Table of Contents

After playing with the tools we have some more options to improve the performance of the particle system. This time, we need to rewrite some parts of the code.

In total, the particle system runs almost twice as fast as initially! Read more to see what pieces of code were changed.

The Series  

Plan for this post  

Start  

We are starting with those numbers, see the previous post (last results)

Core i5 Sandy Bridge

count tunnel attractors fountain
171000 429.195 608.598 460.299
181000 460.649 647.825 490.412
191000 489.206 688.603 520.302

Core i5 Ivy Bridge

count tunnel attractors fountain
171000 529.188 746.594 570.297
181000 565.648 792.824 605.912
191000 593.956 832.478 640.739

(time in milliseconds)

SIMD preparation  

Previously I tried to force the compiler to use SSE2 or AVX instructions. As we saw, there was a nice performance boost (around 10% for AVX). But hey… SIMD should calculate things 4x or 8x times faster… so why we got only a little improvement?

In real life it’s not that simple:

  • SIMD can do 4 or 8 instructions at a time, but we still need to wait for the memory. See my summary of a talk “Native code performance on modern CPUs” for more information. In general, we can get max 2.5x speedup using SSE2/4, assuming we have ideally ‘vectorizable’ code. Not all code is in such a perfect state.
  • Current CPU’s are superscalar that means CPU can run several different instructions in parallel. Sometimes SIMD code can be even slower than the original code created by a compiler.
  • Additional small problem: SIMD registers needs memory chunks to be aligned to 128bits (16-byte alignment). We need to take care of this when we allocate new memory. So not every variable or array is good for SSE code.

Whan can we do?

  • Since particles operate mostly on glm::vec4 there is a high chance to use the full power of SSE. We use 4 floats per vector, 16 bytes.
  • glm adds a very nice feature glm::simdVec4 which basically adds SSE code to common vector functions. So I simply changed glm::vec4 to glm::simdVec4.
  • Memory must be aligned, so I used _aligned_malloc and _aligned_free.

Some code examples:

// particles.h, in ParticleData class declaration
glm::simdVec4 *m_pos;
glm::simdVec4 *m_col;

// in particles.cpp, generate() method:
m_pos = (glm::simdVec4 *)_aligned_malloc(sizeof(glm::vec4)*maxSize, 16);
m_col = (glm::simdVec4 *)_aligned_malloc(sizeof(glm::vec4)*maxSize, 16);

// particles.cpp, destructor
_aligned_free(m_pos);
_aligned_free(m_col);

The results after changes (Visual Studio):

Sandy Bridge:

count tunnel attractors fountain
171000 387.563 495.281 394.641
181000 417.320 529.660 426.330
191000 447.665 563.833 450.416

Ivy Bridge:

count tunnel attractors fountain
171000 476.625 596.313 483.656
181000 514.328 639.664 523.332
191000 552.666 682.333 558.667

Wow: almost 20% of improvement! All by proper data structures (for vectors) and memory alignment.

SSE and AVX instructions  

So far we got a nice speed up… Now, let’s write some SSE code for most critical loops. Will it run faster?

Euler update, SSE:

__m128 ga = globalA.Data;
__m128 *pa, *pb, pc;
__m128 ldt = _mm_set_ps1(localDT);

size_t i;
for (i = 0; i < endId; i++)
{
    pa = (__m128*)(&p->m_acc[i].x);
    *pa = _mm_add_ps(*pa, ga);
}

for (i = 0; i < endId; i ++)
{
    pa = (__m128*)(&p->m_vel[i].x);
    pb = (__m128*)(&p->m_acc[i].x);
    pc = _mm_mul_ps(*pb, ldt);
    *pa = _mm_add_ps(*pa, pc);
}

for (size_t i = 0; i < endId; i++)
{
    pa = (__m128*)(&p->m_pos[i].x);
    pb = (__m128*)(&p->m_vel[i].x);
    pc = _mm_mul_ps(*pb, ldt);
    *pa = _mm_add_ps(*pa, pc);
}

Readability is much worse in this case.

