On Tue, 16 Aug 2022, Hubert Mazur wrote:
Provide optimized implementation of pix_abs8 function for arm64.
Performance comparison tests are shown below.
- pix_abs_1_0_c: 101.2
- pix_abs_1_0_neon: 22.5
- sad_1_c: 101.2
- sad_1_neon: 22.5
Benchmarks and tests are run with checkasm tool on AWS Graviton 3.
---
libavcodec/aarch64/me_cmp_init_aarch64.c | 4 ++
libavcodec/aarch64/me_cmp_neon.S | 49 ++++++++++++++++++++++++
2 files changed, 53 insertions(+)
diff --git a/libavcodec/aarch64/me_cmp_init_aarch64.c
b/libavcodec/aarch64/me_cmp_init_aarch64.c
index 2f51f0497e..e7dbd4cbc5 100644
--- a/libavcodec/aarch64/me_cmp_init_aarch64.c
+++ b/libavcodec/aarch64/me_cmp_init_aarch64.c
@@ -31,6 +31,8 @@ int ff_pix_abs16_x2_neon(MpegEncContext *v, const uint8_t
*pix1, const uint8_t *
ptrdiff_t stride, int h);
int ff_pix_abs16_y2_neon(MpegEncContext *v, const uint8_t *pix1, const uint8_t
*pix2,
ptrdiff_t stride, int h);
+int ff_pix_abs8_neon(MpegEncContext *s, const uint8_t *blk1, const uint8_t
*blk2,
+ ptrdiff_t stride, int h);
Alignment
diff --git a/libavcodec/aarch64/me_cmp_neon.S b/libavcodec/aarch64/me_cmp_neon.S
index 3f4266d4d5..8c396cad21 100644
--- a/libavcodec/aarch64/me_cmp_neon.S
+++ b/libavcodec/aarch64/me_cmp_neon.S
@@ -72,6 +72,55 @@ function ff_pix_abs16_neon, export=1
ret
endfunc
+function ff_pix_abs8_neon, export=1
+ // x0 unused
+ // x1 uint8_t *pix1
+ // x2 uint8_t *pix2
+ // x3 ptrdiff_t stride
+ // x4 int h
w4, not x4
+
+ movi d18, #0
Unused d18
+ movi v30.8h, #0
+ cmp w4, #4
+ b.lt 2f
+
+// make 4 iterations at once
+1:
+ ld1 {v0.8b}, [x1], x3 // Load pix1 for first
iteration
+ ld1 {v1.8b}, [x2], x3 // Load pix2 for first
iteration
+ ld1 {v2.8b}, [x1], x3 // Load pix1 for
second iteration
+ uabal v30.8h, v0.8b, v1.8b // Absolute
difference, first iteration
+ ld1 {v3.8b}, [x2], x3 // Load pix2 for
second iteration
+ ld1 {v4.8b}, [x1], x3 // Load pix1 for third
iteration
+ uabal v30.8h, v2.8b, v3.8b // Absolute
difference, second iteration
+ ld1 {v5.8b}, [x2], x3 // Load pix2 for third
iteration
+ sub w4, w4, #4 // h -= 4
+ uabal v30.8h, v4.8b, v5.8b // Absolute
difference, third iteration
+ ld1 {v6.8b}, [x1], x3 // Load pix1 for
foruth iteration
+ ld1 {v7.8b}, [x2], x3 // Load pix2 for
fourth iteration
+ cmp w4, #4
+ uabal v30.8h, v6.8b, v7.8b // Absolute
difference, foruth iteration
The interleaving here looks mostly quite good, but the last uabal comes
almost directly after the two loads; I moved the second-last uabal from
before the two ld1s to between ld1 and cmp, and got a rather notable
speedup.
Before: Cortex A53 A72 A73
pix_abs_1_0_neon: 65.7 33.7 21.5
After:
pix_abs_1_0_neon: 57.7 33.5 21.5
So this is a 13% speedup on Cortex A53, just by moving one single
instruction. This is why paying attention to scheduling matters, sometimes
a lot.
+ uaddlv s20, v30.8h // Add up vector
+ add d18, d18, d20
+ fmov w0, s18
And finally, by removing the unnecessary add of d18 here, I got this
further reduced to the following runtimes:
Cortex A53 A72 A73
pix_abs_1_0_neon: 54.7 30.7 20.2
// Martin
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
