# N-Body Optimization

I've created a serial C++ code for gravitational N-Body calculation. Because I expect to have upwards of 8-71 sparse bodies (ie, where Barnes-Hut is not necessarily practical) and running for long periods of time, I want to make as much use of parallelization and vectorization as possible. I did try a method with mutex and conditional_variable however, I found that this implementation works significantly faster: locking and unlocking mutexes proved more overhead for relatively short functions for threads. Forgive my probably obnoxious attempt at this, it is my first attempt at anything parallel and/or vectorized and I'm still new with C++, so I expect there will be plenty of criticism.

It is just two classes, Body and NBody and a helper namespace mathx.

Body.h

#pragma once

#include <immintrin.h>
#include <intrin.h>

struct Body {
__m256d pos, vel;
double mu;

Body();
Body(double MU, const __m256d& position, const __m256d& velocity);
Body(const Body& orig);
~Body();

virtual __m256d grav(const __m256d & R) const;
void push(const __m256d & acc, const __m256d & dt);
};



Body.cpp

#include "Body.h"
#include <cmath>

Body::Body() {
mu = 1;
pos = _mm256_setzero_pd();
vel = _mm256_setzero_pd();
}

Body::Body(double MU, const __m256d& position, const __m256d& velocity){
pos = position;
vel = velocity;
mu = MU;
}

Body::Body(const Body& orig) {
pos = orig.pos;
vel = orig.vel;
mu = orig.mu;
}

Body::~Body() {
}

__m256d Body::grav(const __m256d & R) const {
const double g = mu/(R[3]*R[3]*R[3]);
}

void Body::push(const __m256d & acc, const __m256d & dt){
}



NBody.h


#pragma once

#include "orbital/Body.h"
#include <vector>
#include <atomic>
#include <stdint.h>

class alignas(32) NBody {
public:
NBody();
~NBody();

void par_leapfrog(double time);
void par_step();

void setTime(double time);
void setTimestep(double step);
void setTimeInterval(double t_interval);

void output(std::string filename);

private:

// Body Stuff
std::vector< Body > bodies;

std::vector< double > times;
std::vector< std::vector< double * > > positions; // for some reason cant store __m256d

void setup();
void leapfrog_halfstep();

// Time Stuff
double t = 0., dt = 5, time_interval = 3600.0, t_test = 0.;
__m256d _dt;

// Gate / Parallel Stuff
std::atomic<uint_fast8_t> nFinished = 0;
bool done = false;
bool step = false;
bool accelerate = false;
bool push = false;

void worker();

// Internal Variables
std::atomic<uint_fast8_t> idxR, idxBody;
__m256d * R; // array of vector distance between bodies

};



NBody.cpp

#include "NBody.h"
#include <utility>
#include "geometry/mathx.h"
#include <iostream>
#include <string>
#include <cmath>

NBody::NBody() {
}

NBody::~NBody() {
}

bodies.push_back(b);
}

void NBody::par_leapfrog(double time){
setup();

leapfrog_halfstep(); // single threaded half step

for(uint_fast8_t i = 0; i < nThreads; i++){
}

while(t < time) {

par_step();

if(t > t_test) {
times.push_back(t);
t_test += time_interval;
}

t += dt;
}

done = true;
}

void NBody::setup() {
t_test = t;
nBodies = bodies.size();
done = false;
positions.resize(nBodies);
nR = mathx::combination(nBodies,2);
R = new __m256d[nR];

// reset this
step = false;
accelerate = false;
done = false;

}

void NBody::leapfrog_halfstep() {

// single thread this for convenience
__m256d acc;
__m256d dt2 = _mm256_set_pd(dt/2,dt/2,dt/2,dt/2);
for(uint_fast8_t i = 0; i < nBodies;i++) {
acc = _mm256_setzero_pd();
for(uint_fast8_t j = 0; j < nBodies; j++) {
if(i != j) {
__m256d R_tmp = _mm256_sub_pd(bodies[j].pos,bodies[i].pos);
__m256d tmp = _mm256_mul_pd(R_tmp,R_tmp);
R_tmp[3] = sqrt(tmp[0]+tmp[1]+tmp[2]);
}
}
bodies[i].vel = _mm256_fmsub_pd(acc,dt2,bodies[i].vel);
}
}

if (nBodies < max){
} else {
if (max > 0) {
} else {
}
}
}

void NBody::par_step(){
// Gate 1
idxR = 0;
nFinished = 0;
step = true;
step = false;
// Gate 2
idxBody = 0;
nFinished = 0;
accelerate = true;
accelerate = false;
}

void NBody::worker() {
__m256d acc;
uint_fast8_t i_body,j_body,ix,ix1;

