To Replicate

1. Checkout the correct branch pytorch/pytorch#62450
2. Build PyTorch
3. Clone this repo, and run make init in the same directory. Also make sure cmake is looking at the libs for the correct PyTorch (that you just built).
4. Run sudo cpupower frequency-set --governor performance to disable CPU scaling if necessary (I haven't observed a big difference from this)

Results:

----------- Baseline -----------
2021-08-11T19:56:42-04:00
Running ./example-app-baseline
Run on (32 X 3592.81 MHz CPU s)
CPU Caches:
  L1 Data 32 KiB (x16)
  L1 Instruction 32 KiB (x16)
  L2 Unified 512 KiB (x16)
  L3 Unified 16384 KiB (x4)
Load Average: 0.71, 0.63, 0.63
-------------------------------------------------------------------
Benchmark                         Time             CPU   Iterations
-------------------------------------------------------------------
op_call_no_requires_grad        699 ns          699 ns       962985
op_call                         679 ns          679 ns      1002686
op_call_boxed                   736 ns          736 ns       961718
----------- Templated -----------
2021-08-11T19:56:45-04:00
Running ./example-app-templated
Run on (32 X 3598.34 MHz CPU s)
CPU Caches:
  L1 Data 32 KiB (x16)
  L1 Instruction 32 KiB (x16)
  L2 Unified 512 KiB (x16)
  L3 Unified 16384 KiB (x4)
Load Average: 0.71, 0.63, 0.63
-------------------------------------------------------------------
Benchmark                         Time             CPU   Iterations
-------------------------------------------------------------------
op_call_no_requires_grad        847 ns          847 ns       820522
op_call                        1659 ns         1659 ns       421860
op_call_boxed                  1707 ns         1707 ns       404835
----------- Boxed -----------
2021-08-11T19:56:48-04:00
Running ./example-app-boxed
Run on (32 X 4155.37 MHz CPU s)
CPU Caches:
  L1 Data 32 KiB (x16)
  L1 Instruction 32 KiB (x16)
  L2 Unified 512 KiB (x16)
  L3 Unified 16384 KiB (x4)
Load Average: 0.73, 0.64, 0.63
-------------------------------------------------------------------
Benchmark                         Time             CPU   Iterations
-------------------------------------------------------------------
op_call_no_requires_grad        910 ns          910 ns       759083
op_call                        1725 ns         1725 ns       405478
op_call_boxed                  1715 ns         1715 ns       408859