];
     }
 
@@ -928,22 +931,22 @@ JNIEXPORT void JNICALL Java_electrosphere_FluidSim_copyNeighbors
     //
 
     if(ARR_EXISTS(chunk_mask,0,0,2)){
-        source = GET_ARR(env,neighborArray,CK(0,0,2));
+        source = GET_ARR_RAW(env,neighborArray,CK(0,0,2));
         target[IX(0,0,DIM-1)] = source[IX(DIM-2,DIM-2,1)];
     }
 
     if(ARR_EXISTS(chunk_mask,2,0,2)){
-        source = GET_ARR(env,neighborArray,CK(2,0,2));
+        source = GET_ARR_RAW(env,neighborArray,CK(2,0,2));
         target[IX(DIM-1,0,DIM-1)] = source[IX(1,DIM-2,1)];
     }
 
     if(ARR_EXISTS(chunk_mask,0,2,2)){
-        source = GET_ARR(env,neighborArray,CK(0,2,2));
+        source = GET_ARR_RAW(env,neighborArray,CK(0,2,2));
         target[IX(0,DIM-1,DIM-1)] = source[IX(DIM-2,1,1)];
     }
 
     if(ARR_EXISTS(chunk_mask,2,2,2)){
-        source = GET_ARR(env,neighborArray,CK(2,2,2));
+        source = GET_ARR_RAW(env,neighborArray,CK(2,2,2));
         target[IX(DIM-1,DIM-1,DIM-1)] = source[IX(1,1,1)];
     }
 

src/main/java/electrosphere/FluidSim.java

@@ -15,6 +15,7 @@ import org.joml.Vector2i;
 import org.joml.Vector3i;
 import org.lwjgl.BufferUtils;
 import org.lwjgl.PointerBuffer;
+import org.lwjgl.glfw.GLFW;
 import org.lwjgl.system.MemoryUtil;
 
 /**
@@ -40,21 +41,21 @@ public class FluidSim {
     //        +-------------+ (2,0,0)           +---------------> X
 
     //Buffers that contain density for current frame
-    ByteBuffer[] density = new ByteBuffer[27];
+    public ByteBuffer[] density = new ByteBuffer[27];
     //Buffers that contain new density to add to the simulation
-    ByteBuffer[] densityAddition = new ByteBuffer[27];
+    public ByteBuffer[] densityAddition = new ByteBuffer[27];
     //Buffers that contain u vector directions
-    ByteBuffer[] uVector = new ByteBuffer[27];
+    public ByteBuffer[] uVector = new ByteBuffer[27];
     //Buffers that contain v vector directions
-    ByteBuffer[] vVector = new ByteBuffer[27];
+    public ByteBuffer[] vVector = new ByteBuffer[27];
     //Buffers that contain w vector directions
-    ByteBuffer[] wVector = new ByteBuffer[27];
+    public ByteBuffer[] wVector = new ByteBuffer[27];
     //Buffers that contain u vector directions to add to the simulation
-    ByteBuffer[] uAdditionVector = new ByteBuffer[27];
+    public ByteBuffer[] uAdditionVector = new ByteBuffer[27];
     //Buffers that contain v vector directions to add to the simulation
-    ByteBuffer[] vAdditionVector = new ByteBuffer[27];
+    public ByteBuffer[] vAdditionVector = new ByteBuffer[27];
     //Buffers that contain w vector directions to add to the simulation
-    ByteBuffer[] wAdditionVector = new ByteBuffer[27];
+    public ByteBuffer[] wAdditionVector = new ByteBuffer[27];
 
     //The densities for every voxel for the current frame
     float[] densityArrayView = new float[DIM * DIM * DIM];
@@ -69,9 +70,7 @@ public class FluidSim {
     public float[] v0ArrayView = new float[DIM * DIM * DIM];
     float[] w0ArrayView = new float[DIM * DIM * DIM];
 
-    int chunkMask = 0;
-
-    public static long threadpool;
+    public int chunkMask = 0;
     
 
     static final float DIFFUSION_CONSTANT = 0.0f;
@@ -79,8 +78,7 @@ public class FluidSim {
 
     static final int LINEARSOLVERTIMES = 20;
 
