[FFmpeg-devel] [PATCH] lavc/vvc_mc: R-V V avg w_avg
Rémi Denis-Courmont
remi at remlab.net
Sat May 25 11:27:49 EEST 2024
Hi,
Sorry for the delay, this is not my day job.
Le tiistaina 21. toukokuuta 2024, 22.24.21 EEST uk7b at foxmail.com a écrit :
> From: sunyuechi <sunyuechi at iscas.ac.cn>
> ---
> libavcodec/riscv/Makefile | 2 +
> libavcodec/riscv/vvc_mc_rvv.S | 313 +++++++++++++++++++++++++++++++++
> libavcodec/riscv/vvcdsp_init.c | 71 ++++++++
In keeping in line with the rest of the project, that should probably go into
libavcodec/riscv/vvc/
> libavcodec/vvc/dsp.c | 4 +-
> libavcodec/vvc/dsp.h | 1 +
> 5 files changed, 390 insertions(+), 1 deletion(-)
> create mode 100644 libavcodec/riscv/vvc_mc_rvv.S
> create mode 100644 libavcodec/riscv/vvcdsp_init.c
>
> diff --git a/libavcodec/riscv/Makefile b/libavcodec/riscv/Makefile
> index 27b268ae39..6297664fc9 100644
> --- a/libavcodec/riscv/Makefile
> +++ b/libavcodec/riscv/Makefile
> @@ -68,3 +68,5 @@ RV-OBJS-$(CONFIG_VP9_DECODER) += riscv/vp9_intra_rvi.o \
> RVV-OBJS-$(CONFIG_VP9_DECODER) += riscv/vp9_intra_rvv.o
> OBJS-$(CONFIG_VORBIS_DECODER) += riscv/vorbisdsp_init.o
> RVV-OBJS-$(CONFIG_VORBIS_DECODER) += riscv/vorbisdsp_rvv.o
> +OBJS-$(CONFIG_VVC_DECODER) += riscv/vvcdsp_init.o
> +RVV-OBJS-$(CONFIG_VVC_DECODER) += riscv/vvc_mc_rvv.o
> diff --git a/libavcodec/riscv/vvc_mc_rvv.S b/libavcodec/riscv/vvc_mc_rvv.S
> new file mode 100644
> index 0000000000..3c2bac37f7
> --- /dev/null
> +++ b/libavcodec/riscv/vvc_mc_rvv.S
> @@ -0,0 +1,313 @@
> +/*
> + * Copyright (c) 2024 Institue of Software Chinese Academy of Sciences
> (ISCAS). + *
> + * This file is part of FFmpeg.
> + *
> + * FFmpeg is free software; you can redistribute it and/or
> + * modify it under the terms of the GNU Lesser General Public
> + * License as published by the Free Software Foundation; either
> + * version 2.1 of the License, or (at your option) any later version.
> + *
> + * FFmpeg is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
> + * Lesser General Public License for more details.
