[FFmpeg-devel] [PATCH 3/4] x86/vvcdec: inter, add optical flow avx2 code

Nuo Mi nuomi2021 at gmail.com
Sun Aug 18 04:48:06 EEST 2024


BDoF used about 10%–25% of the CPU for some clips.
Here are the FPS for one run; please ignore the negative values, as they may be due to round-to-round variation

clips                                       | before | after | delta
--------------------------------------------|--------|-------|------
RitualDance_1920x1080_60_10_420_37_RA.266   | 310.0  | 363.0 | 14.60%
NovosobornayaSquare_1920x1080.bin           | 322.3  | 339.7 |  5.12%
Tango2_3840x2160_60_10_420_27_LD.266        |  71.0  | 68.7  | -3.35%
RitualDance_1920x1080_60_10_420_32_LD.266   | 250.0  | 245.3 | -1.92%
Chimera_8bit_1080P_1000_frames.vvc          | 359.3  | 422.7 | 15.00%
BQTerrace_1920x1080_60_10_420_22_RA.vvc     | 142.3  | 147.7 |  3.66%
---
 libavcodec/x86/vvc/Makefile      |   1 +
 libavcodec/x86/vvc/vvc_of.asm    | 380 +++++++++++++++++++++++++++++++
 libavcodec/x86/vvc/vvcdsp_init.c |  21 ++
 3 files changed, 402 insertions(+)
 create mode 100644 libavcodec/x86/vvc/vvc_of.asm

diff --git a/libavcodec/x86/vvc/Makefile b/libavcodec/x86/vvc/Makefile
index 04f16bc10c..aa59aa59cf 100644
--- a/libavcodec/x86/vvc/Makefile
+++ b/libavcodec/x86/vvc/Makefile
@@ -6,5 +6,6 @@ OBJS-$(CONFIG_VVC_DECODER)             += x86/vvc/vvcdsp_init.o \
 X86ASM-OBJS-$(CONFIG_VVC_DECODER)      += x86/vvc/vvc_alf.o      \
                                           x86/vvc/vvc_dmvr.o     \
                                           x86/vvc/vvc_mc.o       \
+                                          x86/vvc/vvc_of.o       \
                                           x86/vvc/vvc_sad.o      \
                                           x86/h26x/h2656_inter.o
diff --git a/libavcodec/x86/vvc/vvc_of.asm b/libavcodec/x86/vvc/vvc_of.asm
new file mode 100644
index 0000000000..b70073fa82
--- /dev/null
+++ b/libavcodec/x86/vvc/vvc_of.asm
@@ -0,0 +1,380 @@
+; /*
+; * Provide AVX2 luma optical flow functions for VVC decoding
+; * Copyright (c) 2024 Nuo Mi
+; *
+; * 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"
+
+%define MAX_PB_SIZE             128
+%define SRC_STRIDE              (MAX_PB_SIZE * 2)
+%define SRC_PS                  2                                                       ; source pixel size, sizeof(int16_t)
+%define BDOF_STACK_SIZE         10                                                      ; (4 + 1) * 2, 4 lines + the first line, *2 for h and v
+%define bdof_stack_offset(line) ((line) * 2 % BDOF_STACK_SIZE * mmsize)
+%define SHIFT                   6
+%define SHIFT2                  4
+
+SECTION_RODATA 32
+pd_15                   times 8 dd 15
+pd_m15                  times 8 dd -15
+cextern pb_0
+
+pb_shuffle_w8  times 2 db   0, 1, 0xff, 0xff, 8, 9, 0xff, 0xff, 6, 7, 0xff, 0xff, 14, 15, 0xff, 0xff
+pb_shuffle_w16 times 2 db   0, 1, 0xff, 0xff, 6, 7, 0xff, 0xff, 8, 9, 0xff, 0xff, 14, 15, 0xff, 0xff
+pd_perm_w16            dd   0, 2, 1, 4, 3, 6, 5, 7
+%if ARCH_X86_64
+
+%if HAVE_AVX2_EXTERNAL
+
+SECTION .text
+
+INIT_YMM avx2
+
+; dst = (src0 >> shift) - (src1 >> shift)
+%macro DIFF 5 ; dst, src0, src1, shift, tmp
+    psraw   %1, %2, %4
+    psraw   %5, %3, %4
+    psubw   %1, %5
+%endmacro
+
+%macro LOAD_GRAD_H 4 ; dst, src, off, tmp
+    movu    %1, [%2 + %3 + 2 * SRC_PS]
+    movu    %4, [%2 + %3]
+
+    DIFF    %1, %1, %4, SHIFT, %4
