diff options
| author | Hubert Mazur <hum@semihalf.com> | 2022-09-20 13:01:56 +0200 |
|---|---|---|
| committer | Martin Storsjö <martin@martin.st> | 2022-09-21 12:57:56 +0300 |
| commit | 0ee535b1db392a340a86ce0581dff94624c2da06 (patch) | |
| tree | 7654c8c760ebdef327a106f729982ef3898b7842 /libavcodec/aarch64 | |
| parent | acbb2777e28c462d38f774d317e6cd3ad2a0b215 (diff) | |
| download | ffmpeg-0ee535b1db392a340a86ce0581dff94624c2da06.tar.gz | |
lavc/aarch64: Add neon implementation for pix_median_abs16
Provide optimized implementation for pix_median_abs16 function.
Performance comparison tests are shown below.
- median_sad_0_c: 720.5
- median_sad_0_neon: 127.2
Benchmarks and tests run with checkasm tool on AWS Graviton 3.
Signed-off-by: Hubert Mazur <hum@semihalf.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
Diffstat (limited to 'libavcodec/aarch64')
| -rw-r--r-- | libavcodec/aarch64/me_cmp_init_aarch64.c | 4 | ||||
| -rw-r--r-- | libavcodec/aarch64/me_cmp_neon.S | 78 |
2 files changed, 82 insertions, 0 deletions
diff --git a/libavcodec/aarch64/me_cmp_init_aarch64.c b/libavcodec/aarch64/me_cmp_init_aarch64.c index ade3e9a4c1..fb51a833be 100644 --- a/libavcodec/aarch64/me_cmp_init_aarch64.c +++ b/libavcodec/aarch64/me_cmp_init_aarch64.c @@ -53,6 +53,8 @@ int nsse16_neon(int multiplier, const uint8_t *s, const uint8_t *s2, ptrdiff_t stride, int h); int nsse16_neon_wrapper(MpegEncContext *c, const uint8_t *s1, const uint8_t *s2, ptrdiff_t stride, int h); +int pix_median_abs16_neon(MpegEncContext *v, const uint8_t *pix1, const uint8_t *pix2, + ptrdiff_t stride, int h); av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx) { @@ -78,6 +80,8 @@ av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx) c->vsse[4] = vsse_intra16_neon; c->nsse[0] = nsse16_neon_wrapper; + + c->median_sad[0] = pix_median_abs16_neon; } } diff --git a/libavcodec/aarch64/me_cmp_neon.S b/libavcodec/aarch64/me_cmp_neon.S index f8998749a5..1e08d04869 100644 --- a/libavcodec/aarch64/me_cmp_neon.S +++ b/libavcodec/aarch64/me_cmp_neon.S @@ -969,3 +969,81 @@ function nsse16_neon, export=1 ret endfunc + +function pix_median_abs16_neon, export=1 + // x0 unused + // x1 uint8_t *pix1 + // x2 uint8_t *pix2 + // x3 ptrdiff_t stride + // w4 int h + + ld1 {v2.16b}, [x1], x3 + ld1 {v3.16b}, [x2], x3 + movi v31.8h, #0 + movi v16.8h, #0 + ext v0.16b, v2.16b, v2.16b, #1 + ext v1.16b, v3.16b, v3.16b, #1 + usubl v28.8h, v2.8b, v3.8b + usubl2 v27.8h, v2.16b, v3.16b + usubl v26.8h, v0.8b, v1.8b + usubl2 v25.8h, v0.16b, v1.16b + sub w4, w4, #1 // we need to make h-1 iterations + saba v31.8h, v26.8h, v28.8h + saba v16.8h, v25.8h, v27.8h + mov h18, v28.h[0] + cmp w4, #1 + sqabs h18, h18 + movi v0.8h, #0 + + b.lt 2f +1: + + ld1 {v6.16b}, [x1], x3 // pix1 vector for V(j-1) + ld1 {v7.16b}, [x2], x3 // pix2 vector for V(j-1) + subs w4, w4, #1 + ext v4.16b, v6.16b, v6.16b, #1 // pix1 vector for V(j) + ext v5.16b, v7.16b, v7.16b, #1 // pix2 vector for V(j) + + // protected registers: v30, v29, v28, v27, v26, v25, v24, v23 + // scratch registers: v22, v21, v20, v19, v17 + + // To find median of three values, calculate sum of them + // and subtract max and min value from it. + usubl v30.8h, v6.8b, v7.8b // V(j-1) + usubl2 v29.8h, v6.16b, v7.16b // V(j-1) + usubl v24.8h, v4.8b, v5.8b // V(j) + usubl2 v23.8h, v4.16b, v5.16b // V(j) + saba v0.8h, v30.8h, v28.8h + add v22.8h, v26.8h, v30.8h + smin v20.8h, v26.8h, v30.8h + add v21.8h, v25.8h, v29.8h + smax v19.8h, v26.8h, v30.8h + sub v22.8h, v22.8h, v28.8h + sub v21.8h, v21.8h, v27.8h + smin v17.8h, v19.8h, v22.8h + smin v22.8h, v25.8h, v29.8h + mov v28.16b, v30.16b + smax v20.8h, v20.8h, v17.8h // median values lower half + smax v19.8h, v25.8h, v29.8h + saba v31.8h, v24.8h, v20.8h + mov v27.16b, v29.16b + smin v19.8h, v19.8h, v21.8h + mov v26.16b, v24.16b + smax v17.8h, v22.8h, v19.8h // median values upper half + mov v25.16b, v23.16b + saba v16.8h, v23.8h, v17.8h + + b.ne 1b + +2: + mov h17, v0.h[0] + ins v16.h[7], wzr + add d18, d18, d17 + add v31.8h, v31.8h, v16.8h + uaddlv s17, v31.8h + add d18, d18, d17 + fmov w0, s18 + + ret + +endfunc |