| Message ID | tencent_C1565BD80F4F886AB785C7905C31D75C1906@qq.com |
|---|---|
| State | New |
| Headers | show |
| Series | [FFmpeg-devel] lavc/vp8dsp: R-V V put_bilin_h v unroll | 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 |
I directly copied the VP9 modifications over... Since len <= 16, it seems like it can be improved a bit more <uk7b@foxmail.com> 于2024年5月30日周四 23:27写道： > From: sunyuechi <sunyuechi@iscas.ac.cn> > > Since len < 64, the registers are sufficient, so it can be > directly unrolled (a4 is even). > > Another benefit of unrolling is that it reduces one load operation > vertically compared to horizontally. > > old new > C908 X60 C908 X60 > vp8_put_bilin4_h_c : 6.2 5.5 : 6.2 5.5 > vp8_put_bilin4_h_rvv_i32 : 2.2 2.0 : 1.5 1.5 > vp8_put_bilin4_v_c : 6.5 5.7 : 6.2 5.7 > vp8_put_bilin4_v_rvv_i32 : 2.2 2.0 : 1.2 1.5 > vp8_put_bilin8_h_c : 24.2 21.5 : 24.2 21.5 > vp8_put_bilin8_h_rvv_i32 : 5.2 4.7 : 3.5 3.5 > vp8_put_bilin8_v_c : 24.5 21.7 : 24.5 21.7 > vp8_put_bilin8_v_rvv_i32 : 5.2 4.7 : 3.5 3.2 > vp8_put_bilin16_h_c : 48.0 42.7 : 48.0 42.7 > vp8_put_bilin16_h_rvv_i32 : 5.7 5.0 : 5.2 4.5 > vp8_put_bilin16_v_c : 48.2 43.0 : 48.2 42.7 > vp8_put_bilin16_v_rvv_i32 : 5.7 5.2 : 4.5 4.2 > --- > libavcodec/riscv/vp8dsp_rvv.S | 34 +++++++++++++++++++++++++++++----- > 1 file changed, 29 insertions(+), 5 deletions(-) > > diff --git a/libavcodec/riscv/vp8dsp_rvv.S b/libavcodec/riscv/vp8dsp_rvv.S > index 3360a38cac..5bea6cba9c 100644 > --- a/libavcodec/riscv/vp8dsp_rvv.S > +++ b/libavcodec/riscv/vp8dsp_rvv.S > @@ -172,11 +172,35 @@ func ff_put_vp8_bilin4_\type\()_rvv, zve32x > li t4, 4 > sub t1, t1, \mn > 1: > - addi a4, a4, -1 > - bilin_load v0, \type, \mn > - vse8.v v0, (a0) > - add a2, a2, a3 > - add a0, a0, a1 > + add t0, a2, a3 > + add t2, a0, a1 > + addi a4, a4, -2 > +.ifc \type,v > + add t3, t0, a3 > +.else > + addi t5, a2, 1 > + addi t3, t0, 1 > + vle8.v v2, (t5) > +.endif > + vle8.v v0, (a2) > + vle8.v v4, (t0) > + vle8.v v6, (t3) > + vwmulu.vx v28, v0, t1 > + vwmulu.vx v26, v4, t1 > +.ifc \type,v > + vwmaccu.vx v28, \mn, v4 > +.else > + vwmaccu.vx v28, \mn, v2 > +.endif > + vwmaccu.vx v26, \mn, v6 > + vwaddu.wx v24, v28, t4 > + vwaddu.wx v22, v26, t4 > + vnsra.wi v30, v24, 3 > + vnsra.wi v0, v22, 3 > + vse8.v v30, (a0) > + vse8.v v0, (t2) > + add a2, t0, a3 > + add a0, t2, a1 > bnez a4, 1b > > ret > -- > 2.45.1 > > _______________________________________________ > 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". >