-    static final float GRAVITY = 0.0f;
-    // static final float GRAVITY = -100f; //<-- original value
+    static final float GRAVITY = -100f;
     
     public void setup(Vector3i offset){
         //allocate buffers for this chunk
@@ -157,8 +155,12 @@ public class FluidSim {
         }
     }
 
-    static LinkedList<FluidSim> simArrayList = new LinkedList<FluidSim>();
+    static int i = 0;
+    static double time = 0;
+    static double lastTime = 0;
     public static void simChunks(FluidSim[][][] simArray, int step, float timestep){
+
+        List<FluidSim> chunksToSim = new LinkedList<FluidSim>();
         //
         //init data for upcoming frame
         for(int x = 0; x < simArray.length; x++){
@@ -173,44 +175,25 @@ public class FluidSim {
                     //
                     simArray[x][y][z].writeNewStateIntoBuffers();
                     //
-                    //flag this chunk as being simulated in the upcoming frame
-                    simArrayList.add(simArray[x][y][z]);
+                    // add to queue
+                    //
+                    chunksToSim.add(simArray[x][y][z]);
                 }
             }
         }
 
-        //for each chunk, send them to C land
-        for(FluidSim chunk : simArrayList){
-            queueChunkWrapper(threadpool, chunk, timestep);
+        lastTime = GLFW.glfwGetTime();
+        //
+        //simulate
+        simulateWrapper(chunksToSim,timestep);
+        //clock
+        time = time + (GLFW.glfwGetTime() - lastTime);
+        i++;
+        if(i == 100){
+            System.out.println(time / 100.0 * 1000.0);
         }
 
-        //call for work to be done
-        submitWorkWrapper(threadpool);
         
-        // //
-        // //Vector stage
-        // solveChunkMask(simArray);
-        // addVectorSources(simArray, timestep);
-        // swapAllVectorFields(simArray, timestep);
-        // solveVectorDiffusion(simArray, timestep);
-        // solveProjection(simArray, step, timestep);
-        // swapAllVectorFields(simArray, timestep);
-        // advectVectorsAcrossBoundaries(simArray, timestep);
-        // solveProjection(simArray, step, timestep);
-
-        // //
-        // //Density stage
-        // addDensity(simArray, timestep);
-        // swapAllDensityArrays(simArray, timestep);
-        // diffuseDensity(simArray, timestep);
-        // swapAllDensityArrays(simArray, timestep);
-        // advectDensity(simArray, timestep);
-        // // mirrorNeighborDensities(simArray, timestep);
-
-        //call for work to be done
-        fetchWorkWrapper(threadpool);
-
-        simArrayList.clear();
 
