Totsu

Totsu copied to clipboard

Benchmark, profile and optimize it to speed up.

https://github.com/convexbrain/Totsu/releases/tag/totsu_f32cuda_v0.1.0

A benchmark result of LP

• https://github.com/convexbrain/Totsu/tree/1f5200599ffd8bdf15e6ce672bcc1c2f0bbc11bb/experimental/benchmark_lp
• F32CUDA is faster than FloatGeneric.

• CPU
  • Intel(R) Core(TM) i7-8700 CPU @ 3.20GHz
  • RAM: 32.0 GB
• GPU
  • NVIDIA GeForce RTX 3070
  • CUDA core: 5888
  • Core clock: 1725 MHz
  • Memory bandwidth: 448.06 GB/s
  • Memory: 8192 MB GDDR6

A benchmark result of QP

• https://github.com/convexbrain/Totsu/tree/884e36b4fd32d696ddca046af755ad8a2d120a61/experimental/benchmark_qp
• F32CUDA is slower than FloatGeneric. 😭

Proceed to profiling using this benchmark.

A profiling result of QP benchmark

• Many memory accesses are occurring when projecting onto the cone.

https://github.com/convexbrain/Totsu/tree/b56407463b691a3f2418510bc43e8a72d5186fc1/experimental/benchmark_qp

• CUDA-izing projection onto cones as much as possible.
• 200 vars (100 primals, 100 duals).

• 400 vars (200 primals, 200 duals).

https://github.com/convexbrain/Totsu/tree/77f0e5cc10e7a2d29567352f88135a99ed620be1/experimental/benchmark_qp

• FxHashMap instead of HashMap.
• 200 vars (100 primals, 100 duals).

https://github.com/convexbrain/Totsu/tree/13b8d378f79445c53b9c9f77fbf4389029423d12/experimental/benchmark_qp

• Intermittent criteria checks.
• 200 vars (100 primals, 100 duals).

• The effect of CUDA comes out from about 800 variables.
• The number of iterations is not monotonically increasing; probably because those QPs are generated with random numbers.
• In the first place, the number of iterations is too large.