Well.. because scalar registers are limited, the direct unrolling will be like this for now. We can handle different lengths separately in the future flow gg <hlefthleft@gmail.com> 于2024年5月30日周四 23:36写道： > I directly copied the VP9 modifications over... Since len <= 16, it seems > like it can be improved a bit more > > <uk7b@foxmail.com> 于2024年5月30日周四 23:27写道： > >> From: sunyuechi <sunyuechi@iscas.ac.cn> >> >> Since len < 64, the registers are sufficient, so it can be >> directly unrolled (a4 is even). >> >> Another benefit of unrolling is that it reduces one load operation >> vertically compared to horizontally. >> >> old new >> C908 X60 C908 X60 >> vp8_put_bilin4_h_c : 6.2 5.5 : 6.2 5.5 >> vp8_put_bilin4_h_rvv_i32 : 2.2 2.0 : 1.5 1.5 >> vp8_put_bilin4_v_c : 6.5 5.7 : 6.2 5.7 >> vp8_put_bilin4_v_rvv_i32 : 2.2 2.0 : 1.2 1.5 >> vp8_put_bilin8_h_c : 24.2 21.5 : 24.2 21.5 >> vp8_put_bilin8_h_rvv_i32 : 5.2 4.7 : 3.5 3.5 >> vp8_put_bilin8_v_c : 24.5 21.7 : 24.5 21.7 >> vp8_put_bilin8_v_rvv_i32 : 5.2 4.7 : 3.5 3.2 >> vp8_put_bilin16_h_c : 48.0 42.7 : 48.0 42.7 >> vp8_put_bilin16_h_rvv_i32 : 5.7 5.0 : 5.2 4.5 >> vp8_put_bilin16_v_c : 48.2 43.0 : 48.2 42.7 >> vp8_put_bilin16_v_rvv_i32 : 5.7 5.2 : 4.5 4.2 >> --- >> libavcodec/riscv/vp8dsp_rvv.S | 34 +++++++++++++++++++++++++++++----- >> 1 file changed, 29 insertions(+), 5 deletions(-) >> >> diff --git a/libavcodec/riscv/vp8dsp_rvv.S b/libavcodec/riscv/vp8dsp_rvv.S >> index 3360a38cac..5bea6cba9c 100644 >> --- a/libavcodec/riscv/vp8dsp_rvv.S >> +++ b/libavcodec/riscv/vp8dsp_rvv.S >> @@ -172,11 +172,35 @@ func ff_put_vp8_bilin4_\type\()_rvv, zve32x >> li t4, 4 >> sub t1, t1, \mn >> 1: >> - addi a4, a4, -1 >> - bilin_load v0, \type, \mn >> - vse8.v v0, (a0) >> - add a2, a2, a3 >> - add a0, a0, a1 >> + add t0, a2, a3 >> + add t2, a0, a1 >> + addi a4, a4, -2 >> +.ifc \type,v >> + add t3, t0, a3 >> +.else >> + addi t5, a2, 1 >> + addi t3, t0, 1 >> + vle8.v v2, (t5) >> +.endif >> + vle8.v v0, (a2) >> + vle8.v v4, (t0) >> + vle8.v v6, (t3) >> + vwmulu.vx v28, v0, t1 >> + vwmulu.vx v26, v4, t1 >> +.ifc \type,v >> + vwmaccu.vx v28, \mn, v4 >> +.else >> + vwmaccu.vx v28, \mn, v2 >> +.endif >> + vwmaccu.vx v26, \mn, v6 >> + vwaddu.wx v24, v28, t4 >> + vwaddu.wx v22, v26, t4 >> + vnsra.wi v30, v24, 3 >> + vnsra.wi v0, v22, 3 >> + vse8.v v30, (a0) >> + vse8.v v0, (t2) >> + add a2, t0, a3 >> + add a0, t2, a1 >> bnez a4, 1b >> >> ret >> -- >> 2.45.1 >> >> _______________________________________________ >> 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". >> >
ping <uk7b@foxmail.com> 于2024年5月30日周四 23:27写道： > From: sunyuechi <sunyuechi@iscas.ac.cn> > > Since len < 64, the registers are sufficient, so it can be > directly unrolled (a4 is even). > > Another benefit of unrolling is that it reduces one load operation > vertically compared to horizontally. > > old new > C908 X60 C908 X60 > vp8_put_bilin4_h_c : 6.2 5.5 : 6.2 5.5 > vp8_put_bilin4_h_rvv_i32 : 2.2 2.0 : 1.5 1.5 > vp8_put_bilin4_v_c : 6.5 5.7 : 6.2 5.7 > vp8_put_bilin4_v_rvv_i32 : 2.2 2.0 : 1.2 1.5 > vp8_put_bilin8_h_c : 24.2 21.5 : 24.2 21.5 > vp8_put_bilin8_h_rvv_i32 : 5.2 4.7 : 3.5 3.5 > vp8_put_bilin8_v_c : 24.5 21.7 : 24.5 21.7 > vp8_put_bilin8_v_rvv_i32 : 5.2 4.7 : 3.5 3.2 > vp8_put_bilin16_h_c : 48.0 42.7 : 48.0 42.7 > vp8_put_bilin16_h_rvv_i32 : 5.7 5.0 : 5.2 4.5 > vp8_put_bilin16_v_c : 48.2 43.0 : 48.2 42.7 > vp8_put_bilin16_v_rvv_i32 : 5.7 5.2 : 4.5 4.2 > --- > libavcodec/riscv/vp8dsp_rvv.S | 34 +++++++++++++++++++++++++++++----- > 1 file changed, 29 insertions(+), 5 deletions(-) > > diff --git a/libavcodec/riscv/vp8dsp_rvv.S b/libavcodec/riscv/vp8dsp_rvv.S > index 3360a38cac..5bea6cba9c 100644 > --- a/libavcodec/riscv/vp8dsp_rvv.S > +++ b/libavcodec/riscv/vp8dsp_rvv.S > @@ -172,11 +172,35 @@ func ff_put_vp8_bilin4_\type\()_rvv, zve32x > li t4, 4 > sub t1, t1, \mn > 1: > - addi a4, a4, -1 > - bilin_load v0, \type, \mn > - vse8.v v0, (a0) > - add a2, a2, a3 > - add a0, a0, a1 > + add t0, a2, a3 > + add t2, a0, a1 > + addi a4, a4, -2 > +.ifc \type,v > + add t3, t0, a3 > +.else > + addi t5, a2, 1 > + addi t3, t0, 1 > + vle8.v v2, (t5) > +.endif > + vle8.v v0, (a2) > + vle8.v v4, (t0) > + vle8.v v6, (t3) > + vwmulu.vx v28, v0, t1 > + vwmulu.vx v26, v4, t1 > +.ifc \type,v > + vwmaccu.vx v28, \mn, v4 > +.else > + vwmaccu.vx v28, \mn, v2 > +.endif > + vwmaccu.vx v26, \mn, v6 > + vwaddu.wx v24, v28, t4 > + vwaddu.wx v22, v26, t4 > + vnsra.wi v30, v24, 3 > + vnsra.wi v0, v22, 3 > + vse8.v v30, (a0) > + vse8.v v0, (t2) > + add a2, t0, a3 > + add a0, t2, a1 > bnez a4, 1b > > ret > -- > 2.45.1 > > _______________________________________________ > 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". >
Le torstaina 30. toukokuuta 2024, 18.26.53 EEST uk7b@foxmail.com a écrit : > From: sunyuechi <sunyuechi@iscas.ac.cn> > > Since len < 64, the registers are sufficient, so it can be > directly unrolled (a4 is even). > > Another benefit of unrolling is that it reduces one load operation > vertically compared to horizontally. > > old new > C908 X60 C908 X60 > vp8_put_bilin4_h_c : 6.2 5.5 : 6.2 5.5 > vp8_put_bilin4_h_rvv_i32 : 2.2 2.0 : 1.5 1.5 > vp8_put_bilin4_v_c : 6.5 5.7 : 6.2 5.7 > vp8_put_bilin4_v_rvv_i32 : 2.2 2.0 : 1.2 1.5 > vp8_put_bilin8_h_c : 24.2 21.5 : 24.2 21.5 > vp8_put_bilin8_h_rvv_i32 : 5.2 4.7 : 3.5 3.5 > vp8_put_bilin8_v_c : 24.5 21.7 : 24.5 21.7 > vp8_put_bilin8_v_rvv_i32 : 5.2 4.7 : 3.5 3.2 > vp8_put_bilin16_h_c : 48.0 42.7 : 48.0 42.7 > vp8_put_bilin16_h_rvv_i32 : 5.7 5.0 : 5.2 4.5 > vp8_put_bilin16_v_c : 48.2 43.0 : 48.2 42.7 > vp8_put_bilin16_v_rvv_i32 : 5.7 5.2 : 4.5 4.2 > --- > libavcodec/riscv/vp8dsp_rvv.S | 34 +++++++++++++++++++++++++++++----- > 1 file changed, 29 insertions(+), 5 deletions(-) > > diff --git a/libavcodec/riscv/vp8dsp_rvv.S b/libavcodec/riscv/vp8dsp_rvv.S > index 3360a38cac..5bea6cba9c 100644 > --- a/libavcodec/riscv/vp8dsp_rvv.S > +++ b/libavcodec/riscv/vp8dsp_rvv.S > @@ -172,11 +172,35 @@ func ff_put_vp8_bilin4_\type\()_rvv, zve32x > li t4, 4 > sub t1, t1, \mn > 1: > - addi a4, a4, -1 > - bilin_load v0, \type, \mn Does this not render the type parameter of bilin_load useless (always h)? (Not a blocker for this patch.) > - vse8.v v0, (a0) > - add a2, a2, a3 > - add