I tested chunk sizes 8 and 4, and found them to be much slower than with chunk size 16. A minor tweak to the way that the number of workgroups is calculated for the SMVP shader is needed for these tests to be run. I manually changed the chunk_size variable and ran the benchmark suite to get the following results.
Chunk size: 8
MSM size
1st run
Run 1
Run 2
Run 3
Run 4
Run 5
Average (incl 1st)
Average (excl 1st)
2^16
24303
2052
2075
1952
2142
2011
5756
2046
2^17
3201
3310
3164
3193
3300
3164
3222
3226
2^18
6669
5695
5565
5689
5712
5703
5839
5673
2^19
11392
10419
10368
10404
10361
10372
10553
10385
2^20
20680
19639
19660
19672
19757
19666
19846
19679
Chunk size: 4
2 16: 23820ms
2 17: 45787ms
The results are clear, so I didn't run more benchmarks
I tested chunk sizes 8 and 4, and found them to be much slower than with chunk size 16. A minor tweak to the way that the number of workgroups is calculated for the SMVP shader is needed for these tests to be run. I manually changed the
chunk_sizevariable and ran the benchmark suite to get the following results.Chunk size: 8
243032052207519522142201157562046320133103164319333003164322232266669569555655689571257035839567311392104191036810404103611037210553103852068019639196601967219757196661984619679Chunk size: 4
2 16: 23820ms 2 17: 45787ms
The results are clear, so I didn't run more benchmarks