[FFmpeg-devel] [PATCH v1 3/5] avcodec: [loongarch] Optimize Hevc_idct/lpf with LSX.

Hao Chen chenhao at loongson.cn
Mon Feb 14 10:25:34 EET 2022


From: Lu Wang <wanglu at loongson.cn>

ffmpeg -i 5_h265_1080p_60fps_3Mbps.mkv -f rawvideo -y /dev/null -an
before:  110fps
after : 124fps
---
 libavcodec/loongarch/Makefile                 |    2 +
 libavcodec/loongarch/hevc_idct_lsx.c          |  842 ++++++
 libavcodec/loongarch/hevc_lpf_sao_lsx.c       | 2485 +++++++++++++++++
 libavcodec/loongarch/hevcdsp_init_loongarch.c |   19 +
 libavcodec/loongarch/hevcdsp_lsx.h            |   26 +
 5 files changed, 3374 insertions(+)
 create mode 100644 libavcodec/loongarch/hevc_idct_lsx.c
 create mode 100644 libavcodec/loongarch/hevc_lpf_sao_lsx.c

diff --git a/libavcodec/loongarch/Makefile b/libavcodec/loongarch/Makefile
index 9f416d0c6e..cfc8e3aaff 100644
--- a/libavcodec/loongarch/Makefile
+++ b/libavcodec/loongarch/Makefile
@@ -26,3 +26,5 @@ LSX-OBJS-$(CONFIG_VP9_DECODER)        += loongarch/vp9_mc_lsx.o \
                                          loongarch/vp9_lpf_lsx.o \
                                          loongarch/vp9_idct_lsx.o
 LSX-OBJS-$(CONFIG_HEVC_DECODER)       += loongarch/hevcdsp_lsx.o \
+                                         loongarch/hevc_idct_lsx.o \
+                                         loongarch/hevc_lpf_sao_lsx.o
diff --git a/libavcodec/loongarch/hevc_idct_lsx.c b/libavcodec/loongarch/hevc_idct_lsx.c
new file mode 100644
index 0000000000..2193b27546
--- /dev/null
+++ b/libavcodec/loongarch/hevc_idct_lsx.c
@@ -0,0 +1,842 @@
+/*
+ * Copyright (c) 2022 Loongson Technology Corporation Limited
+ * Contributed by Shiyou Yin <yinshiyou-hf at loongson.cn>
+ *                Hao Chen <chenhao at loongson.cn>
+ *
+ * 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/loongarch/loongson_intrinsics.h"
+#include "hevcdsp_lsx.h"
+
+static const int16_t gt8x8_cnst[16] __attribute__ ((aligned (64))) = {
+    64, 64, 83, 36, 89, 50, 18, 75, 64, -64, 36, -83, 75, -89, -50, -18
+};
+
+static const int16_t gt16x16_cnst[64] __attribute__ ((aligned (64))) = {
+    64, 83, 64, 36, 89, 75, 50, 18, 90, 80, 57, 25, 70, 87, 9, 43,
+    64, 36, -64, -83, 75, -18, -89, -50, 87, 9, -80, -70, -43, 57, -25, -90,
+    64, -36, -64, 83, 50, -89, 18, 75, 80, -70, -25, 90, -87, 9, 43, 57,
+    64, -83, 64, -36, 18, -50, 75, -89, 70, -87, 90, -80, 9, -43, -57, 25
+};
+
+static const int16_t gt32x32_cnst0[256] __attribute__ ((aligned (64))) = {
+    90, 90, 88, 85, 82, 78, 73, 67, 61, 54, 46, 38, 31, 22, 13, 4,
+    90, 82, 67, 46, 22, -4, -31, -54, -73, -85, -90, -88, -78, -61, -38, -13,
+    88, 67, 31, -13, -54, -82, -90, -78, -46, -4, 38, 73, 90, 85, 61, 22,
+    85, 46, -13, -67, -90, -73, -22, 38, 82, 88, 54, -4, -61, -90, -78, -31,
+    82, 22, -54, -90, -61, 13, 78, 85, 31, -46, -90, -67, 4, 73, 88, 38,
+    78, -4, -82, -73, 13, 85, 67, -22, -88, -61, 31, 90, 54, -38, -90, -46,
+    73, -31, -90, -22, 78, 67, -38, -90, -13, 82, 61, -46, -88, -4, 85, 54,
+    67, -54, -78, 38, 85, -22, -90, 4, 90, 13, -88, -31, 82, 46, -73, -61,
+    61, -73, -46, 82, 31, -88, -13, 90, -4, -90, 22, 85, -38, -78, 54, 67,
+    54, -85, -4, 88, -46, -61, 82, 13, -90, 38, 67, -78, -22, 90, -31, -73,
+    46, -90, 38, 54, -90, 31, 61, -88, 22, 67, -85, 13, 73, -82, 4, 78,
+    38, -88, 73, -4, -67, 90, -46, -31, 85, -78, 13, 61, -90, 54, 22, -82,
+    31, -78, 90, -61, 4, 54, -88, 82, -38, -22, 73, -90, 67, -13, -46, 85,
+    22, -61, 85, -90, 73, -38, -4, 46, -78, 90, -82, 54, -13, -31, 67, -88,
+    13, -38, 61, -78, 88, -90, 85, -73, 54, -31, 4, 22, -46, 67, -82, 90,
+    4, -13, 22, -31, 38, -46, 54, -61, 67, -73, 78, -82, 85, -88, 90, -90
