Expand description
Routines for IDCT
Essentially we provide 2 routines for IDCT, a scalar implementation and a not super optimized AVX2 one, i’ll talk about them here.
There are 2 reasons why we have the avx one
- No one compiles with -C target-features=avx2 hence binaries won’t probably take advantage(even if it exists).
- AVX employs zero short circuit in a way the scalar code cannot employ it.
- AVX does this by checking for MCU’s whose 63 AC coefficients are zero and if true, it writes values directly, if false, it goes the long way of calculating.
- Although this can be trivially implemented in the scalar version, it generates code
I’m not happy width(scalar version that basically loops and that is too many branches for me)
The avx one does a better job of using bitwise or’s with (
_mm256_or_si256) which is magnitudes of faster than anything I could come up with
The AVX code also has some cool transpose_u16 instructions which look so complicated to be cool (spoiler alert, i barely understand how it works, that’s why I credited the owner).
Modules§
Functions§
- choose_
idct_ func - Choose an appropriate IDCT function