AltiVec vp9

[classilla]

Some of the idcts and other components in vp9 look like they were written to aid easy vectorization. Assuming the compiler isn't doing that already, these look like good targets for gradual conversion.

[classilla]

Converted:

• vp9_convolve_avg_vmx
• vp9_convolve_copy_vmx
• vp9_idct16x16_1_add_vmx
• vp9_idct32x32_1_add_vmx
• vp9_idct4x4_16_add_vmx
• vp9_idct4x4_1_add_vmx
• vp9_idct8x8_1_add_vmx
• vp9_iht4x4_16_add_vmx

Also killed the loop filter, which helped quite a bit as well.

[classilla]

Whew!
Next intraframe predictors.

idct.txt

[barracuda156]

@classilla Could this be ported to modern libvpx? It dropped its own Altivec-without-VSX support several versions ago :(

[classilla]

Probably not trivially because I suspect much of the underlying structs have changed.

[barracuda156]

Probably not trivially because I suspect much of the underlying structs have changed.

And there is no known way to emulate VSX on Apple Altivec, right?

[classilla]

Not in any efficient way. VSX basically uses (and expands) FPRs. Any equivalent code without it would be larger and much slower.