tree 5fe13cc84e517940534f55ae53e27d767c503984
parent 71ccfafa26cd4f69aed50aecb8d7ddcdaec22a4a
author Ihor Buhaiov <freedbrt@gmail.com> 1792085844 +0200
committer Ihor Buhaiov <freedbrt@gmail.com> 1792085844 +0200

Add a Neon version of the k-means centroid step

The centroid step of av1_k_means() was a static C function. Its loop
updates one counter and one sum per pixel, and when neighboring pixels
pick the same color each update waits on the store of the one before.

Move it into RTCD as av1_calc_centroids_dim1() and dim2(), the same
way as av1_calc_indices, and add a Neon version for dim1 that counts
and sums two centroids per pass, 16 pixels per step. The empty cluster
fallback and the rounding are the same as in C, so the output does not
change. Add a test that checks the Neon version against C.

On real palette inputs the Neon version is 3.9 to 4.8 times faster. On
2560x1440 screen recordings the encoder uses up to 5.5% fewer cycles.

Change-Id: I3b5e8e2c2a4e9358875377ab3297aa2f5a6c63fa