+%endmacro
+
+%macro SUM_GRAD 2 ;(dst/grad0, grad1)
+    paddw  %1, %2
+    psraw  %1, 1    ; shift3
+%endmacro
+
+%macro APPLY_BDOF_MIN_BLOCK_LINE 5 ; dst, vx, vy, tmp, line_num
+%define off bdof_stack_offset(%5)
+    pmullw                        %1, %2, [rsp + off + 0 * mmsize]               ; vx * (gradient_h[0] - gradient_h[1])
+    pmullw                        %4, %3, [rsp + off + 1 * mmsize]               ; vy * (gradient_v[0] - gradient_v[1])
+    paddw                         %1, [src0q + (%5 + 1) * SRC_STRIDE + SRC_PS]
+    paddw                         %4, [src1q + (%5 + 1) * SRC_STRIDE + SRC_PS]
+    paddsw                        %1, %4                                         ; src0[x] + src1[x] + bdof_offset
+    pmulhrsw                      %1, m11
+    CLIPW                         %1, m9, m10
+%endmacro
+
+%macro SAVE_8BPC 2 ; dst, src
+    packuswb                   m%2, m%2
+    vpermq                     m%2, m%2, q0020
+
+    cmp                         wd, 16
+    je                       %%w16
+    movq                        %1, xm%2
+    jmp                     %%wend
+%%w16:
+    movu                        %1, xm%2
+%%wend:
+%endmacro
+
+%macro SAVE_16BPC 2 ; dst, src
+    cmp                         wd, 16
+    je                       %%w16
+    movu                        %1, xm%2
+    jmp                     %%wend
+%%w16:
+    movu                        %1, m%2
+%%wend:
+%endmacro
+
+%macro SAVE 2 ; dst, src
+    cmp                 pixel_maxd, (1 << 8) - 1
+    jne               %%save_16bpc
+    SAVE_8BPC                   %1, %2
+    jmp                      %%end
+%%save_16bpc:
+    SAVE_16BPC                   %1, %2
+%%end:
+%endmacro
+
+; [rsp + even * mmsize] are gradient_h[0] - gradient_h[1]
+; [rsp +  odd * mmsize] are gradient_v[0] - gradient_v[1]
+%macro APPLY_BDOF_MIN_BLOCK 4 ; block_num, vx, vy, bd
+    vpbroadcastd                  m9, [pb_0]
+
+    movd                        xm10, pixel_maxd
+    vpbroadcastw                 m10, xm10
+
+    lea                        tmp0d, [pixel_maxd + 1]
+    movd                        xm11, tmp0d
+    VPBROADCASTW                 m11, xm11                 ;shift_4 for pmulhrsw
+
+    APPLY_BDOF_MIN_BLOCK_LINE    m6, %2, %3, m7, (%1) * 4 + 0
+    SAVE                        [dstq + 0 * dsq], 6
+
+    APPLY_BDOF_MIN_BLOCK_LINE    m6, %2, %3, m7, (%1) * 4 + 1
+    SAVE                        [dstq + 1 * dsq], 6
+
+    APPLY_BDOF_MIN_BLOCK_LINE    m6, %2, %3, m7, (%1) * 4 + 2
+    SAVE                        [dstq + 2 * dsq], 6
+
+    APPLY_BDOF_MIN_BLOCK_LINE    m6, %2, %3, m7, (%1) * 4 + 3
+    SAVE                        [dstq + ds3q], 6
+%endmacro
+
+%macro SUM_MIN_BLOCK_W16 4 ; src/dst, shuffle, perm, tmp
+    pshufb  %4, %1, %2
+    vpermd  %4, %3, %4
+    paddw   %1, %4
+%endmacro
+
+%macro SUM_MIN_BLOCK_W8 3 ; src/dst, shuffle, tmp
+    pshufb  %3, %1, %2
+    paddw   %1, %3
+%endmacro
+
+%macro BDOF_PROF_GRAD 2 ; line_no, last_line
+%assign i0 (%1 + 0) % 3
+%assign j0 (%1 + 1) % 3
+%assign k0 (%1 + 2) % 3
+%assign i1 3 + (%1 + 0) % 3
+%assign j1 3 + (%1 + 1) % 3
+%assign k1 3 + (%1 + 2) % 3
+
+; we cached src0 in m0 to m2
+%define t0 m %+ i0
+%define c0 m %+ j0
+%define b0 m %+ k0
+
+; we cached src1 in m3 to m5
+%define t1 m %+ i1
+%define c1 m %+ j1
+%define b1 m %+ k1
+%define ndiff t1
+%define off bdof_stack_offset(%1)
