[FFmpeg-devel] [PATCH 2/2] lavc/vvc_mc: R-V V dmvr
Rémi Denis-Courmont
remi at remlab.net
Sat Sep 28 16:49:37 EEST 2024
Le 28 septembre 2024 12:42:37 GMT+03:00, flow gg <hlefthleft at gmail.com> a écrit :
>> Is 4x unroll really faster than 2x here? We don't typically unroll 4x
>> manually.
>
>I first did 2x and then changed it to 4x. The test results are similar, and
>I'm not sure how to choose between them...
At similar speed, shorter code is better.
>> t5 seems to be 8-bit, so vwmulu.vx should work better here? Since you
>> leveraged it in the previous function, I'm a bit confused why not here,
>TBH.
>> Likewise vwmaccu.vx.
>
>DMVR doesn't have right shifts, but DMVR_h, _v, and _hv do.
>So DMVR only needs one vset, while the others, if using widen, require vset
>switching.
Sure but so what? vsetvli/vsetivli is pretty fast (unlike vsetvl), and in this case the code would be shorter. Or are you trying to factor the code for different VTYPEs?
>> Missing rounding opportunity, vssra.vi should work better here.
>> Same comments.
>
>Okay, Updated it.
>
>Rémi Denis-Courmont <remi at remlab.net> 于2024年9月28日周六 14:56写道:
>
>> Hi,
>>
>> Le perjantaina 27. syyskuuta 2024, 20.09.30 EEST uk7b at foxmail.com a écrit
>> :
>> > From: sunyuechi <sunyuechi at iscas.ac.cn>
>> >
>> > k230 banana_f3
>> > dmvr_8_12x20_c: 628.5 ( 1.00x) 624.1 ( 1.00x)
>> > dmvr_8_12x20_rvv_i32: 137.5 ( 4.57x) 92.9 ( 6.72x)
>> > dmvr_8_20x12_c: 609.7 ( 1.00x) 655.4 ( 1.00x)
>> > dmvr_8_20x12_rvv_i32: 146.7 ( 4.16x) 82.4 ( 7.95x)
>> > dmvr_8_20x20_c: 998.7 ( 1.00x) 1092.9 ( 1.00x)
>> > dmvr_8_20x20_rvv_i32: 221.0 ( 4.52x) 144.9 ( 7.54x)
>> > dmvr_h_8_12x20_c: 2008.0 ( 1.00x) 1999.2 ( 1.00x)
>> > dmvr_h_8_12x20_rvv_i32: 285.7 ( 7.03x) 207.4 ( 9.64x)
>> > dmvr_h_8_20x12_c: 1989.5 ( 1.00x) 2009.7 ( 1.00x)
>> > dmvr_h_8_20x12_rvv_i32: 322.7 ( 6.16x) 176.2 (11.41x)
>> > dmvr_h_8_20x20_c: 3304.2 ( 1.00x) 3342.9 ( 1.00x)
>> > dmvr_h_8_20x20_rvv_i32: 526.5 ( 6.28x) 290.6 (11.50x)
>> > dmvr_hv_8_12x20_c: 3609.7 ( 1.00x) 3603.4 ( 1.00x)
>> > dmvr_hv_8_12x20_rvv_i32: 554.2 ( 6.51x) 467.9 ( 7.70x)
>> > dmvr_hv_8_20x12_c: 3637.5 ( 1.00x) 3624.4 ( 1.00x)
>> > dmvr_hv_8_20x12_rvv_i32: 489.5 ( 7.43x) 342.6 (10.58x)
>> > dmvr_hv_8_20x20_c: 6794.7 ( 1.00x) 5936.9 ( 1.00x)
>> > dmvr_hv_8_20x20_rvv_i32: 785.7 ( 8.65x) 561.4 (10.58x)
>> > dmvr_v_8_12x20_c: 2156.0 ( 1.00x) 2155.2 ( 1.00x)
>> > dmvr_v_8_12x20_rvv_i32: 295.0 ( 7.31x) 207.4 (10.39x)
>> > dmvr_v_8_20x12_c: 2137.5 ( 1.00x) 2165.7 ( 1.00x)
>> > dmvr_v_8_20x12_rvv_i32: 322.7 ( 6.62x) 186.7 (11.60x)
>> > dmvr_v_8_20x20_c: 3554.2 ( 1.00x) 3593.2 ( 1.00x)
>> > dmvr_v_8_20x20_rvv_i32: 535.7 ( 6.63x) 290.6 (12.36x)
>> > ---
>> > libavcodec/riscv/vvc/vvc_mc_rvv.S | 141 +++++++++++++++++++++++++++++
>> > libavcodec/riscv/vvc/vvcdsp_init.c | 22 +++++
>> > 2 files changed, 163 insertions(+)
>> >
>> > diff --git a/libavcodec/riscv/vvc/vvc_mc_rvv.S
>> > b/libavcodec/riscv/vvc/vvc_mc_rvv.S index 18532616d9..a5e20cbc67 100644
>> > --- a/libavcodec/riscv/vvc/vvc_mc_rvv.S
>> > +++ b/libavcodec/riscv/vvc/vvc_mc_rvv.S
>> > @@ -285,3 +285,144 @@ endfunc
>> > func_w_avg 128
>> > func_w_avg 256
>> > #endif
>> > +
>> > +func dmvr zve32x, zbb, zba
>> > + lpad 0
>> > + li t0, 4
>> > +1:
>> > + add t1, a1, a2
>> > + addi t4, a0, 128*2
>> > + add t2, t1, a2
>> > + addi t5, a0, 128*2*2
>> > + add t3, t2, a2
>> > + addi t6, a0, 128*2*3
>> > + vle8.v v0, (a1)
>> > + vle8.v v4, (t1)
>> > + vle8.v v8, (t2)
>> > + vle8.v v12, (t3)
>> > + addi a3, a3, -4
>> > + vwmulu.vx v16, v0, t0
>> > + vwmulu.vx v20, v4, t0
>> > + vwmulu.vx v24, v8, t0
>> > + vwmulu.vx v28, v12, t0
>> > + vse16.v v16, (a0)
>> > + vse16.v v20, (t4)
>> > + vse16.v v24, (t5)
>> > + vse16.v v28, (t6)
>> > + sh2add a1, a2, a1
>> > + add a0, a0, 128*2*4
>> > + bnez a3, 1b
>> > + ret
>> > +endfunc
>>
>> Is 4x unroll really faster than 2x here? We don't typically unroll 4x
>> manually.
