avcodec/arm/hevcdsp_sao : add NEON optimization for sao
authorMeng Wang <wangmeng.kids@bytedance.com>
Tue, 27 Mar 2018 12:43:06 +0000 (20:43 +0800)
committerMichael Niedermayer <michael@niedermayer.cc>
Mon, 9 Apr 2018 01:45:15 +0000 (03:45 +0200)
Signed-off-by: Meng Wang <wangmeng.kids@bytedance.com>
Reviewed-by: Shengbin Meng <shengbinmeng@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
libavcodec/arm/Makefile
libavcodec/arm/hevcdsp_init_neon.c
libavcodec/arm/hevcdsp_sao_neon.S [new file with mode: 0644]

index fd2401f..e656011 100644 (file)
@@ -138,7 +138,8 @@ 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
index a4628d2..201a088 100644 (file)
 #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 +150,47 @@ 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 = _dst;
+    uint8_t *src = _src;
+    int16_t offset_table[32] = {0};
+    int k;
+
+    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 = _dst;
+    uint8_t *src = _src;
+    int a_stride, b_stride;
+    ptrdiff_t stride_src = (2*MAX_PB_SIZE + AV_INPUT_BUFFER_PADDING_SIZE);
+
+    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 +217,16 @@ 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 (file)
index 0000000..3471679
--- /dev/null
@@ -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