 
         //
@@ -351,620 +334,16 @@ public class FluidSim {
         return rVal;
     }
 
-    // private static void solveChunkMask(FluidSim[][][] simArray){
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-    //                 simArray[x][y][z].calculateChunkMaskWrapper();
-    //             }
-    //         }
-    //     }
-    // }
-
-    // private static void addVectorSources(FluidSim[][][] simArray, float timestep){
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-
-    //                 //Add source to all 3 vectors
-    //                 // add_source(N, u, u0, dt);
-    //                 // add_source(N, v, v0, dt);
-    //                 // add_source(N, w, w0, dt);
-    //                 simArray[x][y][z].addSourceToVectorsWrapper(timestep);
-    //             }
-    //         }
-    //     }
-    //     //swap
-    //     //u <=> u0 etc for u, v, and w
-        
-    // }
-
-    // private static void solveVectorDiffusion(FluidSim[][][] simArray, float timestep){
-    //     //samples u,v,w,u0,v0,w0
-    //     //sets u,v,w
-    //     // for(int x = 0; x < simArray.length; x++){
-    //     //     for(int y = 0; y < simArray[0].length; y++){
-    //     //         for(int z = 0; z < simArray[0][0].length; z++){
-    //     //             simArray[x][y][z].copyNeighborsWrapper(1, simArray[x][y][z].uVector);
-    //     //             simArray[x][y][z].copyNeighborsWrapper(2, simArray[x][y][z].vVector);
-    //     //             simArray[x][y][z].copyNeighborsWrapper(3, simArray[x][y][z].wVector);
-    //     //             simArray[x][y][z].copyNeighborsWrapper(1, simArray[x][y][z].uAdditionVector);
-    //     //             simArray[x][y][z].copyNeighborsWrapper(2, simArray[x][y][z].vAdditionVector);
-    //     //             simArray[x][y][z].copyNeighborsWrapper(3, simArray[x][y][z].wAdditionVector);
-    //     //         }
-    //     //     }
-    //     // }
-    //     for(int l = 0; l < LINEARSOLVERTIMES; l++){
-    //         for(int x = 0; x < simArray.length; x++){
-    //             for(int y = 0; y < simArray[0].length; y++){
-    //                 for(int z = 0; z < simArray[0][0].length; z++){
-    //                     //lin_solve(env, chunk_mask, N, b, x, x0, a, 1+6*a);
-    //                     //+
-    //                     //set_bnd(env, chunk_mask, N, b, x);
-    //                     //for u, v, and w all in 1 shot
-    //                     simArray[x][y][z].solveVectorDiffuseWrapper(timestep);
-    //                 }
-    //             }
-    //         }
-    //         for(int x = 0; x < simArray.length; x++){
-    //             for(int y = 0; y < simArray[0].length; y++){
-    //                 for(int z = 0; z < simArray[0][0].length; z++){
-    //                     simArray[x][y][z].setBoundsToNeighborsWrapper(1, simArray[x][y][z].uVector);
-    //                     simArray[x][y][z].setBoundsToNeighborsWrapper(2, simArray[x][y][z].vVector);
-    //                     simArray[x][y][z].setBoundsToNeighborsWrapper(3, simArray[x][y][z].wVector);
-    //                     simArray[x][y][z].copyNeighborsWrapper(1, x, simArray[x][y][z].uVector);
-    //                     simArray[x][y][z].copyNeighborsWrapper(2, x, simArray[x][y][z].vVector);
-    //                     simArray[x][y][z].copyNeighborsWrapper(3, x, simArray[x][y][z].wVector);
-    //                 }
-    //             }
-    //         }
-    //     }
-    // }
-
-    // private static void solveProjection(FluidSim[][][] simArray, int step, float timestep){
-    //     //samples u,v,w
-    //     //sets u0,v0
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(1, simArray[x][y][z].uVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(2, simArray[x][y][z].vVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(3, simArray[x][y][z].wVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(1, simArray[x][y][z].uAdditionVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(2, simArray[x][y][z].vAdditionVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(1, x, simArray[x][y][z].uVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(2, x, simArray[x][y][z].vVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(3, x, simArray[x][y][z].wVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(1, x, simArray[x][y][z].uAdditionVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(2, x, simArray[x][y][z].vAdditionVector);
-    //             }
-    //         }
-    //     }
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-    //                 // System.out.println("Setup " + x + " " +  y + " " + z);
-    //                 //setup projection across boundaries
-    //                 //...
-    //                 //set boundaries appropriately
-    //                 //...
-    //                 simArray[x][y][z].setupProjectionWrapper(timestep);
-    //             }
-    //         }
-    //     }
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(0, simArray[x][y][z].uAdditionVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(0, simArray[x][y][z].vAdditionVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(0, x, simArray[x][y][z].uAdditionVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(0, x, simArray[x][y][z].vAdditionVector);