> + *
> + * You should have received a copy of the GNU Lesser General Public
> + * License along with FFmpeg; if not, write to the Free Software
> + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301
> USA + */
> +
> +#include "libavutil/riscv/asm.S"
> +
> +.macro vsetvlstatic8 w vlen is_w
> + .if \w <= 2
> + vsetivli zero, \w, e8, mf8, ta, ma
> + .elseif \w <= 4 && \vlen == 128
> + vsetivli zero, \w, e8, mf4, ta, ma
> + .elseif \w <= 4 && \vlen >= 256
> + vsetivli zero, \w, e8, mf8, ta, ma
> + .elseif \w <= 8 && \vlen == 128
> + vsetivli zero, \w, e8, mf2, ta, ma
> + .elseif \w <= 8 && \vlen >= 256
> + vsetivli zero, \w, e8, mf4, ta, ma
> + .elseif \w <= 16 && \vlen == 128
> + vsetivli zero, \w, e8, m1, ta, ma
> + .elseif \w <= 16 && \vlen >= 256
> + vsetivli zero, \w, e8, mf2, ta, ma
> + .elseif \w <= 32 && \vlen >= 256
> + li t0, \w
> + vsetvli zero, t0, e8, m1, ta, ma
> + .elseif \w <= (\vlen / 4) || \is_w
> + li t0, 64
> + vsetvli zero, t0, e8, m2, ta, ma
> + .else
> + li t0, \w
> + vsetvli zero, t0, e8, m4, ta, ma
> + .endif
> +.endm
> +
> +.macro vsetvlstatic16 w vlen is_w
> + .if \w <= 2
> + vsetivli zero, \w, e16, mf4, ta, ma
> + .elseif \w <= 4 && \vlen == 128
> + vsetivli zero, \w, e16, mf2, ta, ma
> + .elseif \w <= 4 && \vlen >= 256
> + vsetivli zero, \w, e16, mf4, ta, ma
> + .elseif \w <= 8 && \vlen == 128
> + vsetivli zero, \w, e16, m1, ta, ma
> + .elseif \w <= 8 && \vlen >= 256
> + vsetivli zero, \w, e16, mf2, ta, ma
> + .elseif \w <= 16 && \vlen == 128
> + vsetivli zero, \w, e16, m2, ta, ma
> + .elseif \w <= 16 && \vlen >= 256
> + vsetivli zero, \w, e16, m1, ta, ma
> + .elseif \w <= 32 && \vlen >= 256
> + li t0, \w
> + vsetvli zero, t0, e16, m2, ta, ma
> + .elseif \w <= (\vlen / 4) || \is_w
> + li t0, 64
> + vsetvli zero, t0, e16, m4, ta, ma
> + .else
> + li t0, \w
> + vsetvli zero, t0, e16, m8, ta, ma
> + .endif
> +.endm
> +
> +.macro vsetvlstatic32 w vlen
> + .if \w <= 2
> + vsetivli zero, \w, e32, mf2, ta, ma
> + .elseif \w <= 4 && \vlen == 128
> + vsetivli zero, \w, e32, m1, ta, ma
> + .elseif \w <= 4 && \vlen >= 256
> + vsetivli zero, \w, e32, mf2, ta, ma
> + .elseif \w <= 8 && \vlen == 128
> + vsetivli zero, \w, e32, m2, ta, ma
> + .elseif \w <= 8 && \vlen >= 256
> + vsetivli zero, \w, e32, m1, ta, ma
> + .elseif \w <= 16 && \vlen == 128
> + vsetivli zero, \w, e32, m4, ta, ma
> + .elseif \w <= 16 && \vlen >= 256
> + vsetivli zero, \w, e32, m2, ta, ma
> + .elseif \w <= 32 && \vlen >= 256
> + li t0, \w
> + vsetvli zero, t0, e32, m4, ta, ma
> + .else
> + li t0, \w
> + vsetvli zero, t0, e32, m8, ta, ma
> + .endif
> +.endm
> +
> +.macro avg_nx1 w vlen
> + vsetvlstatic16 \w, \vlen, 0
> + vle16.v v0, (a2)
> + vle16.v v8, (a3)