a0, a0, a1 > + add t0, a2, a3 > + add t2, a0, a1 > + addi a4, a4, -2 > +.ifc \type,v > + add t3, t0, a3 > +.else > + addi t5, a2, 1 > + addi t3, t0, 1 > + vle8.v v2, (t5) Not sure if I already asked this but is this really faster than slide1? Normally we want to minimise the work of the memory bus. > +.endif > + vle8.v v0, (a2) > + vle8.v v4, (t0) > + vle8.v v6, (t3) > + vwmulu.vx v28, v0, t1 > + vwmulu.vx v26, v4, t1 > +.ifc \type,v > + vwmaccu.vx v28, \mn, v4 > +.else > + vwmaccu.vx v28, \mn, v2 > +.endif > + vwmaccu.vx v26, \mn, v6 > + vwaddu.wx v24, v28, t4 > + vwaddu.wx v22, v26, t4 > + vnsra.wi v30, v24, 3 > + vnsra.wi v0, v22, 3 > + vse8.v v30, (a0) > + vse8.v v0, (t2) > + add a2, t0, a3 > + add a0, t2, a1 > bnez a4, 1b > > ret
> Does this not render the type parameter of bilin_load useless (always h)? > (Not a blocker for this patch.) Yes, this was needed in the initial version, but it is no longer required. I just sent a patch. > Not sure if I already asked this but is this really faster than slide1? > Normally we want to minimise the work of the memory bus. Originally it was slide, but based on your review, it was changed to load, which should be better. review: "Can't we skip the slide and just load the vector at a2+1? Also then, we can keep VL=len and halve the multipler." Rémi Denis-Courmont <remi@remlab.net> 于2024年6月12日周三 22:41写道： > Le torstaina 30. toukokuuta 2024, 18.26.53 EEST uk7b@foxmail.com a écrit : > > From: sunyuechi <sunyuechi@iscas.ac.cn> > > > > Since len < 64, the registers are sufficient, so it can be > > directly unrolled (a4 is even). > > > > Another benefit of unrolling is that it reduces one load operation > > vertically compared to horizontally. > > > > old new > > C908 X60 C908 X60 > > vp8_put_bilin4_h_c : 6.2 5.5 : 6.2 5.5 > > vp8_put_bilin4_h_rvv_i32 : 2.2 2.0 : 1.5 1.5 > > vp8_put_bilin4_v_c : 6.5 5.7 : 6.2 5.7 > > vp8_put_bilin4_v_rvv_i32 : 2.2 2.0 : 1.2 1.5 > > vp8_put_bilin8_h_c : 24.2 21.5 : 24.2 21.5 > > vp8_put_bilin8_h_rvv_i32 : 5.2 4.7 : 3.5 3.5 > > vp8_put_bilin8_v_c : 24.5 21.7 : 24.5 21.7 > > vp8_put_bilin8_v_rvv_i32 : 5.2 4.7 : 3.5 3.2 > > vp8_put_bilin16_h_c : 48.0 42.7 : 48.0 42.7 > > vp8_put_bilin16_h_rvv_i32 : 5.7 5.0 : 5.2 4.5 > > vp8_put_bilin16_v_c : 48.2 43.0 : 48.2 42.7 > > vp8_put_bilin16_v_rvv_i32 : 5.7 5.2 : 4.5 4.2 > > --- > > libavcodec/riscv/vp8dsp_rvv.S | 34 +++++++++++++++++++++++++++++----- > > 1 file changed, 29 insertions(+), 5 deletions(-) > > > > diff --git a/libavcodec/riscv/vp8dsp_rvv.S > b/libavcodec/riscv/vp8dsp_rvv.S > > index 3360a38cac..5bea6cba9c 100644 > > --- a/libavcodec/riscv/vp8dsp_rvv.S > > +++ b/libavcodec/riscv/vp8dsp_rvv.S > > @@ -172,11 +172,35 @@ func ff_put_vp8_bilin4_\type\()_rvv, zve32x > > li t4, 4 > > sub t1, t1, \mn > > 1: > > - addi a4, a4, -1 > > - bilin_load v0, \type, \mn > > Does this not render the type parameter of bilin_load useless (always