[AIROCMLIR-44] Update quick-tune lists for gemm and conv

[mirza-halilcevic]

Motivation

Update gemm and conv quick-tune lists, based on exhaustive and greedy tuning.

Technical Details

The following architectures have been updated:

• gfx950
• gfx942
• gfx90a
• gfx908
• gfx1201
• gfx1101
• gfx1150 (based on exhaustive only)

NOTE: The new quick-tune lists ended up being significantly longer in most cases, even 100+ configs for gfx1201 and gfx1101 f16 convolutions.

Test Plan

Compare performance between old and new quick-tune lists and confirm the new ones are giving better performance.

Test Result

gfx950

gemm

> 20% faster: 46 (9.9%)
5-20% faster: 147 (31.5%)
Within 5%: 257 (55.0%)
5-20% slower: 15 (3.2%)
> 20% slower: 2 (0.4%)

conv

> 20% faster: 101 (7.4%)
5-20% faster: 380 (27.9%)
Within 5%: 842 (61.8%)
5-20% slower: 40 (2.9%)
> 20% slower: 0 (0.0%)

gfx942

gemm

> 20% faster: 16 (3.4%)
5-20% faster: 124 (26.6%)
Within 5%: 297 (63.6%)
5-20% slower: 30 (6.4%)
> 20% slower: 0 (0.0%)

conv

> 20% faster: 220 (16.1%)
5-20% faster: 363 (26.6%)
Within 5%: 723 (53.0%)
5-20% slower: 56 (4.1%)
> 20% slower: 1 (0.1%)

gfx90a

gemm

> 20% faster: 30 (6.4%)
5-20% faster: 80 (17.1%)
Within 5%: 325 (69.6%)
5-20% slower: 26 (5.6%)
> 20% slower: 6 (1.3%)

conv

> 20% faster: 165 (12.1%)
5-20% faster: 355 (26.0%)
Within 5%: 789 (57.9%)
5-20% slower: 48 (3.5%)
> 20% slower: 6 (0.4%)

gfx908

gemm

> 20% faster: 15 (3.2%)
5-20% faster: 98 (21.0%)
Within 5%: 283 (60.6%)
5-20% slower: 62 (13.3%)
> 20% slower: 9 (1.9%)

conv

> 20% faster: 115 (8.4%)
5-20% faster: 312 (22.9%)
Within 5%: 895 (65.7%)
5-20% slower: 41 (3.0%)
> 20% slower: 0 (0.0%)

gfx1201

gemm

> 20% faster: 101 (21.6%)
5-20% faster: 75 (16.1%)
Within 5%: 274 (58.7%)
5-20% slower: 17 (3.6%)
> 20% slower: 0 (0.0%)

conv

> 20% faster: 63 (4.6%)
5-20% faster: 120 (8.8%)
Within 5%: 1107 (81.2%)
5-20% slower: 71 (5.2%)
> 20% slower: 2 (0.1%)

gfx1101

gemm

> 20% faster: 13 (2.8%)
5-20% faster: 146 (31.3%)
Within 5%: 288 (61.7%)
5-20% slower: 20 (4.3%)
> 20% slower: 0 (0.0%)

conv

> 20% faster: 137 (10.1%)
5-20% faster: 345 (25.3%)
Within 5%: 784 (57.5%)
5-20% slower: 96 (7.0%)
> 20% slower: 1 (0.1%)

gfx1150

gemm

> 20% faster: 45 (9.6%)
5-20% faster: 105 (22.5%)
Within 5%: 236 (50.5%)
5-20% slower: 56 (12.0%)
> 20% slower: 25 (5.4%)

conv

> 20% faster: 101 (7.4%)
5-20% faster: 260 (19.1%)
Within 5%: 741 (54.4%)
5-20% slower: 180 (13.2%)
> 20% slower: 81 (5.9%)

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.