From quadratic to linear complexity.

[lordnn]

I've lose my fight vs git and have opened new PR.

[lordnn]

Multithreading may be difficult, Let's try once again:
For example - 3 threads (A, B, C).

Thread A: finding loop 0 (NOT_ENCODED) -> process tile 0
Thread C: finding loop 0 (ON_ENCODING), 1 (NOT_ENCODED) -> process tile 1
Thread B: finding loop 0 (ON_ENCODING), 1 (ON_ENCODING), 2 (NOT_ENCODED) -> process tile 2
// here it doesn't matter what thread will finish first, because next NOT_ENCODED tile has index 3. Let it be C.
Thread C: finding loop 0 (ON_ENCODING), 1 (ENCODED), 2 (ON_ENCODING), 3 (NOT_ENCODED) -> process tile 3

And so on. Tiles processing start in monotonic order. So ctx->tile_idx is an index of the next NOT_ENCODED tile.

Regarding new variable to ctx - you don't need at all oapve_tile_t.stat variable.
So you may save sizeof(s32) * OAPV_MAX_TILES number of bytes, at least for encoder case.

P.S. SATD calculation has constant time (not a funtion of number of tiles). Quadratic will alway beat constant, you only need to increase number of tiles. And finding loop leads to CPU-cache pollution.

[kpchoi]

Adopted in #247 with the original commit authorship preserved. volatile and the critical section around the completion status updates are kept. Closing in favor of #247.

[lordnn]

oapvd_tile->bs_beg used in absolutely the same manner as oapvd_tile->stat - update in one thread and catch updated value in another. But no volatile for bs_beg.

Because volatile IS NOT ABOUT MULTITHREADS SYNCHRONIZATION!!!

volatile is about MMIO-registers in device driver development.