Combine similar kernels using cooperative groups

[huiyuxie]

Several kernels are launched separately to achieve grid-wide synchronization. Cooperative groups can help combine these similar kernels into a single kernel launch to improve performance. Also, kernel size configurators for 1D, 2D, and 3D cooperative kernel launches are added.

Tasks:

• cuda_prolong2mortars! for 3D
• cuda_mortar_flux! with nonconservative_terms::False for 3D
• cuda_mortar_flux! with nonconservative_terms::False for 3D

[huiyuxie]

Here we use Euler mortar 3D as an example for benchmarking.

cuda_prolong2mortars! before kernel combination

BenchmarkTools.Trial: 10000 samples with 1 evaluation.
Range (min … max):   67.100 μs …  48.753 ms  ┊ GC (min … max): 0.00% … 22.12%
Time  (median):     111.200 μs               ┊ GC (median):    0.00%
Time  (mean ± σ):   152.253 μs ± 647.446 μs  ┊ GC (mean ± σ):  1.41% ±  0.34%

    ▁▄▇█▅▁
 ▃▆███████▆▆▅▅▄▃▂▂▂▂▂▂▂▂▁▁▁▁▂▃▂▃▂▂▂▂▂▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁ ▂
 67.1 μs          Histogram: frequency by time          453 μs <

Memory estimate: 15.75 KiB, allocs estimate: 217.

cuda_prolong2mortars! after kernel combination

BenchmarkTools.Trial: 10000 samples with 1 evaluation.
 Range (min … max):   76.600 μs …  1.312 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     112.900 μs              ┊ GC (median):    0.00%
 Time  (mean ± σ):   122.359 μs ± 49.262 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

   ▁█▁    ▁  ▂▁
  ▁███▆▇▇██████▇▅▄▃▃▃▃▂▃▃▂▂▂▂▂▂▂▂▂▂▁▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁ ▂
  76.6 μs         Histogram: frequency by time          300 μs <

 Memory estimate: 12.66 KiB, allocs estimate: 181.

[huiyuxie]

Here we use Euler mortar 3D as an example for benchmarking.

cuda_mortar_flux! with nonconservative_terms::False before kernel combination

BenchmarkTools.Trial: 4405 samples with 1 evaluation.
 Range (min … max):  296.400 μs … 132.412 ms  ┊ GC (min … max): 0.00% … 17.08%
 Time  (median):     980.600 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):     1.128 ms ±   4.445 ms  ┊ GC (mean ± σ):  0.46% ±  0.26%

   █▅                       ▃▄▃▂
  ▃██▆▇▆▄▄▃▁▂▁▁▁▁▁▁▁▁▁▁▂▂▃▄▅█████▆▅▅▄▃▂▂▂▂▂▁▁▁▁▁▁▂▂▂▁▂▁▂▂▁▁▁▁▁▁ ▂
  296 μs           Histogram: frequency by time         1.87 ms <

 Memory estimate: 20.28 KiB, allocs estimate: 265.

cuda_mortar_flux! with nonconservative_terms::False after kernel combination

BenchmarkTools.Trial: 10000 samples with 1 evaluation.
 Range (min … max):  124.800 μs …  55.918 ms  ┊ GC (min … max): 0.00% … 21.18%
 Time  (median):     231.250 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   389.808 μs ± 786.346 μs  ┊ GC (mean ± σ):  0.54% ±  0.28%

  █▃▆▂
  ████▃▂▂▂▂▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▃▄▄▃▃▃▃▃▃▃▃▂▂▂▂▂▂▂▂▁▁ ▂
  125 μs           Histogram: frequency by time          819 μs <

 Memory estimate: 16.00 KiB, allocs estimate: 218.

[huiyuxie]

Here we use MHD mortar 3D as an example for benchmarking.

cuda_mortar_flux! with nonconservative_terms::False before kernel combination

BenchmarkTools.Trial: 3956 samples with 1 evaluation.
 Range (min … max):  746.100 μs …  33.295 ms  ┊ GC (min … max): 0.00% … 21.74%
 Time  (median):       1.374 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):     1.259 ms ± 644.034 μs  ┊ GC (mean ± σ):  0.15% ±  0.35%

     █▄▁    ▁
  ▃▆████▆▆▇▇█▅▄▃▃▂▂▂▂▂▂▂▁▂▂▁▂▁▂▁▁▃▆▆▅▆▇▆▅▅▅▅▃▃▃▃▄▆▆▆▆▆▅▄▄▄▄▃▃▃▂ ▄
  746 μs           Histogram: frequency by time         1.92 ms <

 Memory estimate: 20.81 KiB, allocs estimate: 268.

cuda_mortar_flux! with nonconservative_terms::False after kernel combination

 BenchmarkTools.Trial: 5731 samples with 1 evaluation.
 Range (min … max):  403.400 μs …  37.955 ms  ┊ GC (min … max): 0.00% … 21.76%
 Time  (median):     892.700 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   864.990 μs ± 731.922 μs  ┊ GC (mean ± σ):  0.28% ±  0.37%

   █▃▂ ▁▄▅▂                  ▂        ▁
  ▇███▇█████▇▆▅▄▄▄▃▃▃▂▂▂▁▂▁▁▄█▇▇██▇▇▆██▇▄▄▃▂▃▂▃▄▅▅▅▇▆▇▆▆▇▇▆▅▄▃▂ ▄
  403 μs           Histogram: frequency by time         1.45 ms <

 Memory estimate: 16.34 KiB, allocs estimate: 221.