+
+    movu                        b0, [src0q + (%1 + 2) * SRC_STRIDE + SRC_PS]
+    movu                        b1, [src1q + (%1 + 2) * SRC_STRIDE + SRC_PS]
+
+    ; gradient_v[0], gradient_v[1]
+    DIFF                        m6,  b0,  t0, SHIFT, t0
+    DIFF                        m7,  b1,  t1, SHIFT, t1
+
+    ; save gradient_v[0] - gradient_v[1]
+    psubw                      m10, m6, m7
+    mova      [rsp + off + mmsize], m10
+
+    ; gradient_h[0], gradient_h[1]
+    LOAD_GRAD_H                 m8, src0q, (%1 + 1) * SRC_STRIDE, t0
+    LOAD_GRAD_H                 m9, src1q, (%1 + 1) * SRC_STRIDE, t1
+
+    ; save gradient_h[0] - gradient_h[1]
+    psubw                      m11, m8, m9
+    mova               [rsp + off], m11
+
+    SUM_GRAD                    m8, m9                  ; temph
+    SUM_GRAD                    m6, m7                  ; tempv
+
+    DIFF                     ndiff, c1, c0, SHIFT2, t0  ; -diff
+
+    psignw                      m7, ndiff, m8           ; sgxdi
+    psignw                      m9, ndiff, m6           ; sgydi
+    psignw                     m10, m8, m6              ; sgxgy
+
+    pabsw                       m6, m6                  ; sgy2
+    pabsw                       m8, m8                  ; sgx2
+
+    ; use t0, t1 as temporary buffers
+    cmp                         wd, 16
+
+    je                       %%w16
+    mova                        t0, [pb_shuffle_w8]
+    SUM_MIN_BLOCK_W8            m6, t0, m11
+    SUM_MIN_BLOCK_W8            m7, t0, m11
+    SUM_MIN_BLOCK_W8            m8, t0, m11
+    SUM_MIN_BLOCK_W8            m9, t0, m11
+    SUM_MIN_BLOCK_W8           m10, t0, m11
+    jmp                     %%wend
+
+%%w16:
+    mova                        t0, [pb_shuffle_w16]
+    mova                        t1, [pd_perm_w16]
+    SUM_MIN_BLOCK_W16           m6, t0, t1, m11
+    SUM_MIN_BLOCK_W16           m7, t0, t1, m11
+    SUM_MIN_BLOCK_W16           m8, t0, t1, m11
+    SUM_MIN_BLOCK_W16           m9, t0, t1, m11
+    SUM_MIN_BLOCK_W16          m10, t0, t1, m11
+
+%%wend:
+    phaddw                      m8, m7
+    phaddw                      m6, m9
+    phaddw                      m8, m6                   ; 2 sgx2, 2 sgxdi, sgy2, 2 sgydi, 2 sgx2, 2 sgxdi, 2 sgy2, 2 sgydi
+
+%if (%1) == 0 || (%2)
+    ; pad for top and bottom
+    paddw                       m8, m8
+    paddw                      m10, m10
+%endif
+
+    paddw                      m12, m8
+    paddw                      m13, m10
+%endmacro
+
+
+%macro LOG2 5; log_sum, src, cmp, shift, tmp
+    pcmpgtw               %5, %2, %3
+    pandd                 %5, %4
+    paddw                 %1, %5
+
+    psrlw                 %2, %5
+    psrlw                 %4, 1
+    psrlw                 %3, %4
+%endmacro
+
+%macro LOG2 3; dst0/src0, dst1/src, offset
+    pextrd              tmp0d, xm%1, %3
+    pextrd              tmp1d, xm%2, %3
+    bsr                 tmp0d, tmp0d
+    bsr                 tmp1d, tmp1d
+    pinsrd              xm%1, tmp0d, %3
+    pinsrd              xm%2, tmp1d, %3
+%endmacro
+
+%macro LOG2 2; dst/src, tmp
+    vperm2i128           m%2, m%1, m%1, 1
+    LOG2                 %1,  %2, 0
+    LOG2                 %1,  %2, 1
+    LOG2                 %1,  %2, 2
+    LOG2                 %1,  %2, 3
+    vperm2i128           m%1, m%1, m%2, q0200
+%endmacro
+