// Generate indexes locally
uint_fast8_t is[nR];
uint_fast8_t js[nR];
uint_fast8_t idx_R[nBodies][nBodies];

unsigned int count = 0;
for ( i_body = 0; i_body < nBodies;i_body++) {
for( j_body = i_body+1; j_body < nBodies; j_body++) {
is[count] = i_body;
js[count] = j_body;
count++;
}
}

for(i_body = 0; i_body < nBodies; i_body++){
for(j_body = 0; j_body < nBodies; j_body++) {
if(j_body > i_body) {
idx_R[i_body][j_body] = (i_body*nBodies + j_body - mathx::combination(i_body+2,2));
} else {
idx_R[i_body][j_body] = (j_body*nBodies + i_body - mathx::combination(j_body+2,2));
}
}
}

while (!done) {

while(!step){if(done) return;}

while(idxR < nR) {
if(ix >= nR) {
break;
}

ix1 = ix+1;

__m256d dr1 = _mm256_sub_pd(bodies[js[ix]].pos,bodies[is[ix]].pos);
__m256d dr1_sq = _mm256_mul_pd( dr1,dr1 );

if(ix1 < nR) {

__m256d dr2 = _mm256_sub_pd(bodies[js[ix1]].pos,bodies[is[ix1]].pos);
__m256d dr2_sq = _mm256_mul_pd( dr2,dr2 );

__m256d temp = _mm256_hadd_pd( dr1_sq, dr2_sq );
__m128d hi128 = _mm256_extractf128_pd( temp, 1 );
__m128d dotproduct_sqrt = _mm_sqrt_pd(_mm_add_pd( _mm256_castpd256_pd128(temp), hi128 ));

dr1[3] = dotproduct_sqrt[0];
dr2[3] = dotproduct_sqrt[1];

R[ix] = std::move(dr1);
R[ix1] = std::move(dr2);

} else {

dr1[3] = sqrt(dr1_sq[0]+dr1_sq[1]+dr1_sq[2]);
R[ix] = std::move(dr1);

}
}

nFinished++;

while(!accelerate){}

while(idxBody < nBodies) { // this check is quick and avoids having to fetch add again
i_body = idxBody++;
if(i_body >= nBodies){
break;
}

// Store position prior to push
if (t > t_test) {
double pos[] = new double[3]{bodies[i_body].pos[0],bodies[i_body].pos[1],bodies[i_body].pos[2]};
positions[i_body].push_back(pos));
}

// sum gravitational acclerations
acc = _mm256_setzero_pd();
for(j_body = 0; j_body < nBodies; j_body++) {
// reverse vector (subtract) if index are reverse order
if(j_body > i_body) {
} else if (j_body < i_body) {
acc =_mm256_sub_pd(bodies[j_body].grav(R[idx_R[i_body][j_body]]),acc);
}
}

bodies[i_body].push(acc,_dt);

}

nFinished++;
}

}

void NBody::setTime(double time){
t = time;
}

void NBody::setTimestep(double step){
dt = step;
}

void NBody::setTimeInterval(double t_interval){
time_interval = t_interval;
}



mathx.h

#pragma once

#include <vector>
#include <utility>

#define UINT unsigned int

namespace mathx {

double legendrePoly(UINT n, double x);

double assocLegendrePoly(UINT l, UINT m, double x);

const unsigned long long factorial[] = {1,1,2,6,24,120,720,5040,40320,362880,3628800,39916800,479001600,6227020800,87178291200,1307674368000,20922789888000,355687428096000,6402373705728000,121645100408832000,2432902008176640000};

double generalBinomial(double alpha, UINT k);

const UINT C[11][11] = {{1},{1,1},{1,2,1},{1,3,3,1},{1,4,6,4,1},{1,5,10,10,5,1},{1,6,15,20,15,6,1},{1,7,21,35,35,21,7,1},{1,8,28,56,70,56,28,8,1},{1,9,36,84,126,126,36,9,1},{1,10,45,120,210,252,210,120,45,10,1}};

UINT combination(UINT n, UINT k);

}



mathx.cpp


#include "mathx.h"
#include <cmath>

namespace mathx {

double legendrePoly(UINT n, double x){
if (n == 0)
return 1;
if (n == 1)
return x;

double sums = 0;

for (UINT k = 0; k < n; k++) {
if (k > 3){
sums += pow(x,k) * (combination(n,k) * generalBinomial((n+k-1)*0.5,n));
} else {
if(k == 0) {
sums += generalBinomial((n+k-1)*0.5,n);
} else {
if(k == 1) {
sums += x * n * generalBinomial((n+k-1)*0.5,n);
} else {
sums += x * n * generalBinomial((n+k-1)*0.5,n);
}
}
}
}
return (1<<n) * sums;
}

double assocLegendrePoly(UINT l, UINT m, double x){
int sums = 0;
for (UINT k = m; k <= l; k++) {
int prod = k;
for (UINT j = m; m < k; m++)
prod *= j;
sums += prod* pow(x,k-m) * combination(l,k) * generalBinomial((l+k-1)*0.5,l);
}
if (m % 2 == 0)
return (1<<l) * pow((1-x*x),m/2) *sums;
else
return -1 * (1<<l) * pow((1-x*x),m*0.5) *sums;
}

double generalBinomial(double alpha, UINT k){
// this can be further optimized for half values required by legendre
double res = 1;
for (UINT i = 1; i <= k; ++i)
res = res * (alpha - (k + i)) / i;
return res;
}