+};
+
+static const int16_t gt32x32_cnst1[64] __attribute__ ((aligned (64))) = {
+    90, 87, 80, 70, 57, 43, 25, 9, 87, 57, 9, -43, -80, -90, -70, -25,
+    80, 9, -70, -87, -25, 57, 90, 43, 70, -43, -87, 9, 90, 25, -80, -57,
+    57, -80, -25, 90, -9, -87, 43, 70, 43, -90, 57, 25, -87, 70, 9, -80,
+    25, -70, 90, -80, 43, 9, -57, 87, 9, -25, 43, -57, 70, -80, 87, -90
+};
+
+static const int16_t gt32x32_cnst2[16] __attribute__ ((aligned (64))) = {
+    89, 75, 50, 18, 75, -18, -89, -50, 50, -89, 18, 75, 18, -50, 75, -89
+};
+
+#define HEVC_IDCT4x4_COL(in_r0, in_l0, in_r1, in_l1,          \
+                         sum0, sum1, sum2, sum3, shift)       \
+{                                                             \
+    __m128i vec0, vec1, vec2, vec3, vec4, vec5;               \
+    __m128i cnst64 = __lsx_vldi(0x0840);                      \
+    __m128i cnst83 = __lsx_vldi(0x0853);                      \
+    __m128i cnst36 = __lsx_vldi(0x0824);                      \
+                                                              \
+    vec0 = __lsx_vdp2_w_h(in_r0, cnst64);                     \
+    vec1 = __lsx_vdp2_w_h(in_l0, cnst83);                     \
+    vec2 = __lsx_vdp2_w_h(in_r1, cnst64);                     \
+    vec3 = __lsx_vdp2_w_h(in_l1, cnst36);                     \
+    vec4 = __lsx_vdp2_w_h(in_l0, cnst36);                     \
+    vec5 = __lsx_vdp2_w_h(in_l1, cnst83);                     \
+                                                              \
+    sum0 = __lsx_vadd_w(vec0, vec2);                          \
+    sum1 = __lsx_vsub_w(vec0, vec2);                          \
+    vec1 = __lsx_vadd_w(vec1, vec3);                          \
+    vec4 = __lsx_vsub_w(vec4, vec5);                          \
+    sum2 = __lsx_vsub_w(sum1, vec4);                          \
+    sum3 = __lsx_vsub_w(sum0, vec1);                          \
+    sum0 = __lsx_vadd_w(sum0, vec1);                          \
+    sum1 = __lsx_vadd_w(sum1, vec4);                          \
+                                                              \
+    sum0 = __lsx_vsrari_w(sum0, shift);                       \
+    sum1 = __lsx_vsrari_w(sum1, shift);                       \
+    sum2 = __lsx_vsrari_w(sum2, shift);                       \
+    sum3 = __lsx_vsrari_w(sum3, shift);                       \
+    sum0 = __lsx_vsat_w(sum0, 15);                            \
+    sum1 = __lsx_vsat_w(sum1, 15);                            \
+    sum2 = __lsx_vsat_w(sum2, 15);                            \
+    sum3 = __lsx_vsat_w(sum3, 15);                            \
+}
+
+#define HEVC_IDCT8x8_COL(in0, in1, in2, in3, in4, in5, in6, in7, shift)  \
+{                                                                        \
+    __m128i src0_r, src1_r, src2_r, src3_r;                              \
+    __m128i src0_l, src1_l, src2_l, src3_l;                              \
+    __m128i filter0, filter1, filter2, filter3;                          \
+    __m128i temp0_r, temp1_r, temp2_r, temp3_r, temp4_r, temp5_r;        \
+    __m128i temp0_l, temp1_l, temp2_l, temp3_l, temp4_l, temp5_l;        \
+    __m128i sum0_r, sum1_r, sum2_r, sum3_r;                              \
+    __m128i sum0_l, sum1_l, sum2_l, sum3_l;                              \
+                                                                         \
+    DUP4_ARG2(__lsx_vilvl_h, in4, in0, in6, in2, in5, in1, in3, in7,     \
+              src0_r, src1_r, src2_r, src3_r);                           \
+    DUP4_ARG2(__lsx_vilvh_h, in4, in0, in6, in2, in5, in1, in3, in7,     \
+              src0_l, src1_l, src2_l, src3_l);                           \