>>
>> > +
>> > +.macro dmvr_h_v mn, type
>> > + lla t4, ff_vvc_inter_luma_dmvr_filters
>> > + sh1add t4, \mn, t4
>> > + lbu t5, (t4)
>> > + lbu t6, 1(t4)
>> > +1:
>> > +.ifc \type,h
>> > + addi t0, a1, 1
>> > + addi t1, a1, 2
>> > +.else
>> > + add t0, a1, a2
>> > + add t1, t0, a2
>> > +.endif
>> > + vle8.v v0, (a1)
>> > + vle8.v v4, (t0)
>> > + vle8.v v8, (t1)
>> > + addi a3, a3, -2
>> > + vzext.vf2 v12, v0
>> > + vzext.vf2 v16, v4
>> > + vzext.vf2 v20, v8
>> > + addi t2, a0, 128*2
>> > + vmul.vx v12, v12, t5
>>
>> t5 seems to be 8-bit, so vwmulu.vx should work better here? Since you
>> leveraged it in the previous function, I'm a bit confused why not here,
>> TBH.
>>
>> > + vmul.vx v24, v16, t5
>> > + vmacc.vx v12, t6, v16
>> > + vmacc.vx v24, t6, v20
>>
>> Likewise vwmaccu.vx.
>>
>> > + vadd.vi v12, v12, 2
>> > + vadd.vi v24, v24, 2
>> > + vsra.vi v12, v12, 2
>> > + vsra.vi v24, v24, 2
>>
>> Missing rounding opportunity, vssra.vi should work better here.
>>
>> > + vse16.v v12, (a0)
>> > + vse16.v v24, (t2)
>> > + add a0, a0, 128*4
>> > + sh1add a1, a2, a1
>> > + bnez a3, 1b
>> > + ret
>> > +.endm
>> > +
>> > +func dmvr_h zve32x, zbb, zba
>> > + lpad 0
>> > + dmvr_h_v a4, h
>> > +endfunc
>> > +
>> > +func dmvr_v zve32x, zbb, zba
>> > + lpad 0
>> > + dmvr_h_v a5, v
>> > +endfunc
>> > +
>> > +.macro dmvr_load_h dst, filter0, filter1
>> > + addi a6, a1, 1
>> > + vle8.v \dst, (a1)
>> > + vle8.v v2, (a6)
>> > + vzext.vf2 v4, \dst
>> > + vzext.vf2 v8, v2
>> > + vmul.vx \dst, v4, \filter0
>> > + vmacc.vx \dst, \filter1, v8
>> > + vadd.vi \dst, \dst, 1 << (2 - 1)
>> > + vsra.vi \dst, \dst, 2
>>
>> Same comments.
>>
>> > +.endm
>> > +
>> > +func dmvr_hv zve32x, zbb, zba
>> > + lpad 0
>> > + lla t0, ff_vvc_inter_luma_dmvr_filters
>> > + sh1add t1, a4, t0
>> > + sh1add t2, a5, t0
>> > + lbu t3, (t1) // filter[mx][0]
>> > + lbu t4, 1(t1) // filter[mx][1]
>> > + lbu t5, (t2) // filter[my][0]
>> > + lbu t6, 1(t2) // filter[my][1]
>> > + dmvr_load_h v12, t3, t4
>> > + add a1, a1, a2
>> > +1:
>> > + vmul.vx v28, v12, t5
>> > + addi a3, a3, -1
>> > + dmvr_load_h v12, t3, t4
>> > + vmacc.vx v28, t6, v12
>> > + vadd.vi v28, v28, 1 << (4 - 1)
>> > + vsra.vi v28, v28, 4
>>
>> Ditto.