UINT combination(UINT n, UINT k) {
if(n <= 10) {
return C[n][k];
}
if(k > n/2){
return combination(n,n-k);
}
UINT num = n;
UINT den = k;
//vectorizable
for(UINT i = 1; i < k; i++){
den *= i;
num *= (n-i);
}
return num/den;
}
}



EDIT:

Adding some of my testing calls that I used, really basic stuff I just inserted into a main function.


int test_parallel(int n, double t) {
//unsigned seed1 = std::chrono::system_clock::now().time_since_epoch().count();
std::default_random_engine generator;

std::uniform_real_distribution<double> mus (1.0,2.0);
std::uniform_real_distribution<double> xs (-2.0,2.0);

NBody sim;

for(int i = 0; i<n;i++) {
}

std::cout << "start test 3 \n";
auto t1 = std::chrono::high_resolution_clock::now();
sim.par_leapfrog(t);
auto t2 = std::chrono::high_resolution_clock::now();
std::cout << "test function took " << std::chrono::duration_cast<std::chrono::milliseconds>(t2-t1).count() << " milliseconds \n";
return 0;
}


int testBody() {

Body B = Body(2, _mm256_set_pd(0.0,1.0,1.0,1.0),_mm256_set_pd(0.0,-1.0,-1.0,-1.0));

__m256d dt = _mm256_set_pd(1.0,1.0,1.0,1.0);
__m256d acc = _mm256_set_pd(2.0,2.0,2.0,2.0);

B.push(acc,dt);

if(abs(B.pos[0]-2.0) < 1e-12 && abs(B.pos[1]-2.0) < 1e-12 && abs(B.pos[2]-2.0) < 1e-12) {
if(abs(B.vel[0]-1.0) < 1e-12 && abs(B.vel[1]-1.0) < 1e-12 && abs(B.vel[2]-1.0) < 1e-12) {
return 0;
} else {
return 2;
}
} else {
return 1;
}

}


int testGravity() {

Body B = Body();
B.mu = 16;

__m256d R = _mm256_set_pd(2.0,0.0,2.0,0.0);
__m256d g = B.grav(R);

if(abs(g[1]-4.0) < 1e-12 ) {
if(abs(g[0]) > 1e-12 ) {
return 2;
}
return 0;
} else {
return 1;
}

}

$$$$

• I've unit tested just about everything as part of my serial C++ code. The thing that I haven't been able to test was the NBody.cpp output because I would need to export the output and compare it to my serial, which I am currently working on. I just figured I'd post this now. – Marius Popescu Jun 2 at 16:49
• Since the Vec3 class is completely superfluous, it should not be included as part of this review. You could create a source file demonstrating its use then post it as part of a different question. – 1201ProgramAlarm Jun 2 at 16:57
• Looks like you got something good going, but you're a bit early in the process to get it reviewed. Please, feel free to come back once you've verified your code actually does what it's supposed to do. – Mast Jun 2 at 17:28
• I have rolled back Rev 3 → 2. Please see What to do when someone answers. – Sᴀᴍ Onᴇᴌᴀ Jun 8 at 18:20
• Thanks, new here, will create new one. – Marius Popescu Jun 8 at 18:27

# Data layout

You have already experienced first-hand a disadvantage of using "1 physics vector = 1 SIMD vector" (such as __m256d pos), causing some friction when coordinates come together:

__m256d temp = _mm256_hadd_pd( dr1_sq, dr2_sq );
__m128d hi128 = _mm256_extractf128_pd( temp, 1 );
__m128d dotproduct_sqrt = _mm_sqrt_pd(_mm_add_pd( _mm256_castpd256_pd128(temp), hi128 ));


Mixing different coordinates in the same SIMD vector leads to horizontal addition and shuffles and extraction and such. Horizontal addition is relatively expensive, equivalent to two shuffles plus a normal addition. _mm256_castpd256_pd128 is free, but extracting the upper half is not.

That strategy of using the 4th component for a different value is also a problem, causing even more extract/insert operations. As a rule of thumb, avoid indexing into SIMD vectors. It's fine to use that construct a bit in a pinch, but I would say it's overused here.

