[x265] [arm64] enable luma_hv for all sizes
Pop, Sebastian
spop at amazon.com
Sat Jul 10 22:04:58 UTC 2021
Hi,
I have seen on a linux-perf profile that interp_hv_pp_c was still called:
0.98% ffmpeg_g libx265.so.200 [.] (anonymous namespace)::interp_hv_pp_c<8, 32, 32>
0.87% ffmpeg_g libx265.so.200 [.] (anonymous namespace)::interp_hv_pp_c<8, 16, 16>
0.25% ffmpeg_g libx265.so.200 [.] (anonymous namespace)::interp_hv_pp_c<8, 64, 64>
That was due to the fact that some of the kernels for luma_hvpp were disabled.
The attached patch enables luma_hvpp for all kernel sizes.
With the patch the profile does not contain calls to interp_hv_pp_c.
before the patch:
luma_hv [ 8x8] 16.08x 318.26 5117.80
luma_hv [ 8x4] 14.94x 220.35 3292.54
luma_hv [ 8x16] 16.89x 519.45 8773.31
luma_hv [12x16] 14.70x 962.32 14149.10
luma_hv [ 8x32] 19.10x 917.16 17519.01
with the patch:
luma_hv [ 4x4] 8.29x 203.25 1684.26
luma_hv [ 8x8] 15.92x 319.45 5086.50
luma_hv [16x16] 17.71x 984.39 17429.32
luma_hv [32x32] 18.54x 3411.28 63255.92
luma_hv [64x64] 19.29x 12628.54 243586.52
luma_hv [ 8x4] 14.89x 219.60 3268.85
luma_hv [ 4x8] 8.75x 297.49 2602.33
luma_hv [ 16x8] 16.86x 598.21 10085.83
luma_hv [ 8x16] 16.68x 520.16 8678.79
luma_hv [32x16] 18.16x 1906.65 34631.16
luma_hv [16x32] 17.31x 1842.00 31880.58
luma_hv [64x32] 19.03x 6696.71 127412.67
luma_hv [32x64] 19.02x 6406.64 121846.34
luma_hv [16x12] 17.40x 791.32 13772.11
luma_hv [12x16] 13.54x 964.49 13058.28
luma_hv [ 16x4] 15.78x 410.06 6469.77
luma_hv [ 4x16] 9.22x 482.95 4450.62
luma_hv [32x24] 18.39x 2661.41 48955.65
luma_hv [24x32] 18.51x 2585.15 47854.70
luma_hv [ 32x8] 17.49x 1150.18 20111.31
luma_hv [ 8x32] 17.36x 921.36 15992.87
luma_hv [64x48] 19.21x 9635.09 185077.92
luma_hv [48x64] 19.13x 9513.58 181971.16
luma_hv [64x16] 18.52x 3729.03 69050.16
luma_hv [16x64] 18.49x 3308.71 61182.30
Ok to commit?
Thanks,
Sebastian
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20210710/959be217/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-arm64-enable-luma_hv-for-all-sizes.patch
Type: application/octet-stream
Size: 3682 bytes
Desc: 0001-arm64-enable-luma_hv-for-all-sizes.patch
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20210710/959be217/attachment-0001.obj>
More information about the x265-devel
mailing list