+                                                                         \
+    DUP4_ARG2(__lsx_vldrepl_w, filter, 0, filter, 4, filter, 8,          \
+              filter, 12, filter0, filter1, filter2, filter3);           \
+    DUP4_ARG2(__lsx_vdp2_w_h, src0_r, filter0, src0_l, filter0,          \
+              src1_r, filter1, src1_l, filter1,  temp0_r, temp0_l,       \
+              temp1_r, temp1_l);                                         \
+                                                                         \
+    LSX_BUTTERFLY_4_W(temp0_r, temp0_l, temp1_l, temp1_r, sum0_r, sum0_l,\
+                      sum1_l, sum1_r);                                   \
+    sum2_r = sum1_r;                                                     \
+    sum2_l = sum1_l;                                                     \
+    sum3_r = sum0_r;                                                     \
+    sum3_l = sum0_l;                                                     \
+                                                                         \
+    DUP4_ARG2(__lsx_vdp2_w_h, src2_r, filter2, src2_l, filter2,          \
+              src3_r, filter3, src3_l, filter3,  temp2_r, temp2_l,       \
+              temp3_r, temp3_l);                                         \
+    temp2_r = __lsx_vadd_w(temp2_r, temp3_r);                            \
+    temp2_l = __lsx_vadd_w(temp2_l, temp3_l);                            \
+    sum0_r  = __lsx_vadd_w(sum0_r, temp2_r);                             \
+    sum0_l  = __lsx_vadd_w(sum0_l, temp2_l);                             \
+    sum3_r  = __lsx_vsub_w(sum3_r, temp2_r);                             \
+    sum3_l  = __lsx_vsub_w(sum3_l, temp2_l);                             \
+                                                                         \
+    in0 = __lsx_vssrarni_h_w(sum0_l, sum0_r, shift);                     \
+    in7 = __lsx_vssrarni_h_w(sum3_l, sum3_r, shift);                     \
+                                                                         \
+    DUP4_ARG2(__lsx_vdp2_w_h, src2_r, filter3, src2_l, filter3,          \
+              src3_r, filter2, src3_l, filter2,  temp4_r, temp4_l,       \
+              temp5_r, temp5_l);                                         \
+    temp4_r = __lsx_vsub_w(temp4_r, temp5_r);                            \
+    temp4_l = __lsx_vsub_w(temp4_l, temp5_l);                            \
+    sum1_r  = __lsx_vadd_w(sum1_r, temp4_r);                             \
+    sum1_l  = __lsx_vadd_w(sum1_l, temp4_l);                             \
+    sum2_r  = __lsx_vsub_w(sum2_r, temp4_r);                             \
+    sum2_l  = __lsx_vsub_w(sum2_l, temp4_l);                             \
+                                                                         \
+    in3 = __lsx_vssrarni_h_w(sum1_l, sum1_r, shift);                     \
+    in4 = __lsx_vssrarni_h_w(sum2_l, sum2_r, shift);                     \
+                                                                         \
+    DUP4_ARG2(__lsx_vldrepl_w, filter, 16, filter, 20, filter, 24,       \
+              filter, 28, filter0, filter1, filter2, filter3);           \
+    DUP4_ARG2(__lsx_vdp2_w_h, src0_r, filter0, src0_l, filter0,          \
+              src1_r, filter1, src1_l, filter1,  temp0_r, temp0_l,       \
+              temp1_r, temp1_l);                                         \
+                                                                         \
+    LSX_BUTTERFLY_4_W(temp0_r, temp0_l, temp1_l, temp1_r, sum0_r, sum0_l,\
+                      sum1_l, sum1_r);                                   \