-    //             }
-    //         }
-    //     }
-    //     //samples u0, v0
-    //     //sets u0
-    //     //these should have just been mirrored in the above
-    //     //
-    //     //Perform main projection solver
-    //     for(int l = 0; l < LINEARSOLVERTIMES; l++){
-    //         for(int x = 0; x < simArray.length; x++){
-    //             for(int y = 0; y < simArray[0].length; y++){
-    //                 for(int z = 0; z < simArray[0][0].length; z++){
-    //                     //lin_solve(env, chunk_mask, N, b, x, x0, a, 1+6*a);
-    //                     //for u, v, and w all in 1 shot
-    //                     simArray[x][y][z].solveProjectionWrapper(timestep);
-    //                 }
-    //             }
-    //         }
-    //         //be sure to set boundaries to neighbor chunk values where appropriate
-    //         for(int x = 0; x < simArray.length; x++){
-    //             for(int y = 0; y < simArray[0].length; y++){
-    //                 for(int z = 0; z < simArray[0][0].length; z++){
-    //                     simArray[x][y][z].setBoundsToNeighborsWrapper(0, simArray[x][y][z].uAdditionVector);
-    //                     simArray[x][y][z].copyNeighborsWrapper(0, x, simArray[x][y][z].uAdditionVector);
-    //                 }
-    //             }
-    //         }
-    //     }
-    //     //samples u,v,w,u0
-    //     //sets u,v,w
-    //     //Finalize projection
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-    //                 //Subtract curl field from current vector field
-    //                 //...
-    //                 simArray[x][y][z].finalizeProjectionWrapper(timestep);
-    //             }
-    //         }
-    //     }
-    //     //set boundaries a final time for u,v,w
-    //     //...
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(1, simArray[x][y][z].uVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(2, simArray[x][y][z].vVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(3, simArray[x][y][z].wVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(1, x, simArray[x][y][z].uVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(2, x, simArray[x][y][z].vVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(3, x, simArray[x][y][z].wVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(1, simArray[x][y][z].uAdditionVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(2, simArray[x][y][z].vAdditionVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(3, simArray[x][y][z].wAdditionVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(0, x, simArray[x][y][z].uAdditionVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(0, x, simArray[x][y][z].vAdditionVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(0, x, simArray[x][y][z].wAdditionVector);
-    //             }
-    //         }
-    //     }
-    // }
-
-    // private static void mirrorNeighborDensities(FluidSim[][][] simArray, float timestep){
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(0, simArray[x][y][z].density);
-    //             }
-    //         }
-    //     }
-    // }
-
-    // private static void swapAllVectorFields(FluidSim[][][] simArray, float timestep){
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-    //                 simArray[x][y][z].swapVectorFields();
-    //             }
-    //         }
-    //     }
-    //     //then need to mirror each array as relevant
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-    //                 simArray[x][y][z].copyNeighborsWrapper(1, x, simArray[x][y][z].uVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(2, x, simArray[x][y][z].vVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(3, x, simArray[x][y][z].wVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(1, x, simArray[x][y][z].uAdditionVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(2, x, simArray[x][y][z].vAdditionVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(2, x, simArray[x][y][z].wAdditionVector);
-    //             }
-    //         }
-    //     }
-    // }
-
-    // private void swapVectorFields(){
-    //     ByteBuffer tmp;
-    //     //swap x0 <-> x
-    //     // tmp = densityAddition;
-    //     // densityAddition = density[13];
-    //     // density[13] = tmp;
-    //     //swap u0 <-> u
-    //     for(int i = 0; i < 27; i++){
-    //         tmp = uAdditionVector[i];
-    //         uAdditionVector[i] = uVector[i];
-    //         uVector[i] = tmp;
-    //         //swap v0 <-> v
-    //         tmp = vAdditionVector[i];
-    //         vAdditionVector[i] = vVector[i];
-    //         vVector[i] = tmp;
-    //         //swap w0 <-> w
-    //         tmp = wAdditionVector[i];
-    //         wAdditionVector[i] = wVector[i];
-    //         wVector[i] = tmp;
-    //     }
-    //     //...
-    // }
-
-    // private static void advectVectorsAcrossBoundaries(FluidSim[][][] simArray, float timestep){