+; %1: 2 sgx2, 2 sgxdi, sgy2, 2 sgydi, 2 sgx2, 2 sgxdi, 2 sgy2, 2 sgydi
+; %2: 4 4sgxgy
+%macro BDOF_VX_VY 2       ;
+    pmovsxwd                m6, xm%1
+    vextracti128           xm7, m%1, 1
+    pmovsxwd                m7, xm7
+    punpcklqdq              m8, m6, m7               ; 4 sgx2,  4 sgy2
+    punpckhqdq              m9, m6, m7               ; 4 sgxdi, 4 sgydi
+    mova                   m10, m8
+
+    LOG2                    10, 11
+    psignd                 m11, m9, m8
+    pslld                  m11, 2
+    vpsravd                m11, m11, m10
+    CLIPD                  m11, [pd_m15], [pd_15]   ; 4 vx
+
+    pxor                    m6, m6
+    phaddw                 m%2, m6
+    phaddw                 m%2, m6
+    vpermq                 m%2, m%2, q0020
+    pshufd                 m%2, m%2, q1120
+    pmovsxwd               m%2, xmm%2               ; 4 sgxgy
+
+    pmulld                 m%2, m11                 ; 4 vx * sgxgy
+    psrad                  m%2, 1
+
+    pslld                   m9, 2
+    vextracti128          xmm9, m9, 1               ; 4 (sgydi << 2)
+    psubd                   m9, m%2                 ; 4 ((sgydi << 2) - (vx * sgxgy >> 1))
+
+    vextracti128          xmm8, m8, 1               ; 4 sgy2
+    psignd                  m9, m8
+
+    vextracti128         xmm10, m10, 1              ; 4 log2(sgy2)
+    vpsravd                 m9, m9, m10
+    CLIPD                   m9, [pd_m15], [pd_15]   ; 4 vy
+
+    vpermq                 m%1, m11, q1100
+    pshuflw                m%1, m%1, q2200
+    vpunpckldq             m%1, m%1                 ; 4 x 4vx
+
+    vpermq                 m%2, m9, q1100
+    pshuflw                m%2, m%2, q2200
+    vpunpckldq             m%2, m%2                 ; 4 x 4vy
+%endmacro
+
+
+%macro BDOF_MINI_BLOCKS 2 ; (block_num, last_block)
+
+%if (%1) == 0
+    movu                    m0, [src0q + 0 * SRC_STRIDE + SRC_PS]
+    movu                    m1, [src0q + 1 * SRC_STRIDE + SRC_PS]
+    movu                    m3, [src1q + 0 * SRC_STRIDE + SRC_PS]
+    movu                    m4, [src1q + 1 * SRC_STRIDE + SRC_PS]
+
+    pxor                   m12, m12
+    pxor                   m13, m13
+
+    BDOF_PROF_GRAD           0, 0
+%endif
+
+    mova                   m14, m12
+    mova                   m15, m13
+
+    pxor                   m12, m12
+    pxor                   m13, m13
+    BDOF_PROF_GRAD  %1 * 4 + 1, 0
+    BDOF_PROF_GRAD  %1 * 4 + 2, 0
+    paddw                  m14, m12
+    paddw                  m15, m13
+
+    pxor                   m12, m12
+    pxor                   m13, m13
+    BDOF_PROF_GRAD  %1 * 4 + 3, %2
+%if (%2) == 0
+    BDOF_PROF_GRAD  %1 * 4 + 4, 0
+%endif
+    paddw                  m14, m12
+    paddw                  m15, m13
+
+    BDOF_VX_VY              14, 15
+    APPLY_BDOF_MIN_BLOCK    %1, m14, m15, bd
+    lea                   dstq, [dstq + 4 * dsq]
+%endmacro
+
+;void ff_vvc_apply_bdof_%1(uint8_t *dst, const ptrdiff_t dst_stride, int16_t *src0, int16_t *src1,
+;    const int w, const int h, const int int pixel_max)
+%macro BDOF_AVX2 0
+cglobal vvc_apply_bdof, 7, 10, 16, BDOF_STACK_SIZE*32, dst, ds, src0, src1, w, h, pixel_max, ds3, tmp0, tmp1
+
+    lea                   ds3q, [dsq * 3]
+    sub                  src0q, SRC_STRIDE + SRC_PS
+    sub                  src1q, SRC_STRIDE + SRC_PS
+
