diff mbox

[FFmpeg-devel] avcodec/arm/hevcdsp_sao : add NEON optimization for sao

Message ID 20180308070331.67304-1-fanyingming@bytedance.com
State Superseded
Headers show

Commit Message

Yingming Fan March 8, 2018, 7:03 a.m. UTC
From: Meng Wang <wangmeng.kids@bytedance.com>

Signed-off-by: Meng Wang <wangmeng.kids@bytedance.com>
---
As FFmpeg hevc decoder have no SAO neon optimization, we add sao_band and sao_edge neon codes in this patch.
I have already submit a patch called 'checkasm/hevc_sao : add hevc_sao for checkasm' several days ago.
Results below was printed by hevc_sao checkasm on an armv7 device Nexus 5. 
From the results we can see: hevc_sao_band speed up ~2x, hevc_sao_edge speed up ~4x. 
Also test FATE under armv7 device and MacOS.

hevc_sao_band_8x8_8_c: 804.9
hevc_sao_band_8x8_8_neon: 452.4
hevc_sao_band_16x16_8_c: 2638.1
hevc_sao_band_16x16_8_neon: 1169.9
hevc_sao_band_32x32_8_c: 9259.9
hevc_sao_band_32x32_8_neon: 3956.1
hevc_sao_band_48x48_8_c: 20344.6
hevc_sao_band_48x48_8_neon: 8649.6
hevc_sao_band_64x64_8_c: 35684.6
hevc_sao_band_64x64_8_neon: 15213.1
hevc_sao_edge_8x8_8_c: 1761.6
hevc_sao_edge_8x8_8_neon: 414.6
hevc_sao_edge_16x16_8_c: 6844.4
hevc_sao_edge_16x16_8_neon: 1589.9
hevc_sao_edge_32x32_8_c: 27156.4
hevc_sao_edge_32x32_8_neon: 6116.6
hevc_sao_edge_48x48_8_c: 60004.6
hevc_sao_edge_48x48_8_neon: 13686.4
hevc_sao_edge_64x64_8_c: 106708.1
hevc_sao_edge_64x64_8_neon: 24240.1

 libavcodec/arm/Makefile            |   3 +-
 libavcodec/arm/hevcdsp_init_neon.c |  63 +++++++++++++
 libavcodec/arm/hevcdsp_sao_neon.S  | 181 +++++++++++++++++++++++++++++++++++++
 3 files changed, 246 insertions(+), 1 deletion(-)
 create mode 100644 libavcodec/arm/hevcdsp_sao_neon.S

Comments

Yingming Fan March 11, 2018, 2:27 a.m. UTC | #1
Hi, there. 
I have already pushed a patch which add hevc_sao checkasm and patch was adopted.
You can verify this optimization by using checkasm under arm device, `checkasm --test=hevc_sao --bench`.
hevc_sao_band speed up ~2x, hevc_sao_edge speed up ~4x. Also passed FATE under arm platform.

