Message ID | 20220725111208.43542-1-hum@semihalf.com |
---|---|
State | New |
Headers | show |
Series | [FFmpeg-devel] lavc/aarch64: Add neon implementation for sse16 | 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 |
On Mon, 25 Jul 2022, Hubert Mazur wrote: > Provide neon implementation for sse16 function. > > Performance comparison tests are shown below. > - sse_0_c: 273.0 > - sse_0_neon: 48.2 > > Benchmarks and tests run with checkasm tool on AWS Graviton 3. > > Signed-off-by: Hubert Mazur <hum@semihalf.com> > --- > libavcodec/aarch64/me_cmp_init_aarch64.c | 4 ++ > libavcodec/aarch64/me_cmp_neon.S | 82 ++++++++++++++++++++++++ > 2 files changed, 86 insertions(+) > > diff --git a/libavcodec/aarch64/me_cmp_init_aarch64.c b/libavcodec/aarch64/me_cmp_init_aarch64.c > index 136b008eb7..3ff5767bd0 100644 > --- a/libavcodec/aarch64/me_cmp_init_aarch64.c > +++ b/libavcodec/aarch64/me_cmp_init_aarch64.c > @@ -30,6 +30,9 @@ int ff_pix_abs16_xy2_neon(MpegEncContext *s, uint8_t *blk1, uint8_t *blk2, > int ff_pix_abs16_x2_neon(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2, > ptrdiff_t stride, int h); > > +int sse16_neon(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2, > + ptrdiff_t stride, int h); The signature of these functions has been changed now (right after these patches were submitted); the pix1/pix2 parameters are now const. Also, nitpick; please align the following line ("ptrdiff_t stride, ...") correctly with the parenthese on the line above. > + > av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx) > { > int cpu_flags = av_get_cpu_flags(); > @@ -40,5 +43,6 @@ av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx) > c->pix_abs[0][3] = ff_pix_abs16_xy2_neon; > > c->sad[0] = ff_pix_abs16_neon; > + c->sse[0] = sse16_neon; > } > } > diff --git a/libavcodec/aarch64/me_cmp_neon.S b/libavcodec/aarch64/me_cmp_neon.S > index cda7ce0408..98c912b608 100644 > --- a/libavcodec/aarch64/me_cmp_neon.S > +++ b/libavcodec/aarch64/me_cmp_neon.S > @@ -270,3 +270,85 @@ function ff_pix_abs16_x2_neon, export=1 > > ret > endfunc > + > +function sse16_neon, export=1 > + // x0 - unused > + // x1 - pix1 > + // x2 - pix2 > + // x3 - stride > + // w4 - h > + > + cmp w4, #4 > + movi d18, #0 > + b.lt 2f > + > +// Make 4 iterations at once > +1: > + > + // res = abs(pix1[0] - pix2[0]) > + // res * res > + > + ld1 {v0.16b}, [x1], x3 // Load pix1 vector for first iteration > + ld1 {v1.16b}, [x2], x3 // Load pix2 vector for first iteration > + uabd v30.16b, v0.16b, v1.16b // Absolute difference, first iteration Try to improve the interleaving of this function; I did a quick test on Cortex A53, A72 and A73, and got these numbers: Before: sse_0_neon: 147.7 64.5 64.7 After: sse_0_neon: 133.7 60.7 59.2 Overall, try to avoid having consecutive instructions operating on the same iteration (except for when doing the same operation on different halves of the same iteration), i.e. not "absolute difference third iteration; multiply lower half third iteration, multiply upper half third iteration, pairwise add third iteration", but bundle it up so you have e.g. "absolute difference third iteration; pairwise add first iteration; multiply {upper,lower} half third iteration; pairwise add second iteration; pairwise add third iteration", or something like that. Then secondly, in general, don't serialize the summation down to a single element in each iteration! You can keep the accumulated sum as a vX.4s vector (or maybe even better, two .4s vectors!) throughout the whole algorithm, and then only add them up horizontally (with an uaddv) at the end. For adding vectors, I would instinctively prefer doing "uaddl v0.4s, v2.4h, v3.4h; uaddl2 v1.4s, v2.8h, v3.8h" instead of "uaddlp v0.4s, v1.4h; uadalp v0.4s, v1.8h" etc. I didn't try out this modification, but please do, I'm pretty sure it will be a fair bit faster, and if not, at least more idiomatic SIMD. I didn't check the other patches yet, but if the other sse* functions are implemented similarly, I would expect the same feedback to apply to them too. Let's iterate on the sse16 patch first now at least and get that one great, and then update sse4/sse8 similarly once we have that one settled. I'll try to have a look at the other patches in the set later today/tomorrow. // Martin