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(1, simArray[x][y][z].uVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(2, simArray[x][y][z].vVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(3, simArray[x][y][z].wVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(1, simArray[x][y][z].uAdditionVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(2, simArray[x][y][z].vAdditionVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(3, simArray[x][y][z].wAdditionVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(1, x, simArray[x][y][z].uVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(2, x, simArray[x][y][z].vVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(3, x, simArray[x][y][z].wVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(1, x, simArray[x][y][z].uAdditionVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(2, x, simArray[x][y][z].vAdditionVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(3, x, simArray[x][y][z].wAdditionVector);
-    //             }
-    //         }
-    //     }
-    //     //samples u,v,w,u0,v0,w0
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-    //                 // advect(env, chunk_mask, N, 1, jru, u0, u0, v0, w0, dt);
-    //                 // advect(env, chunk_mask, N, 2, jrv, v0, u0, v0, w0, dt);
-    //                 // advect(env, chunk_mask, N, 3, jrw, w0, u0, v0, w0, dt);
-    //                 //...
-    //                 simArray[x][y][z].advectVectorsWrapper(timestep);
-    //             }
-    //         }
-    //     }
-    //     //mirror neighbor data
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(1, simArray[x][y][z].uVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(2, simArray[x][y][z].vVector);
-    //                 simArray[x][y][z].setBoundsToNeighborsWrapper(3, simArray[x][y][z].wVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(1, x, simArray[x][y][z].uVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(2, x, simArray[x][y][z].vVector);
-    //                 simArray[x][y][z].copyNeighborsWrapper(3, x, simArray[x][y][z].wVector);
-    //             }
-    //         }
-    //     }
-    // }
-
-    // private static void addDensity(FluidSim[][][] simArray, float timestep){
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-    //                 //add_source(N, x, x0, dt);
-    //                 simArray[x][y][z].addDensityWrapper(timestep);
-    //                 //swap x <=> x0
-    //                 //swap arrays in java side...
-    //                 // simArray[x][y][z].swapDensityArrays();
-    //             }
-    //         }
-    //     }
-    // }
-
-    // private static void swapAllDensityArrays(FluidSim[][][] simArray, float timestep){
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-    //                 simArray[x][y][z].swapDensityArrays();
-    //             }
-    //         }
-    //     }
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-    //                 simArray[x][y][z].copyNeighborsWrapper(0, x, simArray[x][y][z].density);
-    //                 simArray[x][y][z].copyNeighborsWrapper(0, x, simArray[x][y][z].densityAddition);
-    //             }
-    //         }
-    //     }
-    // }
-
-    // private void swapDensityArrays(){
-    //     for(int i = 0; i < 27; i++){
-    //         ByteBuffer tmp = density[i];
-    //         density[i] = densityAddition[i];
-    //         densityAddition[i] = tmp;
-    //     }
-    // }
-
-    // private static void diffuseDensity(FluidSim[][][] simArray, float timestep){
-    //     for(int l = 0; l < LINEARSOLVERTIMES; l++){
-    //         for(int x = 0; x < simArray.length; x++){
-    //             for(int y = 0; y < simArray[0].length; y++){
-    //                 for(int z = 0; z < simArray[0][0].length; z++){
-    //                     //lin_solve(env, chunk_mask, N, b, x, x0, a, 1+6*a);
-    //                     //+
-    //                     //set_bnd(env, chunk_mask, N, b, x);
-    //                     simArray[x][y][z].solveDiffuseDensityWrapper(timestep);
-    //                 }
-    //             }
-    //         }
-    //         for(int x = 0; x < simArray.length; x++){
-    //             for(int y = 0; y < simArray[0].length; y++){
-    //                 for(int z = 0; z < simArray[0][0].length; z++){
-    //                     simArray[x][y][z].setBoundsToNeighborsWrapper(0, simArray[x][y][z].density);
-    //                 }
-    //             }
-    //         }
-    //     }
-    // }
-
-    // private static void advectDensity(FluidSim[][][] simArray, float timestep){
-    //     for(int x = 0; x < simArray.length; x++){
-    //         for(int y = 0; y < simArray[0].length; y++){
-    //             for(int z = 0; z < simArray[0][0].length; z++){
-    //                 //swap x <=> x0 again
-    //                 // simArray[x][y][z].swapDensityArrays();
-    //                 //advect density
-    //                 simArray[x][y][z].advectDensityWrapper(timestep);
-    //             }
-    //         }
-    //     }
-    // }
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
 