Yingming Fan

> On 8 Mar 2018, at 3:03 PM, Yingming Fan <yingmingfan@gmail.com> wrote:
> 
> From: Meng Wang <wangmeng.kids@bytedance.com>
> 
> Signed-off-by: Meng Wang <wangmeng.kids@bytedance.com>
> ---
> As FFmpeg hevc decoder have no SAO neon optimization, we add sao_band and sao_edge neon codes in this patch.
> I have already submit a patch called 'checkasm/hevc_sao : add hevc_sao for checkasm' several days ago.
> Results below was printed by hevc_sao checkasm on an armv7 device Nexus 5. 
> From the results we can see: hevc_sao_band speed up ~2x, hevc_sao_edge speed up ~4x. 
> Also test FATE under armv7 device and MacOS.
> 
> hevc_sao_band_8x8_8_c: 804.9
> hevc_sao_band_8x8_8_neon: 452.4
> hevc_sao_band_16x16_8_c: 2638.1
> hevc_sao_band_16x16_8_neon: 1169.9
> hevc_sao_band_32x32_8_c: 9259.9
> hevc_sao_band_32x32_8_neon: 3956.1
> hevc_sao_band_48x48_8_c: 20344.6
> hevc_sao_band_48x48_8_neon: 8649.6
> hevc_sao_band_64x64_8_c: 35684.6
> hevc_sao_band_64x64_8_neon: 15213.1
> hevc_sao_edge_8x8_8_c: 1761.6
> hevc_sao_edge_8x8_8_neon: 414.6
> hevc_sao_edge_16x16_8_c: 6844.4
> hevc_sao_edge_16x16_8_neon: 1589.9
> hevc_sao_edge_32x32_8_c: 27156.4
> hevc_sao_edge_32x32_8_neon: 6116.6
> hevc_sao_edge_48x48_8_c: 60004.6
> hevc_sao_edge_48x48_8_neon: 13686.4
> hevc_sao_edge_64x64_8_c: 106708.1
> hevc_sao_edge_64x64_8_neon: 24240.1
> 
> libavcodec/arm/Makefile            |   3 +-
> libavcodec/arm/hevcdsp_init_neon.c |  63 +++++++++++++
> libavcodec/arm/hevcdsp_sao_neon.S  | 181 +++++++++++++++++++++++++++++++++++++
> 3 files changed, 246 insertions(+), 1 deletion(-)
> create mode 100644 libavcodec/arm/hevcdsp_sao_neon.S
> 
> diff --git a/libavcodec/arm/Makefile b/libavcodec/arm/Makefile
> index 1eeac5449e..2ee913e8a8 100644
> --- a/libavcodec/arm/Makefile
> +++ b/libavcodec/arm/Makefile
> @@ -136,7 +136,8 @@ NEON-OBJS-$(CONFIG_DCA_DECODER)        += arm/synth_filter_neon.o
> NEON-OBJS-$(CONFIG_HEVC_DECODER)       += arm/hevcdsp_init_neon.o       \
>                                           arm/hevcdsp_deblock_neon.o    \
>                                           arm/hevcdsp_idct_neon.o       \
> -                                          arm/hevcdsp_qpel_neon.o
> +                                          arm/hevcdsp_qpel_neon.o       \
> +				          arm/hevcdsp_sao_neon.o
> NEON-OBJS-$(CONFIG_RV30_DECODER)       += arm/rv34dsp_neon.o
> NEON-OBJS-$(CONFIG_RV40_DECODER)       += arm/rv34dsp_neon.o            \
>                                           arm/rv40dsp_neon.o
> diff --git a/libavcodec/arm/hevcdsp_init_neon.c b/libavcodec/arm/hevcdsp_init_neon.c
> index a4628d2a93..3c480f12f8 100644
> --- a/libavcodec/arm/hevcdsp_init_neon.c
> +++ b/libavcodec/arm/hevcdsp_init_neon.c
> @@ -21,8 +21,16 @@
> #include "libavutil/attributes.h"
> #include "libavutil/arm/cpu.h"
> #include "libavcodec/hevcdsp.h"
> +#include "libavcodec/avcodec.h"
> #include "hevcdsp_arm.h"
> 
> +void ff_hevc_sao_band_filter_neon_8_wrapper(uint8_t *_dst, uint8_t *_src,
> +                                  ptrdiff_t stride_dst, ptrdiff_t stride_src,
> +                                  int16_t *sao_offset_val, int sao_left_class,
> +                                  int width, int height);
> +void ff_hevc_sao_edge_filter_neon_8_wrapper(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val,
> +                                  int eo, int width, int height);
> +
> void ff_hevc_v_loop_filter_luma_neon(uint8_t *_pix, ptrdiff_t _stride, int _beta, int *_tc, uint8_t *_no_p, uint8_t *_no_q);
> void ff_hevc_h_loop_filter_luma_neon(uint8_t *_pix, ptrdiff_t _stride, int _beta, int *_tc, uint8_t *_no_p, uint8_t *_no_q);
> void ff_hevc_v_loop_filter_chroma_neon(uint8_t *_pix, ptrdiff_t _stride, int *_tc, uint8_t *_no_p, uint8_t *_no_q);
> @@ -142,6 +150,51 @@ QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h3v2_neon_8);
> QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h3v3_neon_8);
> #undef QPEL_FUNC_UW
> 
> +void ff_hevc_sao_band_filter_neon_8(uint8_t *dst, uint8_t *src, ptrdiff_t stride_dst, ptrdiff_t stride_src, int width, int height, int16_t *offset_table);
> +
> +void ff_hevc_sao_band_filter_neon_8_wrapper(uint8_t *_dst, uint8_t *_src,
> +                                  ptrdiff_t stride_dst, ptrdiff_t stride_src,
> +                                  int16_t *sao_offset_val, int sao_left_class,
> +                                  int width, int height) {
> +    uint8_t *dst = (uint8_t *)_dst;
> +    uint8_t *src = (uint8_t *)_src;
> +    int16_t offset_table[32] = {0};
> +    int k;
> +
> +    stride_dst /= sizeof(uint8_t);
> +    stride_src /= sizeof(uint8_t);
> +
> +    for (k = 0; k < 4; k++) {
> +        offset_table[(k + sao_left_class) & 31] = sao_offset_val[k + 1];
> +    }
> +
> +    ff_hevc_sao_band_filter_neon_8(dst, src, stride_dst, stride_src, width, height, offset_table);
> +}
> +
> +void ff_hevc_sao_edge_filter_neon_8(uint8_t *dst, uint8_t *src, ptrdiff_t stride_dst, ptrdiff_t stride_src, int width, int height,
> +                                    int a_stride, int b_stride, int16_t *sao_offset_val, uint8_t *edge_idx);
> +
> +void ff_hevc_sao_edge_filter_neon_8_wrapper(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val,
> +                                  int eo, int width, int height) {
> +    static uint8_t edge_idx[] = { 1, 2, 0, 3, 4 };
> +    static const int8_t pos[4][2][2] = {
> +        { { -1,  0 }, {  1, 0 } }, // horizontal
> +        { {  0, -1 }, {  0, 1 } }, // vertical
> +        { { -1, -1 }, {  1, 1 } }, // 45 degree
> +        { {  1, -1 }, { -1, 1 } }, // 135 degree
> +    };
> +    uint8_t *dst = (uint8_t *)_dst;
> +    uint8_t *src = (uint8_t *)_src;
> +    int a_stride, b_stride;
> +    ptrdiff_t stride_src = (2*MAX_PB_SIZE + AV_INPUT_BUFFER_PADDING_SIZE) / sizeof(uint8_t);
> +    stride_dst /= sizeof(uint8_t);
> +
> +    a_stride = pos[eo][0][0] + pos[eo][0][1] * stride_src;
> +    b_stride = pos[eo][1][0] + pos[eo][1][1] * stride_src;
> +
> +    ff_hevc_sao_edge_filter_neon_8(dst, src, stride_dst, stride_src, width, height, a_stride, b_stride, sao_offset_val, edge_idx);
> +}
> +
> void ff_hevc_put_qpel_neon_wrapper(int16_t *dst, uint8_t *src, ptrdiff_t srcstride,
>                                    int height, intptr_t mx, intptr_t my, int width) {
> 
> @@ -168,6 +221,16 @@ av_cold void ff_hevc_dsp_init_neon(HEVCDSPContext *c, const int bit_depth)
>         c->hevc_h_loop_filter_luma     = ff_hevc_h_loop_filter_luma_neon;
>         c->hevc_v_loop_filter_chroma   = ff_hevc_v_loop_filter_chroma_neon;
>         c->hevc_h_loop_filter_chroma   = ff_hevc_h_loop_filter_chroma_neon;
> +        c->sao_band_filter[0]          = ff_hevc_sao_band_filter_neon_8_wrapper;
> +        c->sao_band_filter[1]          = ff_hevc_sao_band_filter_neon_8_wrapper;
> +        c->sao_band_filter[2]          = ff_hevc_sao_band_filter_neon_8_wrapper;
> +        c->sao_band_filter[3]          = ff_hevc_sao_band_filter_neon_8_wrapper;
> +        c->sao_band_filter[4]          = ff_hevc_sao_band_filter_neon_8_wrapper;
> +        c->sao_edge_filter[0]          = ff_hevc_sao_edge_filter_neon_8_wrapper;
> +        c->sao_edge_filter[1]          = ff_hevc_sao_edge_filter_neon_8_wrapper;
> +        c->sao_edge_filter[2]          = ff_hevc_sao_edge_filter_neon_8_wrapper;
> +        c->sao_edge_filter[3]          = ff_hevc_sao_edge_filter_neon_8_wrapper;
