[FFmpeg-devel] [PATCH 6/6][BROKEN] avfilter/vf_nlmeans: add x86 SIMD

[James Almer]

On 10/24/2021 5:25 PM, Paul B Mahol wrote:
> Signed-off-by: Paul B Mahol <onemda at gmail.com>
> ---
>   libavfilter/vf_nlmeans.c       |  3 ++
>   libavfilter/vf_nlmeans.h       |  1 +
>   libavfilter/x86/Makefile       |  2 +
>   libavfilter/x86/vf_nlmeans.asm | 89 ++++++++++++++++++++++++++++++++++
>   4 files changed, 95 insertions(+)
>   create mode 100644 libavfilter/x86/vf_nlmeans.asm
> 
> diff --git a/libavfilter/vf_nlmeans.c b/libavfilter/vf_nlmeans.c
> index 93a14bcf19..16171d830a 100644
> --- a/libavfilter/vf_nlmeans.c
> +++ b/libavfilter/vf_nlmeans.c
> @@ -513,6 +513,9 @@ void ff_nlmeans_init(NLMeansDSPContext *dsp)
>   
>       if (ARCH_AARCH64)
>           ff_nlmeans_init_aarch64(dsp);
> +
> +    if (ARCH_X86)
> +        ff_nlmeans_init_x86(dsp);
>   }
>   
>   static av_cold int init(AVFilterContext *ctx)
> diff --git a/libavfilter/vf_nlmeans.h b/libavfilter/vf_nlmeans.h
> index d0d0056163..ae9f450dbf 100644
> --- a/libavfilter/vf_nlmeans.h
> +++ b/libavfilter/vf_nlmeans.h
> @@ -45,5 +45,6 @@ typedef struct NLMeansDSPContext {
>   
>   void ff_nlmeans_init(NLMeansDSPContext *dsp);
>   void ff_nlmeans_init_aarch64(NLMeansDSPContext *dsp);
> +void ff_nlmeans_init_x86(NLMeansDSPContext *dsp);
>   
>   #endif /* AVFILTER_NLMEANS_H */
> diff --git a/libavfilter/x86/Makefile b/libavfilter/x86/Makefile
> index a29941eaeb..e87481bd7a 100644
> --- a/libavfilter/x86/Makefile
> +++ b/libavfilter/x86/Makefile
> @@ -20,6 +20,7 @@ OBJS-$(CONFIG_LIMITER_FILTER)                += x86/vf_limiter_init.o
>   OBJS-$(CONFIG_LUT3D_FILTER)                  += x86/vf_lut3d_init.o
>   OBJS-$(CONFIG_MASKEDCLAMP_FILTER)            += x86/vf_maskedclamp_init.o
>   OBJS-$(CONFIG_MASKEDMERGE_FILTER)            += x86/vf_maskedmerge_init.o
> +OBJS-$(CONFIG_NLMEANS_FILTER)                += x86/vf_nlmeans_init.o
>   OBJS-$(CONFIG_NOISE_FILTER)                  += x86/vf_noise.o
>   OBJS-$(CONFIG_OVERLAY_FILTER)                += x86/vf_overlay_init.o
>   OBJS-$(CONFIG_PP7_FILTER)                    += x86/vf_pp7_init.o
> @@ -61,6 +62,7 @@ X86ASM-OBJS-$(CONFIG_LIMITER_FILTER)         += x86/vf_limiter.o
>   X86ASM-OBJS-$(CONFIG_LUT3D_FILTER)           += x86/vf_lut3d.o
>   X86ASM-OBJS-$(CONFIG_MASKEDCLAMP_FILTER)     += x86/vf_maskedclamp.o
>   X86ASM-OBJS-$(CONFIG_MASKEDMERGE_FILTER)     += x86/vf_maskedmerge.o
> +X86ASM-OBJS-$(CONFIG_NLMEANS_FILTER)         += x86/vf_nlmeans.o
>   X86ASM-OBJS-$(CONFIG_OVERLAY_FILTER)         += x86/vf_overlay.o
>   X86ASM-OBJS-$(CONFIG_PP7_FILTER)             += x86/vf_pp7.o
>   X86ASM-OBJS-$(CONFIG_PSNR_FILTER)            += x86/vf_psnr.o
> diff --git a/libavfilter/x86/vf_nlmeans.asm b/libavfilter/x86/vf_nlmeans.asm
> new file mode 100644
> index 0000000000..aebcc59b54
> --- /dev/null
> +++ b/libavfilter/x86/vf_nlmeans.asm
> @@ -0,0 +1,89 @@
> +;*****************************************************************************
> +;* x86-optimized functions for nlmeans filter
> +;*
> +;* 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/x86/x86util.asm"
> +
> +%if HAVE_AVX2_EXTERNAL
> +
> +SECTION_RODATA
> +
> +SECTION .text
> +
> +; void ff_compute_weights_line(const uint32_t *const iia,
> +;                              const uint32_t *const iib,
> +;                              const uint32_t *const iid,
> +;                              const uint32_t *const iie,
> +;                              const uint8_t *const src,
> +;                              struct weighted_avg *wa,
> +;                              const float *const lut,
> +;                              int max,
> +;                              int startx, int endx);
> +
> +INIT_YMM avx2
> +cglobal compute_weights_line, 11, 11, 7, iia, iib, iid, iie, src, wa, lut, max, startx, endx, x

There are 10 input arguments, not 11.

> +    movsxdifnidn startxq, startxd
> +    movsxdifnidn   endxq, endxd
> +    movsxdifnidn    maxq, maxd

If you're going to load this from memory into m4 below, then don't 
bother loading it into a gpr at all.

> +
> +    sal          startxq, 2
> +    sal            endxq, 2
> +
> +    mov               xq, startxq
> +    sar          startxq, 2
> +    VBROADCASTI128    m4, maxm

vbroadcasti128 reads 128 bits from memory, and max is an int. You want 
vpbroadcastd.

> +    pcmpeqd           m5, m5
> +
> +    .loop:
> +        movu  m0, [iieq + xq]
> +        movu  m1, [iidq + xq]
> +        movu  m2, [iibq + xq]
> +        movu  m3, [iiaq + xq]
> +        vpmovzxbd  m6, [srcq + startxq]
> +        vcvtdq2ps m6, m6
> +
> +        psubd  m0, m1
> +        psubd  m0, m2
> +        paddd  m0, m3
> +        pminud m0, m4
> +        pslld  m0, 2
> +        mova   m3, m5
> +        vpgatherdd m1, [lutq + m0], m3
> +
> +        vmulps    m2, m1, m6
> +        vunpcklps m0, m1, m2
> +        vunpckhps m1, m1, m2
> +
> +        movu m2, [waq + xq * 2]
> +        movu m3, [waq + xq * 2 + 4 * 4]
> +
> +        vaddps m0, m2
> +        vaddps m1, m3
> +
> +        movu [waq + xq * 2], m0
> +        movu [waq + xq * 2 + 4 * 4], m1
> +
> +        add startxq, 1 * 4
> +        add xq, 4 * 4
> +        cmp xq, endxq
> +        jl .loop
> +    RET
> +
> +%endif
>