Remove sortBump atomicAdd

[jayshah1819]

• Removed the old sortBump counter since the atomicAdd there was slowing the
  kernel down a lot. Now the workgroup ID just comes from builtinsUniform.wgid.x,
  so we don’t hit any global atomics. Also cleared the hist/passHist before each run.

• This improves bandwidth graph.

• Fix plot stroke=timing for proper GPU/CPU lines; expand input range + trials.

[jowens]

Now the workgroup ID just comes from builtinsUniform.wgid.x, so we don’t hit any global atomics.

We can't do this. We could lock up the machine with this. We have no guarantee that issue order on workgroups is in any way correlated with wgid.

[jowens]

Consider the case that the hardware issued all the blocks with the largest wgid first and all the small wgids were not issued and were waiting on the larger wgids to finish. If we didn't have fallback in this code, it would lock up for sure (because the resident workgroups would be waiting for workgroup 0 to finish, but workgroup 0 would not have launched yet). This particular code will not lock up, because we do have fallback, but if we fallback, performance will be terrible.

[jowens]

tl;dr: We need to make this work with sortBump.

[jowens]

See: first paragraph of 2.3.3: https://escholarship.org/uc/item/0bk9z4bt