> +        c->sao_edge_filter[4]          = ff_hevc_sao_edge_filter_neon_8_wrapper;
>         c->add_residual[0]             = ff_hevc_add_residual_4x4_8_neon;
>         c->add_residual[1]             = ff_hevc_add_residual_8x8_8_neon;
>         c->add_residual[2]             = ff_hevc_add_residual_16x16_8_neon;
> diff --git a/libavcodec/arm/hevcdsp_sao_neon.S b/libavcodec/arm/hevcdsp_sao_neon.S
> new file mode 100644
> index 0000000000..347167951b
> --- /dev/null
> +++ b/libavcodec/arm/hevcdsp_sao_neon.S
> @@ -0,0 +1,181 @@
> +/*
> + * Copyright (c) 2017 Meng Wang <wangmeng.kids@bytedance.com>
> + *
> + * 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/arm/asm.S"
> +#include "neon.S"
> +
> +function ff_hevc_sao_band_filter_neon_8, export=1
> +        push    {r4-r10}
> +        ldr     r5,  [sp, #28]   // width
> +        ldr     r4,  [sp, #32]   // height
> +        ldr     r8,  [sp, #36]   // offset_table
> +        vpush   {d8-d15}
> +        mov     r12,  r4         // r12 = height
> +        mov     r6,   r0         // r6 = r0 = dst
> +        mov     r7,   r1         // r7 = r1 = src
> +        vldm    r8,   {q0-q3}
> +        vmov.u16    q15,  #1
> +        vmov.u8     q14,  #32
> +0:      pld      [r1]
> +        vld1.8   {d16},  [r1], r3
> +        cmp      r5,    #4
> +        beq      4f
> +8:      subs     r4,    #1
> +        vshr.u8  d17,   d16,  #3   // index = [src>>3]
> +        vshll.u8 q9,    d17,  #1   // lowIndex = 2*index
> +        vadd.u16 q11,   q9,   q15  // highIndex = (2*index+1) << 8
> +        vshl.u16 q10,   q11,  #8   // q10: highIndex;  q9: lowIndex;
> +        vadd.u16 q10,   q9         // combine high and low index;
> +        // Look-up Table Round 1; index range: 0-15
> +        vtbx.8   d24,   {q0-q1},   d20
> +        vtbx.8   d25,   {q0-q1},   d21
> +        // Look-up Table Round 2; index range: 16-31
> +        vsub.u8  q10,   q14        // Look-up with 8bit
> +        vtbx.8   d24,   {q2-q3},   d20
> +        vtbx.8   d25,   {q2-q3},   d21
> +        vaddw.u8 q13,   q12,       d16
> +        vqmovun.s16      d8,         q13
> +        vst1.8    d8,   [r0],      r2
> +        vld1.8   {d16}, [r1],      r3
> +        bne      8b
> +        subs     r5,    #8
> +        beq      99f
> +        mov      r4,    r12
> +        add r6, #8
> +        mov r0, r6
> +        add r7, #8
> +        mov r1, r7
> +        b        0b
> +4:      subs     r4,    #1
> +        vshr.u8  d17,   d16,  #3  // src>>3
> +        vshll.u8 q9,    d17,  #1   // lowIndex = 2*index
> +        vadd.u16 q11,   q9,   q15  // highIndex = (2*index+1) << 8
> +        vshl.u16 q10,   q11,  #8   // q10: highIndex;  q9: lowIndex;
> +        vadd.u16 q10,   q9         // combine high and low index;
> +        // Look-up Table Round 1; index range: 0-15
> +        vtbx.8   d24,   {q0-q1},   d20
> +        vtbx.8   d25,   {q0-q1},   d21
> +        // Look-up Table Round 2; index range: 16-32
> +        vsub.u8  q10,   q14        // Look-up with 8bit
> +        vtbx.8   d24,   {q2-q3},   d20
> +        vtbx.8   d25,   {q2-q3},   d21
> +        vaddw.u8 q13,   q12,       d16
> +        vqmovun.s16     d14,       q13
> +        vst1.32   d14[0],    [r0],     r2
> +        vld1.32   {d16[0]},  [r1],     r3
> +        bne      4b
> +        b        99f
> +99:
> +        vpop {d8-d15}
> +        pop  {r4-r10}
> +        bx   lr
> +endfunc
> +
> +function ff_hevc_sao_edge_filter_neon_8, export=1
> +        push    {r4-r11}
> +        ldr     r5,  [sp, #32]   // width
> +        ldr     r4,  [sp, #36]   // height
> +        ldr     r8,  [sp, #40]   // a_stride
> +        ldr     r9,  [sp, #44]   // b_stride
> +        ldr     r10, [sp, #48]   // sao_offset_val
> +        ldr     r11, [sp, #52]   // edge_idx
> +        vpush   {d8-d15}
> +        mov     r12,  r4         // r12 = height
> +        mov     r6,   r0         // r6 = r0 = dst
> +        mov     r7,   r1         // r7 = r1 = src
> +        vld1.8  {d0}, [r11]      // edge_idx tabel load in d0 5x8bit
> +        vld1.16 {q1}, [r10]      // sao_offset_val table load in q1, 5x16bit
> +        vmov.u8  d1,  #2
> +        vmov.u16 q2,  #1
> +0:      mov      r10,    r1
> +        add      r10,    r8           // src[x + a_stride]
> +        mov      r11,    r1
> +        add      r11,    r9           // src[x + b_stride]
> +        pld      [r1]
> +        vld1.8   {d16},  [r1],  r3    // src[x]  8x8bit
> +        vld1.8   {d17},  [r10], r3    // src[x + a_stride]
> +        vld1.8   {d18},  [r11], r3    // src[x + b_stride]
> +        cmp      r5,     #4
> +        beq      4f
> +8:      subs     r4,     #1
> +        vcgt.u8  d8,     d16,   d17
> +        vshr.u8  d9,     d8,    #7
> +        vclt.u8  d8,     d16,   d17
> +        vadd.u8  d8,     d9           // diff0
> +        vcgt.u8  d10,    d16,   d18
> +        vshr.u8  d11,    d10,   #7
> +        vclt.u8  d10,    d16,   d18
> +        vadd.u8  d10,    d11          // diff1
> +        vadd.s8  d8,     d10
> +        vadd.s8  d8,     d1
> +        vtbx.8   d9,     {d0},  d8    // offset_val
> +        vshll.u8 q6,     d9,    #1    // lowIndex
> +        vadd.u16 q7,     q6,    q2
> +        vshl.u16 q10,    q7,    #8    // highIndex
> +        vadd.u16 q10,    q6           // combine lowIndex and highIndex, offset_val
> +        vtbx.8   d22,    {q1},  d20
> +        vtbx.8   d23,    {q1},  d21
> +        vaddw.u8 q12,    q11,   d16
> +        vqmovun.s16      d26,   q12
> +        vst1.8   d26,    [r0],  r2
> +        vld1.8   {d16},  [r1],  r3    // src[x]  8x8bit
> +        vld1.8   {d17},  [r10], r3    // src[x + a_stride]
> +        vld1.8   {d18},  [r11], r3    // src[x + b_stride]
> +        bne      8b
> +        subs     r5,     #8
> +        beq      99f
> +        mov      r4,     r12
> +        add      r6,     #8
> +        mov      r0,     r6
> +        add      r7,     #8
> +        mov      r1,     r7
> +        b        0b
> +4:      subs     r4,    #1
> +        vcgt.u8  d8,     d16,   d17
> +        vshr.u8  d9,     d8,    #7
> +        vclt.u8  d8,     d16,   d17
> +        vadd.u8  d8,     d9           // diff0
> +        vcgt.u8  d10,    d16,   d18
> +        vshr.u8  d11,    d10,   #7
> +        vclt.u8  d10,    d16,   d18
> +        vadd.u8  d10,    d11          // diff1
> +        vadd.s8  d8,     d10
> +        vadd.s8  d8,     d1
> +        vtbx.8   d9,     {d0},  d8    // offset_val
> +        vshll.u8 q6,     d9,    #1    // lowIndex
> +        vadd.u16 q7,     q6,    q2
> +        vshl.u16 q10,    q7,    #8    // highIndex
> +        vadd.u16 q10,    q6           // combine lowIndex and highIndex, offset_val
> +        vtbx.8   d22,    {q1},  d20
> +        vtbx.8   d23,    {q1},  d21
> +        vaddw.u8 q12,    q11,   d16
> +        vqmovun.s16      d26,   q12
> +        vst1.32  d26[0], [r0],  r2
> +        vld1.32   {d16[0]},  [r1],  r3
> +        vld1.32   {d17[0]},  [r10], r3    // src[x + a_stride]
> +        vld1.32   {d18[0]},  [r11], r3    // src[x + b_stride]
> +        bne      4b
> +        b        99f
> +99:
> +        vpop {d8-d15}
> +        pop  {r4-r11}
> +        bx   lr
> +endfunc
> -- 
> 2.14.3 (Apple Git-98)
>
Carl Eugen Hoyos March 11, 2018, 12:54 p.m. UTC | #2
2018-03-08 8:03 GMT+01:00 Yingming Fan <yingmingfan@gmail.com>:
> From: Meng Wang <wangmeng.kids@bytedance.com>

