Fix Issue 6657 - dotProduct overload for small fixed size arrays

[n8sh]

Benchmark using DMD, 10 million trials:

version	arg type	compiler	time (ms)
old	float[3]	dmd -O	118.503
new	float[3]	dmd -O	64.535
old	float[4]	dmd -O	120.456
new	float[4]	dmd -O	74.575
old	float[5]	dmd -O	141.211
new	float[5]	dmd -O	99.659
old	float[6]	dmd -O	151.537
new	float[6]	dmd -O	107.559
old	float[7]	dmd -O	192.009
new	float[7]	dmd -O	137.358
old	float[8]	dmd -O	176.876
new	float[8]	dmd -O	147.655
old	float[9]	dmd -O	200.092
new	float[9]	dmd -O	167.429
old	float[16]	dmd -O	318.174
new	float[16]	dmd -O	288.255

Benchmark using LDC, 10 million trials:

version	arg type	compiler	time (ms)
old	float[3]	ldc -O2	55.215
new	float[3]	ldc -O2	17.219
old	float[4]	ldc -O2	35.801
new	float[4]	ldc -O2	20.170
old	float[5]	ldc -O2	53.151
new	float[5]	ldc -O2	20.194
old	float[6]	ldc -O2	62.134
new	float[6]	ldc -O2	26.126
old	float[7]	ldc -O2	67.609
new	float[7]	ldc -O2	26.335
old	float[8]	ldc -O2	47.428
new	float[8]	ldc -O2	25.960
old	float[9]	ldc -O2	66.051
new	float[9]	ldc -O2	29.275
old	float[16]	ldc -O2	73.301
new	float[16]	ldc -O2	45.106

[dlang-bot]

Thanks for your pull request, @n8sh!

Bugzilla references

Auto-close	Bugzilla	Severity	Description
✓	6657	enhancement	dotProduct overload for small fixed size arrays

Testing this PR locally

If you don't have a local development environment setup, you can use Digger to test this PR:

dub fetch digger
dub run digger -- build "master + phobos#6990"

[n8sh]

double benchmark.

Benchmark using DMD, 10 million trials:

version	arg type	compiler	time (ms)
old	double[3]	dmd -O	286.851
new	double[3]	dmd -O	167.054
old	double[4]	dmd -O	280.592
new	double[4]	dmd -O	178.395
old	double[8]	dmd -O	342.528
new	double[8]	dmd -O	243.230
old	double[9]	dmd -O	388.862
new	double[9]	dmd -O	263.507
old	double[16]	dmd -O	496.451
new	double[16]	dmd -O	353.159

Benchmark using LDC, 10 million trials:

version	arg type	compiler	time (ms)
old	double[3]	ldc -O2	59.573
new	double[3]	ldc -O2	18.770
old	double[4]	ldc -O2	38.978
new	double[4]	ldc -O2	26.844
old	double[8]	ldc -O2	55.563
new	double[8]	ldc -O2	32.187
old	double[9]	ldc -O2	73.577
new	double[9]	ldc -O2	35.114
old	double[16]	ldc -O2	83.846
new	double[16]	ldc -O2	52.690

[thewilsonator]

Did you do -O3 results for LDC?

[n8sh]

float old version ldc2 -O3

version	arg type	compiler	time (ms)
old	float[2]	ldc -O3	48.728
old	float[3]	ldc -O3	58.374
old	float[4]	ldc -O3	32.802
old	float[5]	ldc -O3	45.219
old	float[6]	ldc -O3	51.883
old	float[7]	ldc -O3	65.833
old	float[8]	ldc -O3	45.824
old	float[9]	ldc -O3	58.899
old	float[10]	ldc -O3	68.776
old	float[11]	ldc -O3	70.097
old	float[12]	ldc -O3	52.271
old	float[13]	ldc -O3	71.376
old	float[14]	ldc -O3	74.872
old	float[15]	ldc -O3	82.841
old	float[16]	ldc -O3	68.458

float new version ldc2 -O3

version	arg type	compiler	time (ms)
new	float[2]	ldc -O3	15.801
new	float[3]	ldc -O3	19.864
new	float[4]	ldc -O3	14.314
new	float[5]	ldc -O3	20.084
new	float[6]	ldc -O3	20.095
new	float[7]	ldc -O3	28.606
new	float[8]	ldc -O3	23.166
new	float[9]	ldc -O3	30.785
new	float[10]	ldc -O3	31.497
new	float[11]	ldc -O3	40.093
new	float[12]	ldc -O3	35.538
new	float[13]	ldc -O3	36.830
new	float[14]	ldc -O3	35.538
new	float[15]	ldc -O3	44.962
new	float[16]	ldc -O3	36.207

