The original lepton library used an order of the 64 coefficients that was
0..48 middle (7x7)
49 DC (0,0)
50 - 56 upper row (7)
57 - 63 left column (7)
This was designed to simplify some of the operations, but it turns out that it was more expensive to move back and forth between the aligned format and the normal JPEG block layout, especially since after vectorizing the operations many of the advantages of the layout disappeared for the middle blocks.
The main benefit goes to the IDCT operations which no longer need to shuffle the coordinates around, which speeds up the IDCT operations by about 30%
Also added AVX2 version to build, which is about 20% faster overall
The original lepton library used an order of the 64 coefficients that was
0..48 middle (7x7) 49 DC (0,0) 50 - 56 upper row (7) 57 - 63 left column (7)
This was designed to simplify some of the operations, but it turns out that it was more expensive to move back and forth between the aligned format and the normal JPEG block layout, especially since after vectorizing the operations many of the advantages of the layout disappeared for the middle blocks.
The main benefit goes to the IDCT operations which no longer need to shuffle the coordinates around, which speeds up the IDCT operations by about 30%
Also added AVX2 version to build, which is about 20% faster overall