Refactor the implementation

[IvanUkhov]

macOS, ARM64

Main:

test compute_0001000 ... bench:       1,696 ns/iter (+/- 12)
test compute_0010000 ... bench:      16,578 ns/iter (+/- 162)
test compute_0100000 ... bench:     164,988 ns/iter (+/- 2,065)
test compute_1000000 ... bench:   1,649,850 ns/iter (+/- 10,186)
test context_0001000 ... bench:       1,697 ns/iter (+/- 16)
test context_0010000 ... bench:      16,582 ns/iter (+/- 203)
test context_0100000 ... bench:     165,100 ns/iter (+/- 2,997)
test context_1000000 ... bench:   1,652,079 ns/iter (+/- 39,340)

This PR:

test compute_0001000 ... bench:       2,088 ns/iter (+/- 26)
test compute_0010000 ... bench:      20,554 ns/iter (+/- 1,331)
test compute_0100000 ... bench:     204,522 ns/iter (+/- 4,461)
test compute_1000000 ... bench:   2,045,050 ns/iter (+/- 8,309)
test context_0001000 ... bench:       2,226 ns/iter (+/- 17)
test context_0010000 ... bench:      21,832 ns/iter (+/- 449)
test context_0100000 ... bench:     217,379 ns/iter (+/- 6,678)
test context_1000000 ... bench:   2,173,283 ns/iter (+/- 18,401)

Linux, AMD64

Main:

test compute_0001000 ... bench:       2,370 ns/iter (+/- 19)
test compute_0010000 ... bench:      23,144 ns/iter (+/- 4,352)
test compute_0100000 ... bench:     230,255 ns/iter (+/- 1,556)
test compute_1000000 ... bench:   2,299,363 ns/iter (+/- 6,198)
test context_0001000 ... bench:       2,368 ns/iter (+/- 32)
test context_0010000 ... bench:      23,136 ns/iter (+/- 101)
test context_0100000 ... bench:     230,174 ns/iter (+/- 760)
test context_1000000 ... bench:   2,302,334 ns/iter (+/- 8,662)

This PR:

test compute_0001000 ... bench:       1,790 ns/iter (+/- 14)
test compute_0010000 ... bench:      17,532 ns/iter (+/- 106)
test compute_0100000 ... bench:     174,332 ns/iter (+/- 572)
test compute_1000000 ... bench:   1,745,888 ns/iter (+/- 5,770)
test context_0001000 ... bench:       2,163 ns/iter (+/- 13)
test context_0010000 ... bench:      21,208 ns/iter (+/- 274)
test context_0100000 ... bench:     211,190 ns/iter (+/- 2,317)
test context_1000000 ... bench:   2,037,844 ns/iter (+/- 5,788)

[IvanUkhov]

@tomhamiltonlambda, not sure how to reconcile this. This morning tried on a laptop with a different architecture and go numbers that contradict the ones we have been seeing so far. What kind of hardware do you have?

[tomhamiltonlambda]

I run on Windows, Intel processor.

Looks like it's pretty inconclusive either direction - a separate point is that the code itself is now a lot clearer - I'd like to avoid going back on that. I can spinning up various EC2 instances and compare speed there. What else do you want to do?

[IvanUkhov]

Yeah, it is pointing in different directions. Thanks to you, the new code looks much cleaner and more intelligible, and it would be nice to retain it. I will keep it open for now and occasionally try to see if I can figure out how to avoid performance degradation. But please do not spent more time on this. Thank you again!

[sammysheep]

M4 Max (unbinned model), MacOS 15.1.1

Main:

test compute_0001000 ... bench:       1,236.41 ns/iter (+/- 82.48)
test compute_0010000 ... bench:      12,129.44 ns/iter (+/- 908.35)
test compute_0100000 ... bench:     120,752.77 ns/iter (+/- 1,339.80)
test compute_1000000 ... bench:   1,205,737.50 ns/iter (+/- 96,453.37)
test context_0001000 ... bench:       1,237.54 ns/iter (+/- 27.23)
test context_0010000 ... bench:      12,136.26 ns/iter (+/- 130.81)
test context_0100000 ... bench:     120,863.10 ns/iter (+/- 9,001.46)
test context_1000000 ... bench:   1,208,620.80 ns/iter (+/- 41,977.97)

This PR:

test compute_0001000 ... bench:       1,482.34 ns/iter (+/- 13.49)
test compute_0010000 ... bench:      14,654.60 ns/iter (+/- 178.37)
test compute_0100000 ... bench:     145,970.15 ns/iter (+/- 2,459.80)
test compute_1000000 ... bench:   1,460,958.30 ns/iter (+/- 20,525.79)
test context_0001000 ... bench:       1,499.25 ns/iter (+/- 25.57)
test context_0010000 ... bench:      14,776.59 ns/iter (+/- 1,921.14)
test context_0100000 ... bench:     146,528.45 ns/iter (+/- 1,815.32)
test context_1000000 ... bench:   1,465,783.30 ns/iter (+/- 11,994.10)

rustc 1.85.0-nightly (7db7489f9 2024-11-25)