 
     /**
-     *  OLD NATIVE DECLARATIONS
+     * Main simulation function
+     * @param timestep
      */
-
-
-
-    // /**
-    //  * The native function call to simulate a frame of fluid
-    //  * @param DIM_X
-    //  * @param DIM_Y
-    //  * @param DIM_Z
-    //  * @param x
-    //  * @param x0
-    //  * @param u
-    //  * @param v
-    //  * @param w
-    //  * @param u0
-    //  * @param v0
-    //  * @param w0
-    //  * @param DIFFUSION_CONSTANT
-    //  * @param VISCOSITY_CONSTANT
-    //  * @param timestep
-    //  */
-    // private native void simulate(
-    //     int DIM_X,
-    //     int chunkMask,
-    //     ByteBuffer[] x,
-    //     ByteBuffer x0,
-    //     ByteBuffer[] u,
-    //     ByteBuffer[] v,
-    //     ByteBuffer[] w,
-    //     ByteBuffer[] u0,
-    //     ByteBuffer[] v0,
-    //     ByteBuffer[] w0,
-    //     float DIFFUSION_CONSTANT,
-    //     float VISCOSITY_CONSTANT,
-    //     float timestep
-    // );
-
-    // private void calculateChunkMaskWrapper(){
-    //     this.chunkMask = this.calculateChunkMask(density);
-    // }
-    // /**
-    //  * Calculates the mask of chunk neighbors
-    //  * @param densityBuffers The neighbor array
-    //  * @return The mask
-    //  */
-    // private native int calculateChunkMask(ByteBuffer[] densityBuffers);
-
-    // /**
-    //  * Add vector values to u, v, and w all at once
-    //  */
-    // private void addSourceToVectorsWrapper(float timestep){
-    //     addSourceToVectors(DIM, chunkMask, uVector, vVector, wVector, uAdditionVector, vAdditionVector, wAdditionVector, DIFFUSION_CONSTANT, VISCOSITY_CONSTANT, timestep);
-    // }
-    // private native void addSourceToVectors(int DIM_X, int chunkMask, ByteBuffer[] u, ByteBuffer v[], ByteBuffer w[], ByteBuffer u0[], ByteBuffer v0[], ByteBuffer w0[], float DIFFUSION_CONSTANT, float VISCOSITY_CONSTANT, float timestep);
-
-    // /**
-    //  * Solves u, v, and w diffusion systems of equations
-    //  */
-    // private void solveVectorDiffuseWrapper(float timestep){
-    //     solveVectorDiffuse(DIM, chunkMask, uVector, vVector, wVector, uAdditionVector, vAdditionVector, wAdditionVector, DIFFUSION_CONSTANT, VISCOSITY_CONSTANT, timestep);
-    // }
-    // private native void solveVectorDiffuse(int DIM_X, int chunkMask, ByteBuffer[] u, ByteBuffer v[], ByteBuffer w[], ByteBuffer u0[], ByteBuffer v0[], ByteBuffer w0[], float DIFFUSION_CONSTANT, float VISCOSITY_CONSTANT, float timestep);
-
-    // /**
-    //  * Setup projection system
-    //  */
-    // private void setupProjectionWrapper(float timestep){
-    //     setupProjection(DIM, chunkMask, uVector, vVector, wVector, uAdditionVector, vAdditionVector, wAdditionVector, DIFFUSION_CONSTANT, VISCOSITY_CONSTANT, timestep);
-    // }
-    // private native void setupProjection(int DIM_X, int chunkMask, ByteBuffer[] u, ByteBuffer v[], ByteBuffer w[], ByteBuffer u0[], ByteBuffer v0[], ByteBuffer w0[], float DIFFUSION_CONSTANT, float VISCOSITY_CONSTANT, float timestep);
-
-    // /**
-    //  * Solve projection system
-    //  */
-    // private void solveProjectionWrapper(float timestep){
-    //     solveProjection(DIM, chunkMask, uVector, vVector, wVector, uAdditionVector, vAdditionVector, wAdditionVector, DIFFUSION_CONSTANT, VISCOSITY_CONSTANT, timestep);
-    // }
-    // private native void solveProjection(int DIM_X, int chunkMask, ByteBuffer[] u, ByteBuffer v[], ByteBuffer w[], ByteBuffer u0[], ByteBuffer v0[], ByteBuffer w0[], float DIFFUSION_CONSTANT, float VISCOSITY_CONSTANT, float timestep);
-
-    // /**
-    //  * Does work like subtracting curl from vector field, setting boundaries, etc
-    //  */
-    // private void finalizeProjectionWrapper(float timestep){
-    //     finalizeProjection(DIM, chunkMask, uVector, vVector, wVector, uAdditionVector, vAdditionVector, wAdditionVector, DIFFUSION_CONSTANT, VISCOSITY_CONSTANT, timestep);
-    // }
-    // private native void finalizeProjection(int DIM_X, int chunkMask, ByteBuffer[] u, ByteBuffer v[], ByteBuffer w[], ByteBuffer u0[], ByteBuffer v0[], ByteBuffer w0[], float DIFFUSION_CONSTANT, float VISCOSITY_CONSTANT, float timestep);
-
-    // /**
-    //  * Advects vectors
-    //  */
-    // private void advectVectorsWrapper(float timestep){
-    //     advectVectors(DIM, chunkMask, uVector, vVector, wVector, uAdditionVector, vAdditionVector, wAdditionVector, DIFFUSION_CONSTANT, VISCOSITY_CONSTANT, timestep);
-    // }
-    // private native void advectVectors(int DIM_X, int chunkMask, ByteBuffer[] u, ByteBuffer v[], ByteBuffer w[], ByteBuffer u0[], ByteBuffer v0[], ByteBuffer w0[], float DIFFUSION_CONSTANT, float VISCOSITY_CONSTANT, float timestep);