> +    stride_dst /= sizeof(uint8_t);
> +    stride_src /= sizeof(uint8_t);

FFmpeg requires sizeof(uint8_t) to be 1, please simplify
your patch accordingly.

Why is the wrapper function needed?

Carl Eugen
Yingming Fan March 11, 2018, 12:59 p.m. UTC | #3
> On 11 Mar 2018, at 8:54 PM, Carl Eugen Hoyos <ceffmpeg@gmail.com> wrote:
> 
> 2018-03-08 8:03 GMT+01:00 Yingming Fan <yingmingfan@gmail.com>:
>> From: Meng Wang <wangmeng.kids@bytedance.com>
> 
>> +    stride_dst /= sizeof(uint8_t);
>> +    stride_src /= sizeof(uint8_t);
> 
> FFmpeg requires sizeof(uint8_t) to be 1, please simplify
> your patch accordingly.
> 
> Why is the wrapper function needed?

We use wrapper because codes in wrapper no need to be written with assembly, C codes more readable.

> 
> Carl Eugen
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
Yingming Fan March 19, 2018, 4:41 a.m. UTC | #4
Hi, is there any review about this patch? What’s your option about wrapper we used in this patch.

Yingming Fan

> On 11 Mar 2018, at 8:59 PM, Yingming Fan <yingmingfan@gmail.com> wrote:
> 
> 
>> On 11 Mar 2018, at 8:54 PM, Carl Eugen Hoyos <ceffmpeg@gmail.com> wrote:
>> 
>> 2018-03-08 8:03 GMT+01:00 Yingming Fan <yingmingfan@gmail.com>:
>>> From: Meng Wang <wangmeng.kids@bytedance.com>
>> 
>>> +    stride_dst /= sizeof(uint8_t);
>>> +    stride_src /= sizeof(uint8_t);
>> 
>> FFmpeg requires sizeof(uint8_t) to be 1, please simplify
>> your patch accordingly.
>> 
>> Why is the wrapper function needed?
> 
> We use wrapper because codes in wrapper no need to be written with assembly, C codes more readable.
> 
>> 
>> Carl Eugen
>> _______________________________________________
>> ffmpeg-devel mailing list
>> ffmpeg-devel@ffmpeg.org
>> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
Shengbin Meng March 22, 2018, 9:16 a.m. UTC | #5
Hi,

By checkasm benchmark, I can see a speedup of ~3x for band mode and ~6x for edge mode on my device (the device has aarch64 CPU, but I configured ffmpeg with `—arch=arm`). And FATE passed as well.

Results of a checkasm run:

$./tests/checkasm/checkasm --test=hevc_sao --bench
$ sudo ./tests/checkasm/checkasm --test=hevc_sao --bench
benchmarking with Linux Perf Monitoring API
nop: 49.8
checkasm: using random seed 1088726844
NEON:
 - hevc_sao.sao_band [OK]
 - hevc_sao.sao_edge [OK]
checkasm: all 10 tests passed
hevc_sao_band_8x8_8_c: 578.0
hevc_sao_band_8x8_8_neon: 215.3
hevc_sao_band_16x16_8_c: 2004.3
hevc_sao_band_16x16_8_neon: 680.8
hevc_sao_band_32x32_8_c: 8363.5
hevc_sao_band_32x32_8_neon: 2579.3
hevc_sao_band_48x48_8_c: 18268.3
hevc_sao_band_48x48_8_neon: 5653.3
hevc_sao_band_64x64_8_c: 32001.8
hevc_sao_band_64x64_8_neon: 9952.0
hevc_sao_edge_8x8_8_c: 1211.0
hevc_sao_edge_8x8_8_neon: 217.5
hevc_sao_edge_16x16_8_c: 4708.5
hevc_sao_edge_16x16_8_neon: 767.0
hevc_sao_edge_32x32_8_c: 18673.0
hevc_sao_edge_32x32_8_neon: 2967.3
hevc_sao_edge_48x48_8_c: 41936.3
hevc_sao_edge_48x48_8_neon: 6642.8
hevc_sao_edge_64x64_8_c: 74015.8
hevc_sao_edge_64x64_8_neon: 11781.8