>>
>> > + vse16.v v28, (a0)
>> > + add a1, a1, a2
>> > + addi a0, a0, 128*2
>> > + bnez a3, 1b
>> > + ret
>> > +endfunc
>> > +
>> > +.macro func_dmvr vlen, name
>> > +func ff_vvc_\name\()_8_rvv_\vlen\(), zve32x, zbb, zba
>> > + lpad 0
>> > + li t0, 20
>> > + beq a6, t0, DMVR20\vlen\name
>> > + .ifc \name, dmvr
>> > + vsetvlstatic8 12, \vlen
>> > + .else
>> > + vsetvlstatic16 12, \vlen
>> > + .endif
>> > + j \name
>> > +DMVR20\vlen\name:
>> > + .ifc \name, dmvr
>> > + vsetvlstatic8 20, \vlen
>> > + .else
>> > + vsetvlstatic16 20, \vlen
>> > + .endif
>> > + j \name
>> > +endfunc
>> > +.endm
>> > +
>> > +.irp vlen,256,128
>> > +func_dmvr \vlen, dmvr
>> > +func_dmvr \vlen, dmvr_h
>> > +func_dmvr \vlen, dmvr_v
>> > +func_dmvr \vlen, dmvr_hv
>> > +.endr
>> > diff --git a/libavcodec/riscv/vvc/vvcdsp_init.c
>> > b/libavcodec/riscv/vvc/vvcdsp_init.c index ac1e7dda7d..7df3ce58db 100644
>> > --- a/libavcodec/riscv/vvc/vvcdsp_init.c
>> > +++ b/libavcodec/riscv/vvc/vvcdsp_init.c
>> > @@ -37,6 +37,26 @@ void bf(ff_vvc_w_avg, bd, opt)(uint8_t *dst, ptrdiff_t
>> > dst_stride, AVG_PROTOTYPES(8, rvv_128)
>> > AVG_PROTOTYPES(8, rvv_256)
>> >
>> > +#define DMVR_PROTOTYPES(bd, opt)
>>
>> > \ +void ff_vvc_dmvr_##bd##_##opt(int16_t *dst,
>> > const uint8_t *src, ptrdiff_t src_stride, \ + int
>> height,
>> > intptr_t mx, intptr_t my, int width);
>>
>> > \ +void ff_vvc_dmvr_h_##bd##_##opt(int16_t *dst, const uint8_t
>> *src,
>> > ptrdiff_t src_stride, \ + int height, intptr_t mx,
>> intptr_t
>> > my, int width); \ +void
>> > ff_vvc_dmvr_v_##bd##_##opt(int16_t *dst, const uint8_t *src, ptrdiff_t
>> > src_stride, \ + int height, intptr_t mx, intptr_t my, int
>> > width); \ +void
>> > ff_vvc_dmvr_hv_##bd##_##opt(int16_t *dst, const uint8_t *src, ptrdiff_t
>> > src_stride, \ + int height, intptr_t mx, intptr_t my, int
>> > width); \ +
>> > +DMVR_PROTOTYPES(8, rvv_128)
>> > +DMVR_PROTOTYPES(8, rvv_256)
>> > +
>> > +#define DMVR_INIT(bd, opt) do { \
>> > + c->inter.dmvr[0][0] = ff_vvc_dmvr_##bd##_##opt; \
>> > + c->inter.dmvr[0][1] = ff_vvc_dmvr_h_##bd##_##opt; \
>> > + c->inter.dmvr[1][0] = ff_vvc_dmvr_v_##bd##_##opt; \
>> > + c->inter.dmvr[1][1] = ff_vvc_dmvr_hv_##bd##_##opt; \
>> > +} while (0)
>> > +
>> > void ff_vvc_dsp_init_riscv(VVCDSPContext *const c, const int bd)
>> > {
>> > #if HAVE_RVV
>> > @@ -51,6 +71,7 @@ void ff_vvc_dsp_init_riscv(VVCDSPContext *const c,
>> const
>> > int bd) # if (__riscv_xlen == 64)
>> > c->inter.w_avg = ff_vvc_w_avg_8_rvv_256;
>> > # endif
>> > + DMVR_INIT(8, rvv_256);
>> > break;
>> > default:
>> > break;
>> > @@ -63,6 +84,7 @@ void ff_vvc_dsp_init_riscv(VVCDSPContext *const c,
>> const
>> > int bd) # if (__riscv_xlen == 64)
>> > c->inter.w_avg = ff_vvc_w_avg_8_rvv_128;
>> > # endif
>> > + DMVR_INIT(8, rvv_128);
>> > break;
>> > default:
>> > break;
>>
>>
>> --
>> 雷米‧德尼-库尔蒙
>> http://www.remlab.net/
>>
>>
>>
>> _______________________________________________
>> ffmpeg-devel mailing list
>> ffmpeg-devel at ffmpeg.org
>> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>>
>> To unsubscribe, visit link above, or email
>> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
>>
>_______________________________________________
>ffmpeg-devel mailing list
>ffmpeg-devel at ffmpeg.org
>https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
>To unsubscribe, visit link above, or email
>ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
More information about the ffmpeg-devel
mailing list