conclusion.md

Conclusions

---

Favourite?

• All of the above!!
• C++ is best when we can blend all its features

---

Performance

• Intel(R) Core(TM) i9-9980XE CPU @ 3.00GHz
• cpupower frequency-set --governor performance
• Single threaded
• GCC 9.2

---

Cornell box scene

256x256 32spp (1 sphere, 38 triangles)

Style	Render time (secs)
Object Oriented	72
Functional	82
Data-Oriented	72

---

Owl scene

128spp (100 spheres, 12 triangles)

Style	Render time (secs)
Object Oriented	96
Functional	101
Data-Oriented	64

---

Suzanne

256x256 8spp (2 spheres, 970 triangles)

Style	Render time (secs)
Object Oriented	80
Functional	119
Data-Oriented	106

What on earth!?!

---

What happened?

---

Object oriented

   252,203,645,149   instructions    #    2.71  insn per cycle         
    23,158,824,048   branches        #  845.438 M/sec                  
       139,741,218   branch-misses   #    0.60% of all branches        

Functional

   238,866,691,159   instructions    #    1.78  insn per cycle         
    21,881,070,251   branches        #  560.513 M/sec                  
     1,105,066,725   branch-misses   #    5.05% of all branches        

Data-oriented Design

   154,821,779,748   instructions    #    1.34  insn per cycle         
    10,353,392,805   branches        #  305.213 M/sec                  
     1,242,670,094   branch-misses   #   12.00% of all branches        

12% of all branches!?!

---

---

for (/* all triangles */) {
  auto u = calcU(/*...*/);
  if (u < 0 || u > 1) {
    continue;
  }
  auto v = calcV(/*...*/);
  if (v < 0 || u + v > 1) {
    continue;
  }
  auto dist = calcD(/*..*/);
  if (dist < nearest) {
    nearest = dist;
  }
}

vcomisd xmm14, xmm0 ; 0 >= u?
ja skip
vcomisd xmm0, [1.0] ; u >= 1.0?
ja skip

---

Branch Prediction

• u < 0 unpredictable
• u > 1 unpredictable
• (u < 0 || u > 1) should be predictable*
• (u < 0 || u > 1 || v < 0 || u + v > 1) more so

* Provided compiler combines conditions...
(u < 0) | (u > 1) | ...

---

auto u = calcU(/*...*/);
if (u < 0 || u > 1) {
  continue;
}
auto v = calcV(/*...*/);
if (v <  0 || u + v > 1) {
  continue;
}

auto u = calcU(/*...*/);
auto v = calcV(/*...*/);
if ((u < 0) | (u > 1) 
    | (v < 0) | (u + v > 1)) {
  continue;
}

---

FINAL STATS

Suzanne DoD: 106s → 68s (36% faster)

---

FINAL STATS

Scene	OO	FP	DoD
Cornell	72 → 64	82 → 64	72 → 56
Owl	96 → 96	101 → 100	64 → 64
Suzanne	80 → 104 (!)	119 → 82	106 → 68

---

If I had more time...

• Code on GitHub
• Threading
• Devirtualisation
• Future directions...
• DoD improvements
• Thanks

---

GO WRITE SOMETHING COOL!