Use logf and powf intrinsics

[sgt0]

This handles edge cases like a pow base of 0 better than the exp(log(x) * y) identity did.

Fixes #11

[yuygfgg]

It seems that this PR is calling cmath functions for log and pow operations instead of using LLVM intrinsics? It might negatively impact the performance since function callings are very hard to vectorize.

[sgt0]

I just followed what upstream swiftshader was doing. Since your version performs better and it's a smaller diff, we can go with it instead.

clip = core.std.BlankClip(None, 1920, 1080, vs.GRAY16, 30000)

set_output(core.std.Expr(clip, "x 1.5 pow"), "std")
set_output(core.akarin.Expr(clip, "x 1.5 pow"), "akarin")

• std: Output 30000 frames in 10.10 seconds (2970.32 fps)
• your PR: Output 30000 frames in 10.04 seconds (2989.49 fps)
• my PR: Output 30000 frames in 11.15 seconds (2691.45 fps)

[Mr-Z-2697]

the speed is worse than halved using the expression from original issue

core.std.BlankClip(width=1920,height=1080,format=vs.GRAY16,length=1e5).std.Expr('x 32768 / 0.86 pow 65535 *').set_output()

std: Output 100000 frames in 29.47 seconds (3392.89 fps)
pr: Output 100000 frames in 67.19 seconds (1488.38 fps)
(both pr have similar speed)

[sgt0]

Seems like it comes down to the exponent being between 0 and 1 (i.e. if I make it 1.86 pow instead, I get 13% better perf)? So I guess now we know why the plugin wasn't using the intrinsics. But, the upstream plugin is outputting incorrect values for that expr (for clips that aren't just blank), so IMO its better perf is moot.

[sgt0]

Erm, idk why I brought up upstream when we're comparing to std. Anyways, closing this PR since I think we either take yuygfgg's or try to improve the current pow/log implementation.