+    sum2_r = sum1_r;                                                     \
+    sum2_l = sum1_l;                                                     \
+    sum3_r = sum0_r;                                                     \
+    sum3_l = sum0_l;                                                     \
+                                                                         \
+    DUP4_ARG2(__lsx_vdp2_w_h, src2_r, filter2, src2_l, filter2,          \
+              src3_r, filter3, src3_l, filter3,  temp2_r, temp2_l,       \
+              temp3_r, temp3_l);                                         \
+    temp2_r = __lsx_vadd_w(temp2_r, temp3_r);                            \
+    temp2_l = __lsx_vadd_w(temp2_l, temp3_l);                            \
+    sum0_r  = __lsx_vadd_w(sum0_r, temp2_r);                             \
+    sum0_l  = __lsx_vadd_w(sum0_l, temp2_l);                             \
+    sum3_r  = __lsx_vsub_w(sum3_r, temp2_r);                             \
+    sum3_l  = __lsx_vsub_w(sum3_l, temp2_l);                             \
+                                                                         \
+    in1 = __lsx_vssrarni_h_w(sum0_l, sum0_r, shift);                     \
+    in6 = __lsx_vssrarni_h_w(sum3_l, sum3_r, shift);                     \
+                                                                         \
+    DUP4_ARG2(__lsx_vdp2_w_h, src2_r, filter3, src2_l, filter3,          \
+              src3_r, filter2, src3_l, filter2,  temp4_r, temp4_l,       \
+              temp5_r, temp5_l);                                         \
+    temp4_r = __lsx_vsub_w(temp4_r, temp5_r);                            \
+    temp4_l = __lsx_vsub_w(temp4_l, temp5_l);                            \
+    sum1_r  = __lsx_vsub_w(sum1_r, temp4_r);                             \
+    sum1_l  = __lsx_vsub_w(sum1_l, temp4_l);                             \
+    sum2_r  = __lsx_vadd_w(sum2_r, temp4_r);                             \
+    sum2_l  = __lsx_vadd_w(sum2_l, temp4_l);                             \
+                                                                         \
+    in2 = __lsx_vssrarni_h_w(sum1_l, sum1_r, shift);                     \
+    in5 = __lsx_vssrarni_h_w(sum2_l, sum2_r, shift);                     \
+}
+
+#define HEVC_IDCT16x16_COL(src0_r, src1_r, src2_r, src3_r,                   \
+                           src4_r, src5_r, src6_r, src7_r,                   \
+                           src0_l, src1_l, src2_l, src3_l,                   \
+                           src4_l, src5_l, src6_l, src7_l, shift)            \
+{                                                                            \
+    int16_t *ptr0, *ptr1;                                                    \
+    __m128i dst0, dst1;                                                      \
+    __m128i filter0, filter1, filter2, filter3;                              \
+    __m128i temp0_r, temp1_r, temp0_l, temp1_l;                              \
+    __m128i sum0_r, sum1_r, sum2_r, sum3_r, sum0_l, sum1_l, sum2_l;          \
+    __m128i sum3_l, res0_r, res1_r, res0_l, res1_l;                          \
+                                                                             \
+    ptr0 = (buf_ptr + 112);                                                  \
+    ptr1 = (buf_ptr + 128);                                                  \
+    k = -1;                                                                  \
+                                                                             \
+    for (j = 0; j < 4; j++)                                                  \
+    {                                                                        \