h)? > (Not a blocker for this patch.) > > > - vse8.v v0, (a0) > > - add a2, a2, a3 > > - add a0, a0, a1 > > + add t0, a2, a3 > > + add t2, a0, a1 > > + addi a4, a4, -2 > > +.ifc \type,v > > + add t3, t0, a3 > > +.else > > + addi t5, a2, 1 > > + addi t3, t0, 1 > > + vle8.v v2, (t5) > > Not sure if I already asked this but is this really faster than slide1? > Normally we want to minimise the work of the memory bus. > > > +.endif > > + vle8.v v0, (a2) > > + vle8.v v4, (t0) > > + vle8.v v6, (t3) > > + vwmulu.vx v28, v0, t1 > > + vwmulu.vx v26, v4, t1 > > +.ifc \type,v > > + vwmaccu.vx v28, \mn, v4 > > +.else > > + vwmaccu.vx v28, \mn, v2 > > +.endif > > + vwmaccu.vx v26, \mn, v6 > > + vwaddu.wx v24, v28, t4 > > + vwaddu.wx v22, v26, t4 > > + vnsra.wi v30, v24, 3 > > + vnsra.wi v0, v22, 3 > > + vse8.v v30, (a0) > > + vse8.v v0, (t2) > > + add a2, t0, a3 > > + add a0, t2, a1 > > bnez a4, 1b > > > > ret > > -- > 雷米‧德尼-库尔蒙 > http://www.remlab.net/ > > > > _______________________________________________ > 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". >
diff --git a/libavcodec/riscv/vp8dsp_rvv.S b/libavcodec/riscv/vp8dsp_rvv.S index 3360a38cac..5bea6cba9c 100644 --- a/libavcodec/riscv/vp8dsp_rvv.S +++ b/libavcodec/riscv/vp8dsp_rvv.S @@ -172,11 +172,35 @@ func ff_put_vp8_bilin4_\type\()_rvv, zve32x li t4, 4 sub t1, t1, \mn 1: - addi a4, a4, -1 - bilin_load v0, \type, \mn - vse8.v v0, (a0) - add a2, a2, a3 - add a0, a0, a1 + add t0, a2, a3 + add t2, a0, a1 + addi a4, a4, -2 +.ifc \type,v + add t3, t0, a3 +.else + addi t5, a2, 1 + addi t3, t0, 1 + vle8.v v2, (t5) +.endif + vle8.v v0, (a2) + vle8.v v4, (t0) + vle8.v v6, (t3) + vwmulu.vx v28, v0, t1 + vwmulu.vx v26, v4, t1 +.ifc \type,v + vwmaccu.vx v28, \mn, v4 +.else + vwmaccu.vx v28, \mn, v2 +.endif + vwmaccu.vx v26, \mn, v6 + vwaddu.wx v24, v28, t4 + vwaddu.wx v22, v26, t4 + vnsra.wi v30, v24, 3 + vnsra.wi v0, v22, 3 + vse8.v v30, (a0) + vse8.v v0, (t2) + add a2, t0, a3 + add a0, t2, a1 bnez a4, 1b ret
From: sunyuechi <sunyuechi@iscas.ac.cn> Since len < 64, the registers are sufficient, so it can be directly unrolled (a4 is even). Another benefit of unrolling is that it reduces one load operation vertically compared to horizontally. old new C908 X60 C908 X60 vp8_put_bilin4_h_c : 6.2 5.5 : 6.2 5.5 vp8_put_bilin4_h_rvv_i32 : 2.2 2.0 : 1.5 1.5 vp8_put_bilin4_v_c : 6.5 5.7 : 6.2 5.7 vp8_put_bilin4_v_rvv_i32 : 2.2 2.0 : 1.2 1.5 vp8_put_bilin8_h_c : 24.2 21.5 : 24.2 21.5 vp8_put_bilin8_h_rvv_i32 : 5.2 4.7 : 3.5 3.5 vp8_put_bilin8_v_c : 24.5 21.7 : 24.5 21.7 vp8_put_bilin8_v_rvv_i32 : 5.2 4.7 : 3.5 3.2 vp8_put_bilin16_h_c : 48.0 42.7 : 48.0 42.7 vp8_put_bilin16_h_rvv_i32 : 5.7 5.0 : 5.2 4.5 vp8_put_bilin16_v_c : 48.2 43.0 : 48.2 42.7 vp8_put_bilin16_v_rvv_i32 : 5.7 5.2 : 4.5 4.2 --- libavcodec/riscv/vp8dsp_rvv.S | 34 +++++++++++++++++++++++++++++----- 1 file changed, 29 insertions(+), 5 deletions(-)