Regards
Shengbin

> On 11 Mar 2018, at 10:27, Yingming Fan <yingmingfan@gmail.com> wrote:
> 
> Hi, there. 
> I have already pushed a patch which add hevc_sao checkasm and patch was adopted.
> You can verify this optimization by using checkasm under arm device, `checkasm --test=hevc_sao --bench`.
> hevc_sao_band speed up ~2x, hevc_sao_edge speed up ~4x. Also passed FATE under arm platform.
> 
> Yingming Fan
> 
>> On 8 Mar 2018, at 3:03 PM, Yingming Fan <yingmingfan@gmail.com> wrote:
>> 
>> From: Meng Wang <wangmeng.kids@bytedance.com>
>> 
>> Signed-off-by: Meng Wang <wangmeng.kids@bytedance.com>
>> ---
>> As FFmpeg hevc decoder have no SAO neon optimization, we add sao_band and sao_edge neon codes in this patch.
>> I have already submit a patch called 'checkasm/hevc_sao : add hevc_sao for checkasm' several days ago.
>> Results below was printed by hevc_sao checkasm on an armv7 device Nexus 5. 
>> From the results we can see: hevc_sao_band speed up ~2x, hevc_sao_edge speed up ~4x. 
>> Also test FATE under armv7 device and MacOS.
>> 
>> hevc_sao_band_8x8_8_c: 804.9
>> hevc_sao_band_8x8_8_neon: 452.4
>> hevc_sao_band_16x16_8_c: 2638.1
>> hevc_sao_band_16x16_8_neon: 1169.9
>> hevc_sao_band_32x32_8_c: 9259.9
>> hevc_sao_band_32x32_8_neon: 3956.1
>> hevc_sao_band_48x48_8_c: 20344.6
>> hevc_sao_band_48x48_8_neon: 8649.6
>> hevc_sao_band_64x64_8_c: 35684.6
>> hevc_sao_band_64x64_8_neon: 15213.1
>> hevc_sao_edge_8x8_8_c: 1761.6
>> hevc_sao_edge_8x8_8_neon: 414.6
>> hevc_sao_edge_16x16_8_c: 6844.4
>> hevc_sao_edge_16x16_8_neon: 1589.9
>> hevc_sao_edge_32x32_8_c: 27156.4
>> hevc_sao_edge_32x32_8_neon: 6116.6
>> hevc_sao_edge_48x48_8_c: 60004.6
>> hevc_sao_edge_48x48_8_neon: 13686.4
>> hevc_sao_edge_64x64_8_c: 106708.1
>> hevc_sao_edge_64x64_8_neon: 24240.1
>> 
>> libavcodec/arm/Makefile            |   3 +-
>> libavcodec/arm/hevcdsp_init_neon.c |  63 +++++++++++++
>> libavcodec/arm/hevcdsp_sao_neon.S  | 181 +++++++++++++++++++++++++++++++++++++
>> 3 files changed, 246 insertions(+), 1 deletion(-)
>> create mode 100644 libavcodec/arm/hevcdsp_sao_neon.S
>> 
>> diff --git a/libavcodec/arm/Makefile b/libavcodec/arm/Makefile
>> index 1eeac5449e..2ee913e8a8 100644
>> --- a/libavcodec/arm/Makefile
>> +++ b/libavcodec/arm/Makefile
>> @@ -136,7 +136,8 @@ NEON-OBJS-$(CONFIG_DCA_DECODER)        += arm/synth_filter_neon.o
>> NEON-OBJS-$(CONFIG_HEVC_DECODER)       += arm/hevcdsp_init_neon.o       \
>>                                          arm/hevcdsp_deblock_neon.o    \
>>                                          arm/hevcdsp_idct_neon.o       \
>> -                                          arm/hevcdsp_qpel_neon.o
>> +                                          arm/hevcdsp_qpel_neon.o       \
>> +				          arm/hevcdsp_sao_neon.o
>> NEON-OBJS-$(CONFIG_RV30_DECODER)       += arm/rv34dsp_neon.o
>> NEON-OBJS-$(CONFIG_RV40_DECODER)       += arm/rv34dsp_neon.o            \
>>                                          arm/rv40dsp_neon.o
>> diff --git a/libavcodec/arm/hevcdsp_init_neon.c b/libavcodec/arm/hevcdsp_init_neon.c
>> index a4628d2a93..3c480f12f8 100644
>> --- a/libavcodec/arm/hevcdsp_init_neon.c
>> +++ b/libavcodec/arm/hevcdsp_init_neon.c
>> @@ -21,8 +21,16 @@
>> #include "libavutil/attributes.h"
>> #include "libavutil/arm/cpu.h"
>> #include "libavcodec/hevcdsp.h"
>> +#include "libavcodec/avcodec.h"
>> #include "hevcdsp_arm.h"
>> 
>> +void ff_hevc_sao_band_filter_neon_8_wrapper(uint8_t *_dst, uint8_t *_src,
>> +                                  ptrdiff_t stride_dst, ptrdiff_t stride_src,
>> +                                  int16_t *sao_offset_val, int sao_left_class,
>> +                                  int width, int height);
>> +void ff_hevc_sao_edge_filter_neon_8_wrapper(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val,
>> +                                  int eo, int width, int height);
>> +
>> void ff_hevc_v_loop_filter_luma_neon(uint8_t *_pix, ptrdiff_t _stride, int _beta, int *_tc, uint8_t *_no_p, uint8_t *_no_q);
>> void ff_hevc_h_loop_filter_luma_neon(uint8_t *_pix, ptrdiff_t _stride, int _beta, int *_tc, uint8_t *_no_p, uint8_t *_no_q);
>> void ff_hevc_v_loop_filter_chroma_neon(uint8_t *_pix, ptrdiff_t _stride, int *_tc, uint8_t *_no_p, uint8_t *_no_q);
>> @@ -142,6 +150,51 @@ QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h3v2_neon_8);
>> QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h3v3_neon_8);
>> #undef QPEL_FUNC_UW
>> 
>> +void ff_hevc_sao_band_filter_neon_8(uint8_t *dst, uint8_t *src, ptrdiff_t stride_dst, ptrdiff_t stride_src, int width, int height, int16_t *offset_table);
>> +
>> +void ff_hevc_sao_band_filter_neon_8_wrapper(uint8_t *_dst, uint8_t *_src,
>> +                                  ptrdiff_t stride_dst, ptrdiff_t stride_src,
>> +                                  int16_t *sao_offset_val, int sao_left_class,
>> +                                  int width, int height) {
>> +    uint8_t *dst = (uint8_t *)_dst;
>> +    uint8_t *src = (uint8_t *)_src;
>> +    int16_t offset_table[32] = {0};
>> +    int k;
>> +
>> +    stride_dst /= sizeof(uint8_t);
>> +    stride_src /= sizeof(uint8_t);
>> +
>> +    for (k = 0; k < 4; k++) {
>> +        offset_table[(k + sao_left_class) & 31] = sao_offset_val[k + 1];
>> +    }
>> +
>> +    ff_hevc_sao_band_filter_neon_8(dst, src, stride_dst, stride_src, width, height, offset_table);
>> +}
>> +
>> +void ff_hevc_sao_edge_filter_neon_8(uint8_t *dst, uint8_t *src, ptrdiff_t stride_dst, ptrdiff_t stride_src, int width, int height,
>> +                                    int a_stride, int b_stride, int16_t *sao_offset_val, uint8_t *edge_idx);
>> +
>> +void ff_hevc_sao_edge_filter_neon_8_wrapper(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val,
>> +                                  int eo, int width, int height) {
>> +    static uint8_t edge_idx[] = { 1, 2, 0, 3, 4 };
>> +    static const int8_t pos[4][2][2] = {
>> +        { { -1,  0 }, {  1, 0 } }, // horizontal
>> +        { {  0, -1 }, {  0, 1 } }, // vertical
>> +        { { -1, -1 }, {  1, 1 } }, // 45 degree
>> +        { {  1, -1 }, { -1, 1 } }, // 135 degree
>> +    };
>> +    uint8_t *dst = (uint8_t *)_dst;
>> +    uint8_t *src = (uint8_t *)_src;
>> +    int a_stride, b_stride;
>> +    ptrdiff_t stride_src = (2*MAX_PB_SIZE + AV_INPUT_BUFFER_PADDING_SIZE) / sizeof(uint8_t);
>> +    stride_dst /= sizeof(uint8_t);
>> +
>> +    a_stride = pos[eo][0][0] + pos[eo][0][1] * stride_src;
>> +    b_stride = pos[eo][1][0] + pos[eo][1][1] * stride_src;
>> +
>> +    ff_hevc_sao_edge_filter_neon_8(dst, src, stride_dst, stride_src, width, height, a_stride, b_stride, sao_offset_val, edge_idx);
>> +}
>> +
>> void ff_hevc_put_qpel_neon_wrapper(int16_t *dst, uint8_t *src, ptrdiff_t srcstride,
>>                                   int height, intptr_t mx, intptr_t my, int width) {
>> 
>> @@ -168,6 +221,16 @@ av_cold void ff_hevc_dsp_init_neon(HEVCDSPContext *c, const int bit_depth)
>>        c->hevc_h_loop_filter_luma     = ff_hevc_h_loop_filter_luma_neon;
>>        c->hevc_v_loop_filter_chroma   = ff_hevc_v_loop_filter_chroma_neon;
>>        c->hevc_h_loop_filter_chroma   = ff_hevc_h_loop_filter_chroma_neon;
>> +        c->sao_band_filter[0]          = ff_hevc_sao_band_filter_neon_8_wrapper;
>> +        c->sao_band_filter[1]          = ff_hevc_sao_band_filter_neon_8_wrapper;
>> +        c->sao_band_filter[2]          = ff_hevc_sao_band_filter_neon_8_wrapper;
>> +        c->sao_band_filter[3]          = ff_hevc_sao_band_filter_neon_8_wrapper;
>> +        c->sao_band_filter[4]          = ff_hevc_sao_band_filter_neon_8_wrapper;
>> +        c->sao_edge_filter[0]          = ff_hevc_sao_edge_filter_neon_8_wrapper;
>> +        c->sao_edge_filter[1]          = ff_hevc_sao_edge_filter_neon_8_wrapper;
