Simplify tests.

[maleadt]

• use @test + ≈ for automatic isapprox
• use mul! directly in order to perform a mixed-mode multiplication on the CPU too

That last point slows down the tests a bit, but I think it's a better comparison. @thomasfaingnaert, was there a reason you compared against multiplications that were done in the output type instead of the input type? Maybe we should just scale down the very big (2048x2048x2048) tests.

[maleadt]

Benchmark results for commit 6abd7c1 (comparing to fa335a1):

ID	before	after	change
["wmma", "Float16*Float16=Float32 (N=16384, A=N, B=T)"]	324.529 ms ± 7.443 ms	334.610 ms ± 1.392 ms	-6.1% ❌

[codecov]

Codecov Report

Patch and project coverage have no change.

Comparison is base (a7f6021) 30.35% compared to head (6abd7c1) 30.35%.

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #124   +/-   ##
=======================================
  Coverage   30.35%   30.35%           
=======================================
  Files          11       11           
  Lines         784      784           
=======================================
  Hits          238      238           
  Misses        546      546           

☔ View full report in Codecov by Sentry.
📢 Do you have feedback about the report comment? Let us know in this issue.

[thomasfaingnaert]

@thomasfaingnaert, was there a reason you compared against multiplications that were done in the output type instead of the input type?

Not sure. I don't remember any particular reason, in any case.