-
-    // /**
-    //  * Adds density to the simulation
-    //  */
-    // private void addDensityWrapper(float timestep){
-    //     addDensity(DIM, chunkMask, density, densityAddition, timestep);
-    // }
-    // private native void addDensity(int DIM_X, int chunkMask, ByteBuffer[] x, ByteBuffer[] x0, float timestep);
-
-    // /**
-    //  * Solve density diffusion
-    //  */
-    // private void solveDiffuseDensityWrapper(float timestep){
-    //     solveDiffuseDensity(DIM, chunkMask, density, densityAddition, uVector, vVector, wVector, DIFFUSION_CONSTANT, VISCOSITY_CONSTANT, timestep);
-    // }
-    // private native void solveDiffuseDensity(int DIM_X, int chunkMask, ByteBuffer[] x, ByteBuffer[] x0, ByteBuffer[] u, ByteBuffer v[], ByteBuffer w[], float DIFFUSION_CONSTANT, float VISCOSITY_CONSTANT, float timestep);
-
-    // /**
-    //  * Solve density diffusion
-    //  */
-    // private void advectDensityWrapper(float timestep){
-    //     advectDensity(DIM, chunkMask, density, densityAddition, uVector, vVector, wVector, DIFFUSION_CONSTANT, VISCOSITY_CONSTANT, timestep);
-    // }
-    // private native void advectDensity(int DIM_X, int chunkMask, ByteBuffer[] x, ByteBuffer[] x0, ByteBuffer[] u, ByteBuffer v[], ByteBuffer w[], float DIFFUSION_CONSTANT, float VISCOSITY_CONSTANT, float timestep);
-
-
-    // /**
-    //  * Sets the bounds of the neighbormap to neighbor values if available
-    //  */
-    // private void setBoundsToNeighborsWrapper(int vectorDir, ByteBuffer[] neighborMap){
-    //     setBoundsToNeighbors(DIM, chunkMask, vectorDir, neighborMap);
-    // }
-    // private native void setBoundsToNeighbors(int DIM_X, int chunkMask, int vectorDir, ByteBuffer[] neighborMap);
-
-    // /**
-    //  * Sets the bounds of the neighbormap to neighbor values if available, otherwise doesn't mess with them.
-    //  * This is to make sure zeroing out doesn't mess up the sim
-    //  */
-    // private void copyNeighborsWrapper(int vectorDir, int x, ByteBuffer[] neighborMap){
-    //     copyNeighbors(DIM, chunkMask, x, vectorDir, neighborMap);
-    // }
-    // private native void copyNeighbors(int DIM_X, int chunkMask, int x, int vectorDir, ByteBuffer[] neighborMap);
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-    
-
-
-
-    /**
-     * Creates the thread pool for handling main work of the c lib
-     */
-    public static long createThreadpoolWrapper(int numThreads){
-        return createThreadpool(numThreads);
+    private static void simulateWrapper(List<FluidSim> chunks, float timestep){
+        simulate(chunks,timestep);
     }
-    private static native long createThreadpool(int numThreads);
-
-
-    /**
-     * Unlocks the C threadpool
-     */
-    public static long unlockThreadsWrapper(){
-        return unlockThreads(threadpool);
-    }
-    private static native long unlockThreads(long threadPoolPtr);
-
-
-    /**
-     * Queues a single chunk to be simulated this frame
-     * @param chunk The chunk
-     * @param timestep The amount of time to simulate it for
-     */
-    public static void queueChunkWrapper(long libraryContext, FluidSim chunk, float timestep){
-        queueChunk(
-            libraryContext,
-            DIM,
-            chunk.chunkMask,
-            chunk.density,
-            chunk.densityAddition,
-            chunk.uVector,
-            chunk.vVector,
-            chunk.wVector,
-            chunk.uAdditionVector,
-            chunk.vAdditionVector,
-            chunk.wAdditionVector,
-            DIFFUSION_CONSTANT,
-            VISCOSITY_CONSTANT,
-            timestep
-        );
-    }
-    public static native void queueChunk(
-        long libraryContext,
-        int DIM_X,
-        int chunkMask,
-        ByteBuffer[] d,
-        ByteBuffer[] d0,
-        ByteBuffer[] u,
-        ByteBuffer[] v,
-        ByteBuffer[] w,
-        ByteBuffer[] u0,
-        ByteBuffer[] v0,
-        ByteBuffer[] w0,
-        float DIFFUSION_CONSTANT,
-        float VISCOSITY_CONSTANT,
-        float timestep
-    );
-
-    /**
-     * Sends work to the threadpool
-     */
-    public static void submitWorkWrapper(long contextPtr){
-        submitWork(contextPtr);
-    }
-    public static native void submitWork(long contextPtr);
-
-    /**
-     * Fetches work from the threadpool
-     */
-    public static void fetchWorkWrapper(long contextPtr){
-        fetchWork(contextPtr);
-    }
-    public static native void fetchWork(long contextPtr);
-
+    private static native void simulate(List<FluidSim> chunks, float timestep);
 
