> 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 <r...@remlab.net> 于2024年6月12日周三 22:41写道： > Le torstaina 30. toukokuuta 2024, 18.26.53 EEST u...@foxmail.com a écrit : > > From: sunyuechi <sunyue...@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-requ...@ffmpeg.org with subject "unsubscribe". > _______________________________________________ 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".
