Turns out Clang's a lot better at auto-vectorizing "obvious" scalar code
into obvious vector code when it's written out the long way, e.g.
F32x4 x = ...;
x = { sqrtf(x[0]), sqrtf(x[1]), sqrtf(x[2]), sqrtf(x[3]) };
vectorizes into sqrtps a lot more reliably than our recurse-onto-scalars
strategy, and also better than the other naive approach,
F32x4 x = ...;
for (int i = 0; i < 4; i++) { x[i] = sqrtf(x[i]); }
So here I've added a map(V, fn) -> V' using C++14 tricks to let the
compiler handle the expansion of x = { fn(x[0]), fn(x[1]), ...
fn(x[N-1]) } for any N, and implemented most skvx scalar fallback code
using that.
With these now vectorizing well at any N, we can remove any
specializations we'd written for particular N, really tidying up.
Over in the SkVM interpreter, this is a big improvement for ceil and
floor, which were being done 2 floats at a time instead of 8. They're
now slimmed way down to
shlq $6, %r13
vroundps $K, (%r12,%r13), %ymm0
vroundps $K, 32(%r12,%r13), %ymm1
jmp ...
where K is 9 or 10 depending on the op.
I haven't found a scalar function that Clang will vectorize to vcvtps2pd
(the rounding one, not truncating vcvttps2pd), so I've kept lrint()
written the long way, updated to the style I've been using lately with
specializations inline.
Change-Id: Ia97abe3c876008228bf62b1daacd6f6140408fc4
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/317375
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
4985db413d