Compute-PI

GPU acceleration

• Device Info

  • Device name: GeForce GTX 770
  • Warp size: 32
  • multiProcessorCount: 8

• Gnuplot command

  ​> set title 'GPU acceleration'
  ​> set ylabel 'Speed up'
  ​> set xlabel 'number of threads'
  ​> set terminal png
  ​> set output 'xxx.png'
  ​> plot "SpeedUp.txt" with points pointtype 7 notitle
  

Speed up for 12800000 slices

• 需注意設定的執行緒數量，會很明顯影響到提昇速率
• 當執行緒數量超過 6144 時，便不再有速率提昇，可參考 Amdahl's law，固定負載下能平行計算的部份有限，因此速度提昇會趨近定值

Error percentage form 1 to 20480000 slices

Time cost for 400000000 slices

• Baseline: 2.999478(s)
• OpenMP
  • 2 threads: 1.460224(s)
  • 4 threads: 0.767469(s)
• AVX: 1.283873(s)
• AVX + Loop unroll: 1.169533(s)
• 6144 GPU threads (threads per block = 256): 0.070991(s)

Source of this article: https://hackmd.io/s/SyEgyhlA