+        DUP4_ARG2(__lsx_vldrepl_w, filter, 0, filter, 4, filter, 16,         \
+                  filter, 20, filter0, filter1, filter2, filter3);           \
+        DUP4_ARG2(__lsx_vdp2_w_h, src0_r, filter0, src0_l, filter0,          \
+                  src4_r, filter2, src4_l, filter2,  sum0_r, sum0_l,         \
+                  sum2_r, sum2_l);                                           \
+        DUP2_ARG2(__lsx_vdp2_w_h, src7_r, filter2, src7_l, filter2,          \
+                  sum3_r, sum3_l);                                           \
+        DUP4_ARG3(__lsx_vdp2add_w_h, sum0_r, src1_r, filter1, sum0_l,        \
+                  src1_l, filter1, sum2_r, src5_r, filter3, sum2_l,          \
+                  src5_l, filter3, sum0_r, sum0_l, sum2_r, sum2_l);          \
+        DUP2_ARG3(__lsx_vdp2add_w_h, sum3_r, src6_r, filter3, sum3_l,        \
+                  src6_l, filter3, sum3_r, sum3_l);                          \
+                                                                             \
+        sum1_r = sum0_r;                                                     \
+        sum1_l = sum0_l;                                                     \
+                                                                             \
+        DUP4_ARG2(__lsx_vldrepl_w, filter, 8, filter, 12, filter, 24,        \
+                  filter, 28, filter0, filter1, filter2, filter3);           \
+        filter += 16;                                                        \
+        DUP2_ARG2(__lsx_vdp2_w_h, src2_r, filter0, src2_l, filter0,          \
+                  temp0_r, temp0_l);                                         \
+        DUP2_ARG3(__lsx_vdp2add_w_h, sum2_r, src6_r, filter2, sum2_l,        \
+                  src6_l, filter2, sum2_r, sum2_l);                          \
+        DUP2_ARG2(__lsx_vdp2_w_h, src5_r, filter2, src5_l, filter2,          \
+                  temp1_r, temp1_l);                                         \
+                                                                             \
+        sum0_r = __lsx_vadd_w(sum0_r, temp0_r);                              \
+        sum0_l = __lsx_vadd_w(sum0_l, temp0_l);                              \
+        sum1_r = __lsx_vsub_w(sum1_r, temp0_r);                              \
+        sum1_l = __lsx_vsub_w(sum1_l, temp0_l);                              \
+        sum3_r = __lsx_vsub_w(temp1_r, sum3_r);                              \
+        sum3_l = __lsx_vsub_w(temp1_l, sum3_l);                              \
+                                                                             \
+        DUP2_ARG2(__lsx_vdp2_w_h, src3_r, filter1, src3_l, filter1,          \
+                  temp0_r, temp0_l);                                         \
+        DUP4_ARG3(__lsx_vdp2add_w_h, sum2_r, src7_r, filter3, sum2_l,        \
+                  src7_l, filter3, sum3_r, src4_r, filter3, sum3_l,          \
+                  src4_l, filter3, sum2_r, sum2_l, sum3_r, sum3_l);          \
+                                                                             \
+        sum0_r = __lsx_vadd_w(sum0_r, temp0_r);                              \
+        sum0_l = __lsx_vadd_w(sum0_l, temp0_l);                              \
+        sum1_r = __lsx_vsub_w(sum1_r, temp0_r);                              \
+        sum1_l = __lsx_vsub_w(sum1_l, temp0_l);                              \
+                                                                             \
+        LSX_BUTTERFLY_4_W(sum0_r, sum0_l, sum2_l, sum2_r, res0_r, res0_l,    \
+                          res1_l, res1_r);                                   \