>> +        c->sao_edge_filter[2]          = ff_hevc_sao_edge_filter_neon_8_wrapper;
>> +        c->sao_edge_filter[3]          = ff_hevc_sao_edge_filter_neon_8_wrapper;
>> +        c->sao_edge_filter[4]          = ff_hevc_sao_edge_filter_neon_8_wrapper;
>>        c->add_residual[0]             = ff_hevc_add_residual_4x4_8_neon;
>>        c->add_residual[1]             = ff_hevc_add_residual_8x8_8_neon;
>>        c->add_residual[2]             = ff_hevc_add_residual_16x16_8_neon;
>> diff --git a/libavcodec/arm/hevcdsp_sao_neon.S b/libavcodec/arm/hevcdsp_sao_neon.S
>> new file mode 100644
>> index 0000000000..347167951b
>> --- /dev/null
>> +++ b/libavcodec/arm/hevcdsp_sao_neon.S
>> @@ -0,0 +1,181 @@
>> +/*
>> + * Copyright (c) 2017 Meng Wang <wangmeng.kids@bytedance.com>
>> + *
>> + * 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/arm/asm.S"
>> +#include "neon.S"
>> +
>> +function ff_hevc_sao_band_filter_neon_8, export=1
>> +        push    {r4-r10}
>> +        ldr     r5,  [sp, #28]   // width
>> +        ldr     r4,  [sp, #32]   // height
>> +        ldr     r8,  [sp, #36]   // offset_table
>> +        vpush   {d8-d15}
>> +        mov     r12,  r4         // r12 = height
>> +        mov     r6,   r0         // r6 = r0 = dst
>> +        mov     r7,   r1         // r7 = r1 = src
>> +        vldm    r8,   {q0-q3}
>> +        vmov.u16    q15,  #1
>> +        vmov.u8     q14,  #32
>> +0:      pld      [r1]
>> +        vld1.8   {d16},  [r1], r3
>> +        cmp      r5,    #4
>> +        beq      4f
>> +8:      subs     r4,    #1
>> +        vshr.u8  d17,   d16,  #3   // index = [src>>3]
>> +        vshll.u8 q9,    d17,  #1   // lowIndex = 2*index
>> +        vadd.u16 q11,   q9,   q15  // highIndex = (2*index+1) << 8
>> +        vshl.u16 q10,   q11,  #8   // q10: highIndex;  q9: lowIndex;
>> +        vadd.u16 q10,   q9         // combine high and low index;
>> +        // Look-up Table Round 1; index range: 0-15
>> +        vtbx.8   d24,   {q0-q1},   d20
>> +        vtbx.8   d25,   {q0-q1},   d21
>> +        // Look-up Table Round 2; index range: 16-31
>> +        vsub.u8  q10,   q14        // Look-up with 8bit
>> +        vtbx.8   d24,   {q2-q3},   d20
>> +        vtbx.8   d25,   {q2-q3},   d21
>> +        vaddw.u8 q13,   q12,       d16
>> +        vqmovun.s16      d8,         q13
>> +        vst1.8    d8,   [r0],      r2
>> +        vld1.8   {d16}, [r1],      r3
>> +        bne      8b
>> +        subs     r5,    #8
>> +        beq      99f
>> +        mov      r4,    r12
>> +        add r6, #8
>> +        mov r0, r6
>> +        add r7, #8
>> +        mov r1, r7
>> +        b        0b
>> +4:      subs     r4,    #1
>> +        vshr.u8  d17,   d16,  #3  // src>>3
>> +        vshll.u8 q9,    d17,  #1   // lowIndex = 2*index
>> +        vadd.u16 q11,   q9,   q15  // highIndex = (2*index+1) << 8
>> +        vshl.u16 q10,   q11,  #8   // q10: highIndex;  q9: lowIndex;
>> +        vadd.u16 q10,   q9         // combine high and low index;
>> +        // Look-up Table Round 1; index range: 0-15
>> +        vtbx.8   d24,   {q0-q1},   d20
>> +        vtbx.8   d25,   {q0-q1},   d21
>> +        // Look-up Table Round 2; index range: 16-32
>> +        vsub.u8  q10,   q14        // Look-up with 8bit
>> +        vtbx.8   d24,   {q2-q3},   d20
>> +        vtbx.8   d25,   {q2-q3},   d21
>> +        vaddw.u8 q13,   q12,       d16
>> +        vqmovun.s16     d14,       q13
>> +        vst1.32   d14[0],    [r0],     r2
>> +        vld1.32   {d16[0]},  [r1],     r3
>> +        bne      4b
>> +        b        99f
>> +99:
>> +        vpop {d8-d15}
>> +        pop  {r4-r10}
>> +        bx   lr
>> +endfunc
>> +
>> +function ff_hevc_sao_edge_filter_neon_8, export=1
>> +        push    {r4-r11}
>> +        ldr     r5,  [sp, #32]   // width
>> +        ldr     r4,  [sp, #36]   // height
>> +        ldr     r8,  [sp, #40]   // a_stride
>> +        ldr     r9,  [sp, #44]   // b_stride
>> +        ldr     r10, [sp, #48]   // sao_offset_val
>> +        ldr     r11, [sp, #52]   // edge_idx
>> +        vpush   {d8-d15}
>> +        mov     r12,  r4         // r12 = height
>> +        mov     r6,   r0         // r6 = r0 = dst
>> +        mov     r7,   r1         // r7 = r1 = src
>> +        vld1.8  {d0}, [r11]      // edge_idx tabel load in d0 5x8bit
>> +        vld1.16 {q1}, [r10]      // sao_offset_val table load in q1, 5x16bit
>> +        vmov.u8  d1,  #2
>> +        vmov.u16 q2,  #1
>> +0:      mov      r10,    r1
>> +        add      r10,    r8           // src[x + a_stride]
>> +        mov      r11,    r1
>> +        add      r11,    r9           // src[x + b_stride]
>> +        pld      [r1]
>> +        vld1.8   {d16},  [r1],  r3    // src[x]  8x8bit
>> +        vld1.8   {d17},  [r10], r3    // src[x + a_stride]
>> +        vld1.8   {d18},  [r11], r3    // src[x + b_stride]
>> +        cmp      r5,     #4
>> +        beq      4f
>> +8:      subs     r4,     #1
>> +        vcgt.u8  d8,     d16,   d17
>> +        vshr.u8  d9,     d8,    #7
>> +        vclt.u8  d8,     d16,   d17
>> +        vadd.u8  d8,     d9           // diff0
>> +        vcgt.u8  d10,    d16,   d18
>> +        vshr.u8  d11,    d10,   #7
>> +        vclt.u8  d10,    d16,   d18
>> +        vadd.u8  d10,    d11          // diff1
>> +        vadd.s8  d8,     d10
>> +        vadd.s8  d8,     d1
>> +        vtbx.8   d9,     {d0},  d8    // offset_val
>> +        vshll.u8 q6,     d9,    #1    // lowIndex
>> +        vadd.u16 q7,     q6,    q2
>> +        vshl.u16 q10,    q7,    #8    // highIndex
>> +        vadd.u16 q10,    q6           // combine lowIndex and highIndex, offset_val
>> +        vtbx.8   d22,    {q1},  d20
>> +        vtbx.8   d23,    {q1},  d21
>> +        vaddw.u8 q12,    q11,   d16
>> +        vqmovun.s16      d26,   q12
>> +        vst1.8   d26,    [r0],  r2
>> +        vld1.8   {d16},  [r1],  r3    // src[x]  8x8bit
>> +        vld1.8   {d17},  [r10], r3    // src[x + a_stride]
>> +        vld1.8   {d18},  [r11], r3    // src[x + b_stride]
>> +        bne      8b
>> +        subs     r5,     #8
>> +        beq      99f
>> +        mov      r4,     r12
>> +        add      r6,     #8
>> +        mov      r0,     r6
>> +        add      r7,     #8
>> +        mov      r1,     r7
>> +        b        0b
>> +4:      subs     r4,    #1
>> +        vcgt.u8  d8,     d16,   d17
>> +        vshr.u8  d9,     d8,    #7
>> +        vclt.u8  d8,     d16,   d17
>> +        vadd.u8  d8,     d9           // diff0
>> +        vcgt.u8  d10,    d16,   d18
>> +        vshr.u8  d11,    d10,   #7
>> +        vclt.u8  d10,    d16,   d18
>> +        vadd.u8  d10,    d11          // diff1
>> +        vadd.s8  d8,     d10
>> +        vadd.s8  d8,     d1
>> +        vtbx.8   d9,     {d0},  d8    // offset_val
>> +        vshll.u8 q6,     d9,    #1    // lowIndex
>> +        vadd.u16 q7,     q6,    q2
>> +        vshl.u16 q10,    q7,    #8    // highIndex
>> +        vadd.u16 q10,    q6           // combine lowIndex and highIndex, offset_val
>> +        vtbx.8   d22,    {q1},  d20
>> +        vtbx.8   d23,    {q1},  d21
>> +        vaddw.u8 q12,    q11,   d16
>> +        vqmovun.s16      d26,   q12
>> +        vst1.32  d26[0], [r0],  r2
>> +        vld1.32   {d16[0]},  [r1],  r3
>> +        vld1.32   {d17[0]},  [r10], r3    // src[x + a_stride]
>> +        vld1.32   {d18[0]},  [r11], r3    // src[x + b_stride]
>> +        bne      4b
>> +        b        99f
>> +99:
>> +        vpop {d8-d15}
>> +        pop  {r4-r11}
>> +        bx   lr
>> +endfunc
>> -- 
>> 2.14.3 (Apple Git-98)
>> 
> 
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
Shengbin Meng March 22, 2018, 9:31 a.m. UTC | #6
The code looks good to me. I think the wrapper is fine, because that part of code is not suitable for NEON assembly.