On Mon, 25 Jul 2022, Hubert Mazur wrote: > Provide neon implementation for sse16 function. > > Performance comparison tests are shown below. > - sse_0_c: 273.0 > - sse_0_neon: 48.2 > > Benchmarks and tests run with checkasm tool on AWS Graviton 3. > > Signed-off-by: Hubert Mazur <hum@semihalf.com> > --- > libavcodec/aarch64/me_cmp_init_aarch64.c | 4 ++ > libavcodec/aarch64/me_cmp_neon.S | 82 ++++++++++++++++++++++++ > 2 files changed, 86 insertions(+) > +// iterate by one > +2: > + > + ld1 {v0.16b}, [x1], x3 // Load pix1 > + ld1 {v1.16b}, [x2], x3 // Load pix2 > + > + uabd v30.16b, v0.16b, v1.16b > + umull v29.8h, v0.8b, v1.8b > + umull2 v28.8h, v0.16b, v1.16b This should probably be using v30 instead of v0/v1 in the umull here. The whole codepath for non-modulo-4 heights is untested in practice. You can apply the patches from https://patchwork.ffmpeg.org/project/ffmpeg/list/?series=7028 to make checkasm test it, so please make sure that the uncommon codepaths in the patches do work too. // Martin
diff --git a/libavcodec/aarch64/me_cmp_init_aarch64.c b/libavcodec/aarch64/me_cmp_init_aarch64.c index 136b008eb7..3ff5767bd0 100644 --- a/libavcodec/aarch64/me_cmp_init_aarch64.c +++ b/libavcodec/aarch64/me_cmp_init_aarch64.c @@ -30,6 +30,9 @@ int ff_pix_abs16_xy2_neon(MpegEncContext *s, uint8_t *blk1, uint8_t *blk2, int ff_pix_abs16_x2_neon(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2, ptrdiff_t stride, int h); +int sse16_neon(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2, + ptrdiff_t stride, int h); + av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx) { int cpu_flags = av_get_cpu_flags(); @@ -40,5 +43,6 @@ av_cold void ff_me_cmp_init_aarch64(MECmpContext *c, AVCodecContext *avctx) c->pix_abs[0][3] = ff_pix_abs16_xy2_neon; c->sad[0] = ff_pix_abs16_neon; + c->sse[0] = sse16_neon; } } diff --git a/libavcodec/aarch64/me_cmp_neon.S b/libavcodec/aarch64/me_cmp_neon.S index cda7ce0408..98c912b608 100644 --- a/libavcodec/aarch64/me_cmp_neon.S +++ b/libavcodec/aarch64/me_cmp_neon.S @@ -270,3 +270,85 @@ function ff_pix_abs16_x2_neon, export=1 ret endfunc + +function sse16_neon, export=1 + // x0 - unused + // x1 - pix1 + // x2 - pix2 + // x3 - stride + // w4 - h + + cmp w4, #4 + movi d18, #0 + b.lt 2f + +// Make 4 iterations at once +1: + + // res = abs(pix1[0] - pix2[0]) + // res * res + + ld1 {v0.16b}, [x1], x3 // Load pix1 vector for first iteration + ld1 {v1.16b}, [x2], x3 // Load pix2 vector for first iteration + uabd v30.16b, v0.16b, v1.16b // Absolute difference, first iteration + ld1 {v2.16b}, [x1], x3 // Load pix1 vector for second iteration + umull v29.8h, v30.8b, v30.8b // Multiply lower half of vectors, first iteration + ld1 {v3.16b}, [x2], x3 // Load pix2 vector for second iteration + umull2 v28.8h, v30.16b, v30.16b // Multiply upper half of vectors, first iteration + uabd v27.16b, v2.16b, v3.16b // Absolute difference, second iteration + uaddlp v17.4s, v29.8h // Pairwise add, first iteration + umull v26.8h, v27.8b, v27.8b // Mulitply lower half, second iteration + umull2 v25.8h, v27.16b, v27.16b // Multiply upper half, second iteration + ld1 {v4.16b}, [x1], x3 // Load pix1 for third iteration + uadalp v17.4s, v26.8h // Pairwise add and accumulate, second iteration + ld1 {v5.16b}, [x2], x3 // Load pix2 for third iteration + uadalp v17.4s, v25.8h // Pairwise add andd accumulate, second iteration + uabd v24.16b, v4.16b, v5.16b // Absolute difference, third iteration + ld1 {v6.16b}, [x1], x3 // Load pix1 for fourth iteration + umull v23.8h, v24.8b, v24.8b // Multiply lower half, third iteration + umull2 v22.8h, v24.16b, v24.16b // Multiply upper half, third iteration + uadalp v17.4s, v23.8h // Pairwise add and accumulate, third iteration + uadalp v17.4s, v22.8h // Pairwise add and accumulate, third iteration + ld1 {v7.16b}, [x2], x3 // Load pix2 for fouth iteration + uadalp v17.4s, v28.8h // Pairwise add and accumulate, first iteration + uabd v21.16b, v6.16b, v7.16b // Absolute difference, fourth iteration + umull v20.8h, v21.8b, v21.8b // Multiply lower half, fourth iteration + uadalp v17.4s, v20.8h // Pairwise add and accumulate, fourth iteration + umull2 v19.8h, v21.16b, v21.16b // Multiply upper half, fourth iteration + uadalp v17.4s, v19.8h // Pairwise add and accumulate, fourth iteration + + sub w4, w4, #4 // h -= 4 + uaddlv d16, v17.4s // add up accumulator vector + cmp w4, #4 + add d18, d18, d16 + + b.ge 1b + + cbnz w4, 2f + fmov w0, s18 + + ret + +// iterate by one +2: + + ld1 {v0.16b}, [x1], x3 // Load pix1 + ld1 {v1.16b}, [x2], x3 // Load pix2 + + uabd v30.16b, v0.16b, v1.16b + umull v29.8h, v0.8b, v1.8b + umull2 v28.8h, v0.16b, v1.16b + uaddlp v17.4s, v29.8h + uadalp v17.4s, v28.8h + + + subs w4, w4, #1 + uaddlv d16, v17.4s + add d18, d18, d16 + + b.ne 2b + fmov w0, s18 + + ret + +endfunc
Provide neon implementation for sse16 function. Performance comparison tests are shown below. - sse_0_c: 273.0 - sse_0_neon: 48.2 Benchmarks and tests run with checkasm tool on AWS Graviton 3. Signed-off-by: Hubert Mazur <hum@semihalf.com> --- libavcodec/aarch64/me_cmp_init_aarch64.c | 4 ++ libavcodec/aarch64/me_cmp_neon.S | 82 ++++++++++++++++++++++++ 2 files changed, 86 insertions(+)