 
 
@@ -1114,7 +493,7 @@ public class FluidSim {
                 for(int k = 0; k < DIM; k++){
                     index = ((i)+(DIM)*(j) + (DIM)*(DIM)*(k));
                     u0ArrayView[index] = 0;
-                    v0ArrayView[index] = 0;
+                    v0ArrayView[index] = densityArrayView[index] * GRAVITY;
                     w0ArrayView[index] = 0;
                 }
             }

src/main/java/electrosphere/Main.java

@@ -22,7 +22,7 @@ public class Main {
 
     public static void main(String args[]){
 
-        int dim = 10;
+        int dim = 5;
         int i = 0;
         long time = 0;
         long lastTime = 0;
@@ -36,9 +36,9 @@ public class Main {
             Mesh.initShaderProgram();
 
 
-            FluidSim[][][] simArray = initFluidSim(dim,1,1);
+            FluidSim[][][] simArray = initFluidSim(dim,1,dim);
             
-            Mesh[][][] meshArray = initMeshes(dim,1,1,simArray);
+            Mesh[][][] meshArray = initMeshes(dim,1,dim,simArray);
 
 
 
@@ -55,10 +55,7 @@ public class Main {
                 //
                 //Simulate
                 //
-                FluidSim.unlockThreadsWrapper();
-                FluidSim.simChunks(simArray,i,0.0001f);
-                FluidSim.unlockThreadsWrapper();
-                time = time + (System.currentTimeMillis() - lastTime);
+                FluidSim.simChunks(simArray,i,0.01f);
                 //
                 //Remesh
                 //
@@ -69,11 +66,12 @@ public class Main {
                         }
                     }
                 }
+                time = time + (System.currentTimeMillis() - lastTime);
                 //redraw
                 GLFWContext.redraw(meshArray);
                 i++;
                 if(i == 100){
-                    System.out.println(time / 100.0);
+                    System.out.println("overall time: " + time / 100.0);
                 }
                 if(i > 3){
                     // scan.next();
@@ -85,7 +83,6 @@ public class Main {
     }
 
     private static FluidSim[][][] initFluidSim(int dimx, int dimy, int dimz){
-        FluidSim.threadpool = FluidSim.createThreadpoolWrapper(5);
         FluidSim[][][] simArray = new FluidSim[dimx][dimy][dimz];
         for(int x = 0; x < simArray.length; x++){
             for(int y = 0; y < simArray[0].length; y++){