[sldev] [patch] OpenJPEG fix_mul() optimization

[Dzonatas]

alissa_sabre at yahoo.co.jp wrote:
> > Attached is a simple change that gives up to 10% extra performance in
> > DWT time.
>
> Hmm. If this fix_mul() is so important, why don't you make it as
> follows? It's simpler, does same thing, and should be slightly faster
> in general...
>
> static INLINE int fix_mul(int a, int b) {
> return (int) (((int64) a * (int64) b + 4096) >> 13);
> }
>
Very good.

I tested it, and it gave me the same output. However, it gave me these
times:

[INFO] tile 1 of 1
[INFO] - tiers-1 took 1.640000 s
[INFO] - dwt took 10.250000 s
[INFO] - tile decoded in 15.630000 s
Generated Outfile bird.bmp
% cumulative self self total
time seconds seconds calls s/call s/call name
49.04 4.36 4.36 3 1.45 2.32 dwt_decode_tile
29.13 6.95 2.59 29760 0.00 0.00 dwt_decode_1_real
5.74 7.46 0.51 1 0.51 0.51 imagetobmp
5.62 7.96 0.50 1 0.50 8.36 tcd_decode_tile
5.17 8.42 0.46 1 0.46 0.46 mct_decode_real
% cumulative self self total
time seconds seconds calls s/call s/call name
49.59 13.24 13.24 9 1.47 2.32 dwt_decode_tile
28.58 20.87 7.63 89280 0.00 0.00 dwt_decode_1_real
6.63 22.64 1.77 3 0.59 8.48 tcd_decode_tile
5.28 24.05 1.41 3 0.47 0.47 mct_decode_real
4.57 25.27 1.22 3 0.41 0.41 imagetobmp


The second column second row of each is the one to look at (6.95s total
on a single run and 20.87s total after 3 runs)

Here is this version:

static INLINE int fix_mul(int a, int b) {
int64 temp = (int64) a * (int64) b ;
return (int) ((temp + (temp & 4096)) >> 13) ;
}

The results:

[INFO] tile 1 of 1
[INFO] - tiers-1 took 1.590000 s
[INFO] - dwt took 9.790000 s
[INFO] - tile decoded in 15.110000 s
Generated Outfile bird.bmp
% cumulative self self total
time seconds seconds calls s/call s/call name
52.78 4.56 4.56 3 1.52 2.20 dwt_decode_tile
23.50 6.59 2.03 29760 0.00 0.00 dwt_decode_1_real
7.29 7.22 0.63 1 0.63 8.20 tcd_decode_tile
5.79 7.72 0.50 1 0.50 0.50 mct_decode_real
4.86 8.14 0.42 1 0.42 0.42 imagetobmp
% cumulative self self total
time seconds seconds calls s/call s/call name
53.17 13.69 13.69 9 1.52 2.20 dwt_decode_tile
23.61 19.77 6.08 89280 0.00 0.00 dwt_decode_1_real
6.83 21.53 1.76 3 0.59 8.15 tcd_decode_tile
5.67 22.99 1.46 3 0.49 0.49 mct_decode_real
4.85 24.24 1.25 3 0.42 0.42 imagetobmp


You would expect your version to run faster, but this shows 6.59s total
on a single run and 19.77s total after 3 runs).

About a second difference in DWT time to decode "bird.jp2".

It just depends on how the compiler sees the variables and knows how to
optimize them away.