But you can remove the using of `sizeof(uint8_t)` as suggested by Carl.

Shengbin Meng

> On 19 Mar 2018, at 12:41, Yingming Fan <yingmingfan@gmail.com> wrote:
> 
> Hi, is there any review about this patch? What’s your option about wrapper we used in this patch.
> 
> Yingming Fan
> 
>> On 11 Mar 2018, at 8:59 PM, Yingming Fan <yingmingfan@gmail.com> wrote:
>> 
>> 
>>> On 11 Mar 2018, at 8:54 PM, Carl Eugen Hoyos <ceffmpeg@gmail.com> wrote:
>>> 
>>> 2018-03-08 8:03 GMT+01:00 Yingming Fan <yingmingfan@gmail.com>:
>>>> From: Meng Wang <wangmeng.kids@bytedance.com>
>>> 
>>>> +    stride_dst /= sizeof(uint8_t);
>>>> +    stride_src /= sizeof(uint8_t);
>>> 
>>> FFmpeg requires sizeof(uint8_t) to be 1, please simplify
>>> your patch accordingly.
>>> 
>>> Why is the wrapper function needed?
>> 
>> We use wrapper because codes in wrapper no need to be written with assembly, C codes more readable.
>> 
>>> 
>>> Carl Eugen
>>> _______________________________________________
>>> ffmpeg-devel mailing list
>>> ffmpeg-devel@ffmpeg.org
>>> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>> 
> 
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
diff mbox

Patch

diff --git a/libavcodec/arm/Makefile b/libavcodec/arm/Makefile
index 1eeac5449e..2ee913e8a8 100644
--- a/libavcodec/arm/Makefile
+++ b/libavcodec/arm/Makefile
@@ -136,7 +136,8 @@  NEON-OBJS-$(CONFIG_DCA_DECODER)        += arm/synth_filter_neon.o
 NEON-OBJS-$(CONFIG_HEVC_DECODER)       += arm/hevcdsp_init_neon.o       \
                                           arm/hevcdsp_deblock_neon.o    \
                                           arm/hevcdsp_idct_neon.o       \
-                                          arm/hevcdsp_qpel_neon.o
+                                          arm/hevcdsp_qpel_neon.o       \
+				          arm/hevcdsp_sao_neon.o
 NEON-OBJS-$(CONFIG_RV30_DECODER)       += arm/rv34dsp_neon.o
 NEON-OBJS-$(CONFIG_RV40_DECODER)       += arm/rv34dsp_neon.o            \
                                           arm/rv40dsp_neon.o
diff --git a/libavcodec/arm/hevcdsp_init_neon.c b/libavcodec/arm/hevcdsp_init_neon.c
index a4628d2a93..3c480f12f8 100644
--- a/libavcodec/arm/hevcdsp_init_neon.c
+++ b/libavcodec/arm/hevcdsp_init_neon.c
@@ -21,8 +21,16 @@ 
 #include "libavutil/attributes.h"
 #include "libavutil/arm/cpu.h"
 #include "libavcodec/hevcdsp.h"
+#include "libavcodec/avcodec.h"
 #include "hevcdsp_arm.h"
 
+void ff_hevc_sao_band_filter_neon_8_wrapper(uint8_t *_dst, uint8_t *_src,
+                                  ptrdiff_t stride_dst, ptrdiff_t stride_src,
+                                  int16_t *sao_offset_val, int sao_left_class,
+                                  int width, int height);
+void ff_hevc_sao_edge_filter_neon_8_wrapper(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val,
+                                  int eo, int width, int height);
+
 void ff_hevc_v_loop_filter_luma_neon(uint8_t *_pix, ptrdiff_t _stride, int _beta, int *_tc, uint8_t *_no_p, uint8_t *_no_q);
 void ff_hevc_h_loop_filter_luma_neon(uint8_t *_pix, ptrdiff_t _stride, int _beta, int *_tc, uint8_t *_no_p, uint8_t *_no_q);
 void ff_hevc_v_loop_filter_chroma_neon(uint8_t *_pix, ptrdiff_t _stride, int *_tc, uint8_t *_no_p, uint8_t *_no_q);
@@ -142,6 +150,51 @@  QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h3v2_neon_8);
 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h3v3_neon_8);
 #undef QPEL_FUNC_UW
 
+void ff_hevc_sao_band_filter_neon_8(uint8_t *dst, uint8_t *src, ptrdiff_t stride_dst, ptrdiff_t stride_src, int width, int height, int16_t *offset_table);
+
+void ff_hevc_sao_band_filter_neon_8_wrapper(uint8_t *_dst, uint8_t *_src,
+                                  ptrdiff_t stride_dst, ptrdiff_t stride_src,
+                                  int16_t *sao_offset_val, int sao_left_class,
+                                  int width, int height) {
+    uint8_t *dst = (uint8_t *)_dst;
+    uint8_t *src = (uint8_t *)_src;
+    int16_t offset_table[32] = {0};
+    int k;
+
+    stride_dst /= sizeof(uint8_t);
+    stride_src /= sizeof(uint8_t);
+
+    for (k = 0; k < 4; k++) {
+        offset_table[(k + sao_left_class) & 31] = sao_offset_val[k + 1];
+    }
+
+    ff_hevc_sao_band_filter_neon_8(dst, src, stride_dst, stride_src, width, height, offset_table);
+}
+
+void ff_hevc_sao_edge_filter_neon_8(uint8_t *dst, uint8_t *src, ptrdiff_t stride_dst, ptrdiff_t stride_src, int width, int height,
+                                    int a_stride, int b_stride, int16_t *sao_offset_val, uint8_t *edge_idx);
+
+void ff_hevc_sao_edge_filter_neon_8_wrapper(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val,
+                                  int eo, int width, int height) {
+    static uint8_t edge_idx[] = { 1, 2, 0, 3, 4 };
+    static const int8_t pos[4][2][2] = {
+        { { -1,  0 }, {  1, 0 } }, // horizontal
+        { {  0, -1 }, {  0, 1 } }, // vertical
+        { { -1, -1 }, {  1, 1 } }, // 45 degree
+        { {  1, -1 }, { -1, 1 } }, // 135 degree
+    };
+    uint8_t *dst = (uint8_t *)_dst;
+    uint8_t *src = (uint8_t *)_src;
+    int a_stride, b_stride;
+    ptrdiff_t stride_src = (2*MAX_PB_SIZE + AV_INPUT_BUFFER_PADDING_SIZE) / sizeof(uint8_t);
+    stride_dst /= sizeof(uint8_t);
+
+    a_stride = pos[eo][0][0] + pos[eo][0][1] * stride_src;
+    b_stride = pos[eo][1][0] + pos[eo][1][1] * stride_src;
+
+    ff_hevc_sao_edge_filter_neon_8(dst, src, stride_dst, stride_src, width, height, a_stride, b_stride, sao_offset_val, edge_idx);
+}
+
 void ff_hevc_put_qpel_neon_wrapper(int16_t *dst, uint8_t *src, ptrdiff_t srcstride,
                                    int height, intptr_t mx, intptr_t my, int width) {
 