Had to tweak the test a bit for int because LDC was too clever about optimizing away work. In the integer case by N=16 there is no longer a benefit with LDC but with DMD there still is.

int old version ldc2 -O3

version	arg type	compiler	time (ms)
old	int[2]	ldc -O3	52.763
old	int[3]	ldc -O3	70.344
old	int[4]	ldc -O3	57.306
old	int[5]	ldc -O3	69.355
old	int[6]	ldc -O3	79.647
old	int[7]	ldc -O3	92.875
old	int[8]	ldc -O3	74.591
old	int[9]	ldc -O3	80.862
old	int[10]	ldc -O3	87.535
old	int[11]	ldc -O3	107.287
old	int[12]	ldc -O3	88.301
old	int[13]	ldc -O3	105.696
old	int[14]	ldc -O3	105.883
old	int[15]	ldc -O3	117.358
old	int[16]	ldc -O3	107.997

int new version ldc2 -O3

version	arg type	compiler	time (ms)
new	int[2]	ldc -O3	19.510
new	int[3]	ldc -O3	24.590
new	int[4]	ldc -O3	24.458
new	int[5]	ldc -O3	29.813
new	int[6]	ldc -O3	37.051
new	int[7]	ldc -O3	37.410
new	int[8]	ldc -O3	65.825
new	int[9]	ldc -O3	68.708
new	int[10]	ldc -O3	64.281
new	int[11]	ldc -O3	70.540
new	int[12]	ldc -O3	67.946
new	int[13]	ldc -O3	80.485
new	int[14]	ldc -O3	101.257
new	int[15]	ldc -O3	112.384
new	int[16]	ldc -O3	128.460

int old version dmd -O

version	arg type	compiler	time (ms)
old	int[2]	dmd -O	120.922
old	int[3]	dmd -O	138.149
old	int[4]	dmd -O	147.788
old	int[5]	dmd -O	150.252
old	int[6]	dmd -O	181.878
old	int[7]	dmd -O	192.617
old	int[8]	dmd -O	204.361
old	int[9]	dmd -O	215.713
old	int[10]	dmd -O	231.395
old	int[11]	dmd -O	248.746
old	int[12]	dmd -O	240.910
old	int[13]	dmd -O	274.690
old	int[14]	dmd -O	287.599
old	int[15]	dmd -O	310.941
old	int[16]	dmd -O	330.701

int new version dmd -O

version	arg type	compiler	time (ms)
new	int[2]	dmd -O	78.174
new	int[3]	dmd -O	105.965
new	int[4]	dmd -O	114.752
new	int[5]	dmd -O	129.529
new	int[6]	dmd -O	133.568
new	int[7]	dmd -O	149.295
new	int[8]	dmd -O	186.759
new	int[9]	dmd -O	194.025
new	int[10]	dmd -O	207.918
new	int[11]	dmd -O	212.257
new	int[12]	dmd -O	229.368
new	int[13]	dmd -O	250.924
new	int[14]	dmd -O	270.067
new	int[15]	dmd -O	281.089
new	int[16]	dmd -O	305.268

[thewilsonator]

Very nice! Please merge when you are happy with it.

[PetarKirov]

Nice! Can you add a unit test to prevent possible future regressions?

[n8sh]

You mean a performance benchmark?

[PetarKirov]

No, just a simple test that calls the function with static arrays of even and odd length.
Performance regressions testing is a very good idea in general, but we don't have any infrastructure to support such.

[n8sh]

@ZombineDev makes sense.