There is an alternative: put the X components of 4 physics vectors together into a SIMD vector, Y in an other SIMD vector, etc. You could have groups of 4 bodies together (AoSoA), or a big array of just X and an other of Y and so on (SoA).

That's a significant rewrite, but I recommend it. That Vec3 that was mentioned, I recommend against the entire idea. It's still using SIMD against the grain. It's a really "attractive looking trap", letting you express the computation in a way that feels nice, but it's not a way that results in good code.

# Unnecessary move

Moving SIMD vectors is not useful. They're trivial to copy and hold no resource.

# Alignment

Aligning NBody aligns its first field, which is an std::vector (so the vector object itself, not the data it holds). That's not useful, but also not harmful. std::vector should, as of C++17, respect the alignment of the data inside it (before 17, that was simply broken).

# Scary synchronization

bool accelerate should not be used for synchronization, it makes this construct unsafe: while(!accelerate){}. That loop may not terminate, or it may work as intended, it's not reliable. Using atomic<bool> would make the threads communicate safely.

• I'll just convert all the bools to atomics, thanks! As for the rewrite of the vectors... I figured my version was suboptimal, that's part of the reason I dropped the vec3 idea too because I figured I would load all x components into an array and vectorize that big array. Should I align the individual fields including integer variables (nBodies, nThreads etc) instead? – Marius Popescu Jun 2 at 20:09
• @MariusPopescu don't bother aligning random things, the main data should be aligned to its natural alignment (eg __m256d should be aligned 32 bytes) but in C++17 that is supposed to happen automatically – harold Jun 2 at 21:02
• Thanks for your help! – Marius Popescu Jun 4 at 13:47

Basics:

Body.h/Body.cpp

The class Body is extremely simple and all its functions are under 5 lines. Calling a function is a relatively heavy operation and calling a virtual function is even more so. Putting but a few operations inside a function will make it an inefficient call. Unless, the function is inlined. The compiler cannot inline functions that are hidden from compilation - so you should move all the quick functions to the header and keep cpp for the heavier stuff.

P.S. why does this class even have a virtual function? you don't utilize the property anywhere.

Inherently, when you multithread your code, the computer has to do more work. All the data synchronization and memory-ownership swapping is not cheap for low-level code. So it is quite possible that the single threaded version would run faster - or at the same speed just with single core at maximal capacity instead of all of them.

If the number of bodies would be huge, like a few thousands, then perhaps multi-threading will improve performance. Though, the exact numbers surely depends on the platform and implementation.

You should read more on std::atomic as regular operations like ++, --, +=, -=, = are slow and usually unnecessarily so. You should read its memory model and use operations like load, store, fetch_add... with appropriate memory instructions.

Linear Algebra:

As suggested by @harold, you shouldn't use __m256d for storing x,y,z coordinates of the body but rather store's n-body's coordinates in a 3xn matrix. Also this way you can perform matrix level operations and utilize SIMD types more efficiently: e.g., you don't waste a coordinate and you can utilize AVX512 instructions which holds twice as much data as __m256d.

Algorithm:

You use a very basic and inaccurate algorithm for N-Body computation: V(t+dt) = V(t) +dt*a(t) and P(t+dt) = P(t)+dt*V(t+dt). I think this is like first order of inaccuracy. What's the point of running the simulation for a long time if it is of such a low accuracy?

You should check out better solutions like Runge–Kutta methods.

• The virtual function will be used later, some planets have spherical harmonics. Even this multithreaded approach is 3x faster than serial. As I understand it, += is the same as fetch add. I am rewriting the way I use SIMD. Leapfrog is symplectic, runge kutta is not, so better for long term angular momentum conservation. There are multi step symplectic methods, but I mostly wanted to figure out the parallelization first. – Marius Popescu Jun 3 at 17:01
• += is same as fetch_add with memory instruction memory_order_seq_cst`. The difference lies in memory instructions. Did you compare it with atomic-less straightforward single core in both terms of efficiency and accuracy? It could be fast because it produces wrong results. – ALX23z Jun 3 at 17:30
• I wanted to keep the most strict memory instruction so that worked out for me anyway. But yes, originally posted the code produced the same results as single core, (and no instrinsic instructions) – Marius Popescu Jun 4 at 13:49
• @MariusPopescu about virtual functions and low level code - if you have only a couple of low level types with only minor variations then it is better to just distinguish them via an enum or store in different vectors with different types. Keep the virtual functions / polymorphism for higher level objects. It is not advisable for it to be applied to low level types. – ALX23z Jun 4 at 14:51
• The reason for the virtual function is the that gravitational calculations for bodies with off spherical values are far more complicated, hence the reason for the polymorphism. We typically only model them for Earth, Moon, Mars, but even Sun and any body we may do longer term maneuvers in. Also perturbation terms can be added in the push calculation if needed. – Marius Popescu Jun 8 at 13:10