> + vadd.vv v8, v8, v0
> + vmax.vx v8, v8, zero
> + vsetvlstatic8 \w, \vlen, 0
> + vnclipu.wi v8, v8, 7
> + vse8.v v8, (a0)
> +.endm
> +
> +.macro avg w h vlen
> + csrw vxrm, zero
> +
> +.if \w <= (\vlen / 4) && \h >= 4
> +.rept (\h / 4)
> + addi t1, a3, 128*2
> + addi t3, a2, 128*2*2
> + add t2, a0, a1
> + sh1add t5, a1, a0
> + addi t4, a3, 128*2*2
> + addi a7, a3, 128*2*3
> + addi t6, a2, 128*2*3
> + vsetvlstatic16 \w, \vlen, 0
> + addi t0, a2, 128*2
> + add a6, t5, a1
> + vle16.v v0, (a2)
> + vle16.v v4, (a3)
> + vle16.v v8, (t0)
> + vle16.v v12, (t1)
> + vle16.v v16, (t3)
> + vle16.v v20, (t4)
> + vle16.v v24, (t6)
> + vle16.v v28, (a7)
> + vadd.vv v4, v4, v0
> + vadd.vv v12, v12, v8
> + vadd.vv v20, v20, v16
> + vadd.vv v28, v28, v24
> + vmax.vx v4, v4, zero
> + vmax.vx v12, v12, zero
> + vmax.vx v20, v20, zero
> + vmax.vx v28, v28, zero
> + vsetvlstatic8 \w, \vlen, 0
> + addi a2, a2, 128*8
> + vnclipu.wi v4, v4, 7
> + vnclipu.wi v12, v12, 7
> + vnclipu.wi v20, v20, 7
> + vnclipu.wi v28, v28, 7
> + addi a3, a3, 128*8
> + vse8.v v4, (a0)
> + vse8.v v12, (t2)
> + vse8.v v20, (t5)
> + sh2add a0, a1, a0
> + vse8.v v28, (a6)
> +.endr
> +
> +.elseif (\w <= (\vlen / 4) && \h == 2) || (\w == (\vlen / 2))
> +.rept (\h / 2)
> + vsetvlstatic16 \w, \vlen, 0
> + addi t0, a2, 128*2
> + addi t1, a3, 128*2
> + add t2, a0, a1
> + vle16.v v0, (a2)
> + vle16.v v8, (a3)
> + vle16.v v16, (t0)
> + vle16.v v24, (t1)
> + vadd.vv v8, v8, v0
> + vadd.vv v24, v24, v16
> + vmax.vx v8, v8, zero
> + vmax.vx v24, v24, zero
> + vsetvlstatic8 \w, \vlen, 0
> + addi a2, a2, 128*4
> + vnclipu.wi v8, v8, 7
> + vnclipu.wi v24, v24, 7
> + addi a3, a3, 128*4
> + vse8.v v8, (a0)
> + vse8.v v24, (t2)
> + sh1add a0, a1, a0
> +.endr
> +
> +.else
> +.rept \h
> + avg_nx1 \w, \vlen
> + .if \w == 128 && \vlen == 128
> + addi a2, a2, 64*2
> + addi a3, a3, 64*2
> + addi a0, a0, 64
> + avg_nx1 \w, \vlen
> + addi a0, a0, -64
> + addi a2, a2, 128
> + addi a3, a3, 128
> + .else
> + addi a2, a2, 128*2
> + addi a3, a3, 128*2
> + .endif
> + add a0, a0, a1
> +.endr
> +.endif
> +.endm
> +
> +.macro w_avg_nx1 w vlen
> + vsetvlstatic16 \w, \vlen, 1
> + vle16.v v0, (a2)
> + vle16.v v8, (a3)
> + vwmul.vx v16, v0, a7
> + vwmacc.vx v16, t3, v8
> + vsetvlstatic32 \w, \vlen
> + vadd.vx v16, v16, t4
> + vsetvlstatic16 \w, \vlen, 1
> + vnsrl.wx v16, v16, t6
> + vmax.vx v16, v16, zero
> + vsetvlstatic8 \w, \vlen, 1
> + vnclipu.wi v16, v16, 0
> + vse8.v v16, (a0)
> +.endm
> +
> +#if (__riscv_xlen == 64)
> +.macro w_avg w h vlen
> + csrw vxrm, zero
> + addi t6, a6, 7
> + ld t3, (sp)
> + ld t4, 8(sp)
> + ld t5, 16(sp)
> + add t4, t4, t5
> + addi t4, t4, 1 // o0 + o1 + 1
> + addi t5, t6, -1 // shift - 1
> + sll t4, t4, t5