+        dst0 = __lsx_vssrarni_h_w(res0_l, res0_r, shift);                    \
+        dst1 = __lsx_vssrarni_h_w(res1_l, res1_r, shift);                    \
+        __lsx_vst(dst0, buf_ptr, 0);                                         \
+        __lsx_vst(dst1, (buf_ptr + ((15 - (j * 2)) << 4)), 0);               \
+                                                                             \
+        LSX_BUTTERFLY_4_W(sum1_r, sum1_l, sum3_l, sum3_r, res0_r, res0_l,    \
+                          res1_l, res1_r);                                   \
+                                                                             \
+        dst0 = __lsx_vssrarni_h_w(res0_l, res0_r, shift);                    \
+        dst1 = __lsx_vssrarni_h_w(res1_l, res1_r, shift);                    \
+        __lsx_vst(dst0, (ptr0 + ((((j + 1) >> 1) * 2 * k) << 4)), 0);        \
+        __lsx_vst(dst1, (ptr1 - ((((j + 1) >> 1) * 2 * k) << 4)), 0);        \
+                                                                             \
+        k *= -1;                                                             \
+        buf_ptr += 16;                                                       \
+    }                                                                        \
+}
+
+#define HEVC_EVEN16_CALC(input, sum0_r, sum0_l, load_idx, store_idx)  \
+{                                                                     \
+    tmp0_r = __lsx_vld(input + load_idx * 8, 0);                      \
+    tmp0_l = __lsx_vld(input + load_idx * 8, 16);                     \
+    tmp1_r = sum0_r;                                                  \
+    tmp1_l = sum0_l;                                                  \
+    sum0_r = __lsx_vadd_w(sum0_r, tmp0_r);                            \
+    sum0_l = __lsx_vadd_w(sum0_l, tmp0_l);                            \
+    __lsx_vst(sum0_r, (input + load_idx * 8), 0);                     \
+    __lsx_vst(sum0_l, (input + load_idx * 8), 16);                    \
+    tmp1_r = __lsx_vsub_w(tmp1_r, tmp0_r);                            \
+    tmp1_l = __lsx_vsub_w(tmp1_l, tmp0_l);                            \
+    __lsx_vst(tmp1_r, (input + store_idx * 8), 0);                    \
+    __lsx_vst(tmp1_l, (input + store_idx * 8), 16);                   \
+}
+
+#define HEVC_IDCT_LUMA4x4_COL(in_r0, in_l0, in_r1, in_l1,     \
+                              res0, res1, res2, res3, shift)  \
+{                                                             \
+    __m128i vec0, vec1, vec2, vec3;                           \
+    __m128i cnst74 = __lsx_vldi(0x84a);                       \
+    __m128i cnst55 = __lsx_vldi(0x837);                       \
+    __m128i cnst29 = __lsx_vldi(0x81d);                       \
+                                                              \
+    vec0 = __lsx_vadd_w(in_r0, in_r1);                        \
+    vec2 = __lsx_vsub_w(in_r0, in_l1);                        \
+    res0 = __lsx_vmul_w(vec0, cnst29);                        \
+    res1 = __lsx_vmul_w(vec2, cnst55);                        \
+    res2 = __lsx_vsub_w(in_r0, in_r1);                        \
+    vec1 = __lsx_vadd_w(in_r1, in_l1);                        \
+    res2 = __lsx_vadd_w(res2, in_l1);                         \
+    vec3 = __lsx_vmul_w(in_l0, cnst74);                       \
+    res3 = __lsx_vmul_w(vec0, cnst55);                        \
+                                                              \
+    res0 = __lsx_vadd_w(res0, __lsx_vmul_w(vec1, cnst55));    \
+    res1 = __lsx_vsub_w(res1, __lsx_vmul_w(vec1, cnst29));    \
+    res2 = __lsx_vmul_w(res2, cnst74);                        \
