| Message ID | 20220420085753.2019156-1-martin@martin.st |
|---|---|
| State | Accepted |
| Commit | 70db14376c206f0fdcbff11c17875f95885c73d9 |
| Headers | show |
| Series | [FFmpeg-devel] swscale: aarch64: Optimize the final summation in the hscale routine | expand |
| Context | Check | Description |
|---|---|---|
| yinshiyou/make_loongarch64 | success | Make finished |
| yinshiyou/make_fate_loongarch64 | success | Make fate finished |
| andriy/make_x86 | success | Make finished |
| andriy/make_fate_x86 | success | Make fate finished |
| andriy/make_armv7_RPi4 | success | Make finished |
| andriy/make_fate_armv7_RPi4 | success | Make fate finished |
Thanks for making this improvement. I will rebase my patches on your change. I also measured the performance on AWS Graviton 2 and 3. I added the numbers to your table.
Before: Cortex A53 A72 A73 Graviton 2 Graviton 3
hscale_8_to_15_width8_neon: 8273.0 4602.5 4289.5 2429.7 1629.1
hscale_8_to_15_width16_neon: 12405.7 6803.0 6359.0 3549.0 2378.4
hscale_8_to_15_width32_neon: 21258.7 11491.7 11469.2 5797.2 3919.6
hscale_8_to_15_width40_neon: 25652.0 14173.7 12488.2 6893.5 4810.4
After:
hscale_8_to_15_width8_neon: 7633.0 3981.5 3350.2 1980.7 1261.1
hscale_8_to_15_width16_neon: 11666.7 5951.0 5512.0 3080.7 2131.4
hscale_8_to_15_width32_neon: 20900.7 10733.2 9481.7 5275.2 3862.1
hscale_8_to_15_width40_neon: 24826.0 13536.2 11502.0 6397.2 4731.9
--
Jonathan Swinney
On 4/20/22, 3:58 AM, "ffmpeg-devel on behalf of Martin Storsjö" <ffmpeg-devel-bounces@ffmpeg.org on behalf of martin@martin.st> wrote:
CAUTION: This email originated from outside of the organization. Do not click links or open attachments unless you can confirm the sender and know the content is safe.
Before: Cortex A53 A72 A73
hscale_8_to_15_width8_neon: 8273.0 4602.5 4289.5
hscale_8_to_15_width16_neon: 12405.7 6803.0 6359.0
hscale_8_to_15_width32_neon: 21258.7 11491.7 11469.2
hscale_8_to_15_width40_neon: 25652.0 14173.7 12488.2
After:
hscale_8_to_15_width8_neon: 7633.0 3981.5 3350.2
hscale_8_to_15_width16_neon: 11666.7 5951.0 5512.0
hscale_8_to_15_width32_neon: 20900.7 10733.2 9481.7
hscale_8_to_15_width40_neon: 24826.0 13536.2 11502.0
Thus, this gives overall a 8-28% speedup for the smaller filter
sizes, around 3-8% for the smaller filter sizes.
Inspired by a patch by Jonathan Swinney <jswinney@amazon.com>.
Signed-off-by: Martin Storsjö <martin@martin.st>
---
I'll go ahead and apply this patch within a few days if there's no
opposition, as it should be a fairly uncontroversial change.
---
libswscale/aarch64/hscale.S | 14 +++-----------
1 file changed, 3 insertions(+), 11 deletions(-)
diff --git a/libswscale/aarch64/hscale.S b/libswscale/aarch64/hscale.S
index af55ffe2b7..da34f1cb8d 100644
--- a/libswscale/aarch64/hscale.S
+++ b/libswscale/aarch64/hscale.S
@@ -61,17 +61,9 @@ function ff_hscale_8_to_15_neon, export=1
smlal v3.4S, v18.4H, v19.4H // v3 accumulates srcp[filterPos[3] + {0..3}] * filter[{0..3}]
smlal2 v3.4S, v18.8H, v19.8H // v3 accumulates srcp[filterPos[3] + {4..7}] * filter[{4..7}]
b.gt 2b // inner loop if filterSize not consumed completely
- addp v0.4S, v0.4S, v0.4S // part0 horizontal pair adding
- addp v1.4S, v1.4S, v1.4S // part1 horizontal pair adding
- addp v2.4S, v2.4S, v2.4S // part2 horizontal pair adding
- addp v3.4S, v3.4S, v3.4S // part3 horizontal pair adding
- addp v0.4S, v0.4S, v0.4S // part0 horizontal pair adding
- addp v1.4S, v1.4S, v1.4S // part1 horizontal pair adding
- addp v2.4S, v2.4S, v2.4S // part2 horizontal pair adding
- addp v3.4S, v3.4S, v3.4S // part3 horizontal pair adding
- zip1 v0.4S, v0.4S, v1.4S // part01 = zip values from part0 and part1
- zip1 v2.4S, v2.4S, v3.4S // part23 = zip values from part2 and part3