+    BDOF_MINI_BLOCKS         0, 0
+
+    cmp                     hd, 16
+    je                    .h16
+    BDOF_MINI_BLOCKS         1, 1
+    jmp                   .end
+
+.h16:
+    BDOF_MINI_BLOCKS         1, 0
+    BDOF_MINI_BLOCKS         2, 0
+    BDOF_MINI_BLOCKS         3, 1
+
+.end:
+    RET
+%endmacro
+
+%macro VVC_OF_AVX2 0
+    BDOF_AVX2
+%endmacro
+
+VVC_OF_AVX2
+
+%endif ; HAVE_AVX2_EXTERNAL
+
+%endif ; ARCH_X86_64
diff --git a/libavcodec/x86/vvc/vvcdsp_init.c b/libavcodec/x86/vvc/vvcdsp_init.c
index d5b4f4f8a5..f3e2e3a27b 100644
--- a/libavcodec/x86/vvc/vvcdsp_init.c
+++ b/libavcodec/x86/vvc/vvcdsp_init.c
@@ -102,6 +102,20 @@ DMVR_PROTOTYPES( 8, avx2)
 DMVR_PROTOTYPES(10, avx2)
 DMVR_PROTOTYPES(12, avx2)
 
+void ff_vvc_apply_bdof_avx2(uint8_t *dst, ptrdiff_t dst_stride,                                     \
+    const int16_t *src0, const int16_t *src1, int w, int h, int pixel_max);                         \
+
+#define OF_PROTOTYPES(bd, opt)                                                                      \
+static void ff_vvc_apply_bdof_##bd##_##opt(uint8_t *dst, ptrdiff_t dst_stride,                      \
+    const int16_t *src0, const int16_t *src1, int w, int h)                                         \
+{                                                                                                   \
+    ff_vvc_apply_bdof##_##opt(dst, dst_stride, src0, src1, w, h, (1 << bd)  - 1);                   \
+}                                                                                                   \
+
+OF_PROTOTYPES( 8, avx2)
+OF_PROTOTYPES(10, avx2)
+OF_PROTOTYPES(12, avx2)
+
 #define ALF_BPC_PROTOTYPES(bpc, opt)                                                                                     \
 void BF(ff_vvc_alf_filter_luma, bpc, opt)(uint8_t *dst, ptrdiff_t dst_stride,                                            \
     const uint8_t *src, ptrdiff_t src_stride, ptrdiff_t width, ptrdiff_t height,                                         \
@@ -328,6 +342,10 @@ ALF_FUNCS(16, 12, avx2)
     c->inter.dmvr[1][1]   = ff_vvc_dmvr_hv_##bd##_avx2;              \
 } while (0)
 
+#define OF_INIT(bd) do {                                             \
+    c->inter.apply_bdof   = ff_vvc_apply_bdof_##bd##_avx2;           \
+} while (0)
+
 #define ALF_INIT(bd) do {                                            \
     c->alf.filter[LUMA]   = ff_vvc_alf_filter_luma_##bd##_avx2;      \
     c->alf.filter[CHROMA] = ff_vvc_alf_filter_chroma_##bd##_avx2;    \
@@ -352,6 +370,7 @@ void ff_vvc_dsp_init_x86(VVCDSPContext *const c, const int bd)
             ALF_INIT(8);
             AVG_INIT(8, avx2);
             MC_LINKS_AVX2(8);
+            OF_INIT(8);
             DMVR_INIT(8);
             SAD_INIT();
         }
@@ -365,6 +384,7 @@ void ff_vvc_dsp_init_x86(VVCDSPContext *const c, const int bd)
             AVG_INIT(10, avx2);
             MC_LINKS_AVX2(10);
             MC_LINKS_16BPC_AVX2(10);
+            OF_INIT(10);
             DMVR_INIT(10);
             SAD_INIT();
         }
@@ -378,6 +398,7 @@ void ff_vvc_dsp_init_x86(VVCDSPContext *const c, const int bd)
             AVG_INIT(12, avx2);
             MC_LINKS_AVX2(12);
             MC_LINKS_16BPC_AVX2(12);
+            OF_INIT(12);
             DMVR_INIT(12);
             SAD_INIT();
         }
-- 
2.34.1



More information about the ffmpeg-devel mailing list