> +
> +.if \w <= (\vlen / 8)
> + .rept (\h / 2)
> + vsetvlstatic16 \w, \vlen, 1
> + addi t0, a2, 128*2
> + addi t1, a3, 128*2
> + vle16.v v0, (a2)
> + vle16.v v8, (a3)
> + vle16.v v20, (t0)
> + vle16.v v24, (t1)
> + vwmul.vx v16, v0, a7
> + vwmul.vx v28, v20, a7
> + vwmacc.vx v16, t3, v8
> + vwmacc.vx v28, t3, v24
> + vsetvlstatic32 \w, \vlen
> + add t2, a0, a1
> + vadd.vx v16, v16, t4
> + vadd.vx v28, v28, t4
> + vsetvlstatic16 \w, \vlen, 1
> + vnsrl.wx v16, v16, t6
> + vnsrl.wx v28, v28, t6
> + vmax.vx v16, v16, zero
> + vmax.vx v28, v28, zero
> + vsetvlstatic8 \w, \vlen, 1
> + addi a2, a2, 128*4
> + vnclipu.wi v16, v16, 0
> + vnclipu.wi v28, v28, 0
> + vse8.v v16, (a0)
> + addi a3, a3, 128*4
> + vse8.v v28, (t2)
> + sh1add a0, a1, a0
> + .endr
> +.else
> + .rept \h
> + w_avg_nx1 \w, \vlen
> + .if \w == (\vlen / 2)
> + addi a2, a2, (\vlen / 2)
> + addi a3, a3, (\vlen / 2)
> + addi a0, a0, (\vlen / 4)
> + w_avg_nx1 \w, \vlen
> + addi a2, a2, -(\vlen / 2)
> + addi a3, a3, -(\vlen / 2)
> + addi a0, a0, -(\vlen / 4)
> + .elseif \w == 128 && \vlen == 128
> + .rept 3
> + addi a2, a2, (\vlen / 2)
> + addi a3, a3, (\vlen / 2)
> + addi a0, a0, (\vlen / 4)
> + w_avg_nx1 \w, \vlen
> + .endr
> + addi a2, a2, -(\vlen / 2) * 3
> + addi a3, a3, -(\vlen / 2) * 3
> + addi a0, a0, -(\vlen / 4) * 3
> + .endif
> +
> + addi a2, a2, 128*2
> + addi a3, a3, 128*2
> + add a0, a0, a1
> + .endr
> +.endif
> +.endm
> +#endif
> +
> +.macro func_avg name vlen
> +func ff_vvc_\name\()_8_rvv_\vlen\(), zve32x
> +.irp w,2,4,8,16,32,64,128
> + li t3, \w
> + bne a4, t3, \name\vlen\()end\w
> +.irp h,2,4,8,16,32,64,128
> + li t4, \h
> + bne a5, t4, \name\vlen\()end\w\h
Expanding the macro 49 times, with up to 14 branches to get there is maybe not
such a great idea. It might look nice on the checkasm µbenchmarks because the
branches under test get predicted and cached.
But in real use, branch prediction will not work so well, and the I-cache will
be filled with all variants of the same function. Indeed, this seems to result
in about .5 MiB of code. Even if only one half is needed (128-bit or 256+-bit
variants). that's a lot. For comparison, x86 uses just about 10 KiB, also with
two variants. What I make out from the arcane forbidden CISC arts there:
- functions are specialised only in one dimension, not both,
- dispatch tables avoid multiplying branches.
> + \name \w \h \vlen
> + ret
> +\name\vlen\()end\w\h:
> +.endr
> +\name\vlen\()end\w:
> +.endr
> +endfunc
> +.endm
> +
> +func_avg avg 256
> +func_avg avg 128
> +#if (__riscv_xlen == 64)
> +func_avg w_avg 256
> +func_avg w_avg 128
> +#endif
--
雷米‧德尼-库尔蒙
http://www.remlab.net/
More information about the ffmpeg-devel
mailing list