@@ -168,6 +221,16 @@  av_cold void ff_hevc_dsp_init_neon(HEVCDSPContext *c, const int bit_depth)
         c->hevc_h_loop_filter_luma     = ff_hevc_h_loop_filter_luma_neon;
         c->hevc_v_loop_filter_chroma   = ff_hevc_v_loop_filter_chroma_neon;
         c->hevc_h_loop_filter_chroma   = ff_hevc_h_loop_filter_chroma_neon;
+        c->sao_band_filter[0]          = ff_hevc_sao_band_filter_neon_8_wrapper;
+        c->sao_band_filter[1]          = ff_hevc_sao_band_filter_neon_8_wrapper;
+        c->sao_band_filter[2]          = ff_hevc_sao_band_filter_neon_8_wrapper;
+        c->sao_band_filter[3]          = ff_hevc_sao_band_filter_neon_8_wrapper;
+        c->sao_band_filter[4]          = ff_hevc_sao_band_filter_neon_8_wrapper;
+        c->sao_edge_filter[0]          = ff_hevc_sao_edge_filter_neon_8_wrapper;
+        c->sao_edge_filter[1]          = ff_hevc_sao_edge_filter_neon_8_wrapper;
+        c->sao_edge_filter[2]          = ff_hevc_sao_edge_filter_neon_8_wrapper;
+        c->sao_edge_filter[3]          = ff_hevc_sao_edge_filter_neon_8_wrapper;
+        c->sao_edge_filter[4]          = ff_hevc_sao_edge_filter_neon_8_wrapper;
         c->add_residual[0]             = ff_hevc_add_residual_4x4_8_neon;
         c->add_residual[1]             = ff_hevc_add_residual_8x8_8_neon;
         c->add_residual[2]             = ff_hevc_add_residual_16x16_8_neon;
diff --git a/libavcodec/arm/hevcdsp_sao_neon.S b/libavcodec/arm/hevcdsp_sao_neon.S
new file mode 100644
index 0000000000..347167951b
--- /dev/null
+++ b/libavcodec/arm/hevcdsp_sao_neon.S
@@ -0,0 +1,181 @@ 
+/*
+ * Copyright (c) 2017 Meng Wang <wangmeng.kids@bytedance.com>
+ *
+ * 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/arm/asm.S"
+#include "neon.S"
+
+function ff_hevc_sao_band_filter_neon_8, export=1
+        push    {r4-r10}
+        ldr     r5,  [sp, #28]   // width
+        ldr     r4,  [sp, #32]   // height
+        ldr     r8,  [sp, #36]   // offset_table
+        vpush   {d8-d15}
+        mov     r12,  r4         // r12 = height
+        mov     r6,   r0         // r6 = r0 = dst
+        mov     r7,   r1         // r7 = r1 = src
+        vldm    r8,   {q0-q3}
+        vmov.u16    q15,  #1
+        vmov.u8     q14,  #32
+0:      pld      [r1]
+        vld1.8   {d16},  [r1], r3
+        cmp      r5,    #4
+        beq      4f
+8:      subs     r4,    #1
+        vshr.u8  d17,   d16,  #3   // index = [src>>3]
+        vshll.u8 q9,    d17,  #1   // lowIndex = 2*index
+        vadd.u16 q11,   q9,   q15  // highIndex = (2*index+1) << 8
+        vshl.u16 q10,   q11,  #8   // q10: highIndex;  q9: lowIndex;
+        vadd.u16 q10,   q9         // combine high and low index;
+        // Look-up Table Round 1; index range: 0-15
+        vtbx.8   d24,   {q0-q1},   d20
+        vtbx.8   d25,   {q0-q1},   d21
+        // Look-up Table Round 2; index range: 16-31
+        vsub.u8  q10,   q14        // Look-up with 8bit
+        vtbx.8   d24,   {q2-q3},   d20
+        vtbx.8   d25,   {q2-q3},   d21
+        vaddw.u8 q13,   q12,       d16
+        vqmovun.s16      d8,         q13
+        vst1.8    d8,   [r0],      r2
+        vld1.8   {d16}, [r1],      r3
+        bne      8b
+        subs     r5,    #8
+        beq      99f
+        mov      r4,    r12
+        add r6, #8
+        mov r0, r6
+        add r7, #8
+        mov r1, r7
+        b        0b
+4:      subs     r4,    #1
+        vshr.u8  d17,   d16,  #3  // src>>3
+        vshll.u8 q9,    d17,  #1   // lowIndex = 2*index
+        vadd.u16 q11,   q9,   q15  // highIndex = (2*index+1) << 8
+        vshl.u16 q10,   q11,  #8   // q10: highIndex;  q9: lowIndex;
+        vadd.u16 q10,   q9         // combine high and low index;
+        // Look-up Table Round 1; index range: 0-15
+        vtbx.8   d24,   {q0-q1},   d20
+        vtbx.8   d25,   {q0-q1},   d21
+        // Look-up Table Round 2; index range: 16-32
+        vsub.u8  q10,   q14        // Look-up with 8bit
+        vtbx.8   d24,   {q2-q3},   d20
+        vtbx.8   d25,   {q2-q3},   d21
+        vaddw.u8 q13,   q12,       d16
+        vqmovun.s16     d14,       q13
+        vst1.32   d14[0],    [r0],     r2
+        vld1.32   {d16[0]},  [r1],     r3
+        bne      4b
+        b        99f
+99:
+        vpop {d8-d15}
+        pop  {r4-r10}
+        bx   lr
+endfunc
+
+function ff_hevc_sao_edge_filter_neon_8, export=1
+        push    {r4-r11}
+        ldr     r5,  [sp, #32]   // width
+        ldr     r4,  [sp, #36]   // height
+        ldr     r8,  [sp, #40]   // a_stride
+        ldr     r9,  [sp, #44]   // b_stride
+        ldr     r10, [sp, #48]   // sao_offset_val
+        ldr     r11, [sp, #52]   // edge_idx
+        vpush   {d8-d15}
+        mov     r12,  r4         // r12 = height
+        mov     r6,   r0         // r6 = r0 = dst
+        mov     r7,   r1         // r7 = r1 = src
+        vld1.8  {d0}, [r11]      // edge_idx tabel load in d0 5x8bit
+        vld1.16 {q1}, [r10]      // sao_offset_val table load in q1, 5x16bit
+        vmov.u8  d1,  #2
+        vmov.u16 q2,  #1
+0:      mov      r10,    r1
+        add      r10,    r8           // src[x + a_stride]
+        mov      r11,    r1
+        add      r11,    r9           // src[x + b_stride]
+        pld      [r1]
+        vld1.8   {d16},  [r1],  r3    // src[x]  8x8bit
+        vld1.8   {d17},  [r10], r3    // src[x + a_stride]
+        vld1.8   {d18},  [r11], r3    // src[x + b_stride]
+        cmp      r5,     #4
+        beq      4f
+8:      subs     r4,     #1
+        vcgt.u8  d8,     d16,   d17
+        vshr.u8  d9,     d8,    #7
+        vclt.u8  d8,     d16,   d17
+        vadd.u8  d8,     d9           // diff0
+        vcgt.u8  d10,    d16,   d18
+        vshr.u8  d11,    d10,   #7
+        vclt.u8  d10,    d16,   d18
+        vadd.u8  d10,    d11          // diff1
+        vadd.s8  d8,     d10
+        vadd.s8  d8,     d1
+        vtbx.8   d9,     {d0},  d8    // offset_val
+        vshll.u8 q6,     d9,    #1    // lowIndex
+        vadd.u16 q7,     q6,    q2
+        vshl.u16 q10,    q7,    #8    // highIndex
+        vadd.u16 q10,    q6           // combine lowIndex and highIndex, offset_val
+        vtbx.8   d22,    {q1},  d20
+        vtbx.8   d23,    {q1},  d21
+        vaddw.u8 q12,    q11,   d16
+        vqmovun.s16      d26,   q12
+        vst1.8   d26,    [r0],  r2
+        vld1.8   {d16},  [r1],  r3    // src[x]  8x8bit
+        vld1.8   {d17},  [r10], r3    // src[x + a_stride]
+        vld1.8   {d18},  [r11], r3    // src[x + b_stride]
+        bne      8b
+        subs     r5,     #8
+        beq      99f
+        mov      r4,     r12
+        add      r6,     #8
+        mov      r0,     r6
+        add      r7,     #8
+        mov      r1,     r7
+        b        0b
+4:      subs     r4,    #1
+        vcgt.u8  d8,     d16,   d17
+        vshr.u8  d9,     d8,    #7
+        vclt.u8  d8,     d16,   d17
+        vadd.u8  d8,     d9           // diff0
+        vcgt.u8  d10,    d16,   d18
+        vshr.u8  d11,    d10,   #7
+        vclt.u8  d10,    d16,   d18
+        vadd.u8  d10,    d11          // diff1
+        vadd.s8  d8,     d10
+        vadd.s8  d8,     d1
+        vtbx.8   d9,     {d0},  d8    // offset_val
+        vshll.u8 q6,     d9,    #1    // lowIndex
+        vadd.u16 q7,     q6,    q2
+        vshl.u16 q10,    q7,    #8    // highIndex
+        vadd.u16 q10,    q6           // combine lowIndex and highIndex, offset_val
+        vtbx.8   d22,    {q1},  d20
+        vtbx.8   d23,    {q1},  d21
+        vaddw.u8 q12,    q11,   d16
+        vqmovun.s16      d26,   q12
+        vst1.32  d26[0], [r0],  r2
+        vld1.32   {d16[0]},  [r1],  r3
+        vld1.32   {d17[0]},  [r10], r3    // src[x + a_stride]
+        vld1.32   {d18[0]},  [r11], r3    // src[x + b_stride]
+        bne      4b
+        b        99f
+99:
+        vpop {d8-d15}
+        pop  {r4-r11}
+        bx   lr
+endfunc