The results:

Sandy Bridge

count tunnel attractors fountain
171000 386.453 492.727 393.363
181000 416.182 529.591 423.795
191000 444.398 564.199 450.099

Ivy Bridge:

count tunnel attractors fountain
171000 481.172 584.086 486.543
181000 516.271 623.136 514.068
191000 547.034 656.517 541.258

Not much, unfortunately. This is because of glm::simdVec4 which uses SSE code. So there is no point in rewriting it. We lose readability and the performance gain is questionable.

Pointer aliasing: __restrict keyword  

In my previous post I got a very interesting comment from Matías N. Goldberg:

… In my experience you get massive performance improvements by just telling the compiler the pointers do not alias to each other…

Matias suggest using __restrict keyword to tell the compiler that pointers are not aliasing. For instance:

glm::vec4 * __restrict acc = p->m_acc;
glm::vec4 * __restrict vel = p->m_vel;
glm::vec4 * __restrict pos = p->m_pos;

And then, instead of p->m_pos just use pos pointer.

When I did such change in all the updaters (and generators) code I got the following results:

Sandy Bridge

count tunnel attractors fountain
171000 372.641 476.820 376.410
181000 401.705 508.353 404.176
191000 427.588 542.794 432.397

Ivy Bridge

count tunnel attractors fountain
171000 475.609 591.805 480.402
181000 502.201 620.601 512.300
191000 534.150 667.575 541.788

This is not a massive improvement, but it’s still worth testing.

Random Number Generator  

I focused mostly on the updaters part so far. But, generators could also be improved a bit. In this module a random number generator is heavily used. What if we changed it?

Right now there is standard C rand() function called. For a particle system we, probably, do not need to use something more advanced (like normal distribution random generator) - uniform distribution is just fine… maybe there are some faster generators than the default one?

I’ve searched and found something: here, here and here

I’ve tried to use this generator:

// http://www.rgba.org/articles/sfrand/sfrand.htm
static unsigned int mirand = 1;
float sfrand(void) {
    unsigned int a;
    mirand *= 16807;
    a = (mirand & 0x007fffff) | 0x40000000;
    return(*((float*)&a) - 3.0f);
}

It has a uniform distribution and 23 bits of precision (C rand() has only 16 bits).

The results:

Sandy Bridge:

count tunnel attractors fountain
171000 334.633 443.816 348.908
181000 363.954 474.477 372.739
191000 384.869 501.435 394.217

Ivy Bridge:

count tunnel attractors fountain
171000 412.172 531.586 429.293
181000 450.146 573.073 463.037
191000 473.518 606.759 484.880

Wow! Now it is around 28% of total improvement for Sandy Bridge and almost the same for Ivy Bridge.

Wrap up  

Final results

CPU count tunnel attractors fountain
Sandy 191000 384.869 (-21.3%) 501.435 (-27.2%) 394.217 (-24.2%)
Ivy 191000 473.518 (-20.3%) 606.759 (-27.1%) 484.880 (-24.3%)

Total (taking times before tools optimization):

CPU tunnel attractors fountain
Sandy 35.5% 43,5% 39,7%
Ivy 33.2% 38,2% 35,6%

We can ‘reverse’ those numbers and say that now the attractor effect runs almost two times faster! Not that bad!

Conclusion:

  • Memory alignment and proper data structures are the key factors.
  • Write SIMD code only if needed, usually it is better to rely on a compiler and third party libraries.
  • Describe your code better: for instance using __restrict keyword. That way a compiler can generate better code.
  • Random number generator can make a difference

What’s Next  

The renderer is very simple so far. Maybe, there are some options to improve its code. For sure, we have to look at CPU to GPU memory transfers and better buffers usage.

Read next: Renderer Optimizations

References