- mov v0.d[1], v2.d[0] // part0123 = zip values from part01 and part23
+ addp v0.4S, v0.4S, v1.4S // part01 horizontal pair adding
+ addp v2.4S, v2.4S, v3.4S // part23 horizontal pair adding
+ addp v0.4S, v0.4S, v2.4S // part0123 horizontal pair adding
subs w2, w2, #4 // dstW -= 4
sqshrn v0.4H, v0.4S, #7 // shift and clip the 2x16-bit final values
st1 {v0.4H}, [x1], #8 // write to destination part0123
--
2.25.1
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
On Thu, 21 Apr 2022, Swinney, Jonathan wrote: > Thanks for making this improvement. I will rebase my patches on your change. I also measured the performance on AWS Graviton 2 and 3. I added the numbers to your table. > > Before: Cortex A53 A72 A73 Graviton 2 Graviton 3 > hscale_8_to_15_width8_neon: 8273.0 4602.5 4289.5 2429.7 1629.1 > hscale_8_to_15_width16_neon: 12405.7 6803.0 6359.0 3549.0 2378.4 > hscale_8_to_15_width32_neon: 21258.7 11491.7 11469.2 5797.2 3919.6 > hscale_8_to_15_width40_neon: 25652.0 14173.7 12488.2 6893.5 4810.4 > > After: > hscale_8_to_15_width8_neon: 7633.0 3981.5 3350.2 1980.7 1261.1 > hscale_8_to_15_width16_neon: 11666.7 5951.0 5512.0 3080.7 2131.4 > hscale_8_to_15_width32_neon: 20900.7 10733.2 9481.7 5275.2 3862.1 > hscale_8_to_15_width40_neon: 24826.0 13536.2 11502.0 6397.2 4731.9 Thanks for the benchmarks! I pushed this patch now, with those numbers included. // Martin
diff --git a/libswscale/aarch64/hscale.S b/libswscale/aarch64/hscale.S index af55ffe2b7..da34f1cb8d 100644 --- a/libswscale/aarch64/hscale.S +++ b/libswscale/aarch64/hscale.S @@ -61,17 +61,9 @@ function ff_hscale_8_to_15_neon, export=1 smlal v3.4S, v18.4H, v19.4H // v3 accumulates srcp[filterPos[3] + {0..3}] * filter[{0..3}] smlal2 v3.4S, v18.8H, v19.8H // v3 accumulates srcp[filterPos[3] + {4..7}] * filter[{4..7}] b.gt 2b // inner loop if filterSize not consumed completely - addp v0.4S, v0.4S, v0.4S // part0 horizontal pair adding - addp v1.4S, v1.4S, v1.4S // part1 horizontal pair adding - addp v2.4S, v2.4S, v2.4S // part2 horizontal pair adding - addp v3.4S, v3.4S, v3.4S // part3 horizontal pair adding - addp v0.4S, v0.4S, v0.4S // part0 horizontal pair adding - addp v1.4S, v1.4S, v1.4S // part1 horizontal pair adding - addp v2.4S, v2.4S, v2.4S // part2 horizontal pair adding - addp v3.4S, v3.4S, v3.4S // part3 horizontal pair adding - zip1 v0.4S, v0.4S, v1.4S // part01 = zip values from part0 and part1 - zip1 v2.4S, v2.4S, v3.4S // part23 = zip values from part2 and part3 - mov v0.d[1], v2.d[0] // part0123 = zip values from part01 and part23 + addp v0.4S, v0.4S, v1.4S // part01 horizontal pair adding + addp v2.4S, v2.4S, v3.4S // part23 horizontal pair adding + addp v0.4S, v0.4S, v2.4S // part0123 horizontal pair adding subs w2, w2, #4 // dstW -= 4 sqshrn v0.4H, v0.4S, #7 // shift and clip the 2x16-bit final values st1 {v0.4H}, [x1], #8 // write to destination part0123
Before: Cortex A53 A72 A73 hscale_8_to_15_width8_neon: 8273.0 4602.5 4289.5 hscale_8_to_15_width16_neon: 12405.7 6803.0 6359.0 hscale_8_to_15_width32_neon: 21258.7 11491.7 11469.2 hscale_8_to_15_width40_neon: 25652.0 14173.7 12488.2 After: hscale_8_to_15_width8_neon: 7633.0 3981.5 3350.2 hscale_8_to_15_width16_neon: 11666.7 5951.0 5512.0 hscale_8_to_15_width32_neon: 20900.7 10733.2 9481.7 hscale_8_to_15_width40_neon: 24826.0 13536.2 11502.0 Thus, this gives overall a 8-28% speedup for the smaller filter sizes, around 3-8% for the smaller filter sizes. Inspired by a patch by Jonathan Swinney <jswinney@amazon.com>. Signed-off-by: Martin Storsjö <martin@martin.st> --- I'll go ahead and apply this patch within a few days if there's no opposition, as it should be a fairly uncontroversial change. --- libswscale/aarch64/hscale.S | 14 +++----------- 1 file changed, 3 insertions(+), 11 deletions(-)