+    res3 = __lsx_vadd_w(res3, __lsx_vmul_w(vec2, cnst29));    \
+                                                              \
+    res0 = __lsx_vadd_w(res0, vec3);                          \
+    res1 = __lsx_vadd_w(res1, vec3);                          \
+    res3 = __lsx_vsub_w(res3, vec3);                          \
+                                                              \
+    res0 = __lsx_vsrari_w(res0, shift);                       \
+    res1 = __lsx_vsrari_w(res1, shift);                       \
+    res2 = __lsx_vsrari_w(res2, shift);                       \
+    res3 = __lsx_vsrari_w(res3, shift);                       \
+    res0 = __lsx_vsat_w(res0, 15);                            \
+    res1 = __lsx_vsat_w(res1, 15);                            \
+    res2 = __lsx_vsat_w(res2, 15);                            \
+    res3 = __lsx_vsat_w(res3, 15);                            \
+}
+
+void ff_hevc_idct_4x4_lsx(int16_t *coeffs, int col_limit)
+{
+    __m128i in0, in1;
+    __m128i in_r0, in_l0, in_r1, in_l1;
+    __m128i sum0, sum1, sum2, sum3;
+    __m128i zero = __lsx_vldi(0x00);
+
+    in0   = __lsx_vld(coeffs, 0);
+    in1   = __lsx_vld(coeffs, 16);
+    in_r0 = __lsx_vilvl_h(zero, in0);
+    in_l0 = __lsx_vilvh_h(zero, in0);
+    in_r1 = __lsx_vilvl_h(zero, in1);
+    in_l1 = __lsx_vilvh_h(zero, in1);
+
+    HEVC_IDCT4x4_COL(in_r0, in_l0, in_r1, in_l1, sum0, sum1, sum2, sum3, 7);
+    LSX_TRANSPOSE4x4_W(sum0, sum1, sum2, sum3, in_r0, in_l0, in_r1, in_l1);
+    HEVC_IDCT4x4_COL(in_r0, in_l0, in_r1, in_l1, sum0, sum1, sum2, sum3, 12);
+
+    /* Pack and transpose */
+    in0  = __lsx_vpickev_h(sum2, sum0);
+    in1  = __lsx_vpickev_h(sum3, sum1);
+    sum0 = __lsx_vilvl_h(in1, in0);
+    sum1 = __lsx_vilvh_h(in1, in0);
+    in0  = __lsx_vilvl_w(sum1, sum0);
+    in1  = __lsx_vilvh_w(sum1, sum0);
+
+    __lsx_vst(in0, coeffs, 0);
+    __lsx_vst(in1, coeffs, 16);
+}
+
+void ff_hevc_idct_8x8_lsx(int16_t *coeffs, int col_limit)
+{
+    const int16_t *filter = &gt8x8_cnst[0];
+    __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+
+    DUP4_ARG2(__lsx_vld, coeffs, 0, coeffs, 16, coeffs, 32,
+              coeffs, 48, in0, in1, in2, in3);
+    DUP4_ARG2(__lsx_vld, coeffs, 64, coeffs, 80, coeffs, 96,
+              coeffs, 112, in4, in5, in6, in7);
+    HEVC_IDCT8x8_COL(in0, in1, in2, in3, in4, in5, in6, in7, 7);
+    LSX_TRANSPOSE8x8_H(in0, in1, in2, in3, in4, in5, in6, in7,
+                       in0, in1, in2, in3, in4, in5, in6, in7);
+    HEVC_IDCT8x8_COL(in0, in1, in2, in3, in4, in5, in6, in7, 12);
+    LSX_TRANSPOSE8x8_H(in0, in1, in2, in3, in4, in5, in6, in7,
+                       in0, in1, in2, in3, in4, in5, in6, in7);
+
+    __lsx_vst(in0, coeffs, 0);
+    __lsx_vst(in1, coeffs, 16);
+    __lsx_vst(in2, coeffs, 32);
+    __lsx_vst(in3, coeffs, 48);
+    __lsx_vst(in4, coeffs, 64);
+    __lsx_vst(in5, coeffs, 80);
+    __lsx_vst(in6, coeffs, 96);
+    __lsx_vst(in7, coeffs, 112);
+}
+
+void ff_hevc_idct_16x16_lsx(int16_t *coeffs, int col_limit)
+{
+    int16_t i, j, k;
+    int16_t buf[256];
+    int16_t *buf_ptr = &buf[0];
+    int16_t *src = coeffs;
+    const int16_t *filter = &gt16x16_cnst[0];
+    __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+    __m128i in8, in9, in10, in11, in12, in13, in14, in15;
+    __m128i vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+    __m128i src0_r, src1_r, src2_r, src3_r, src4_r, src5_r, src6_r, src7_r;
+    __m128i src0_l, src1_l, src2_l, src3_l, src4_l, src5_l, src6_l, src7_l;
+
+    for (i = 2; i--;) {
+        DUP4_ARG2(__lsx_vld, src, 0, src, 32, src, 64, src, 96,
+                  in0, in1, in2, in3);
+        DUP4_ARG2(__lsx_vld, src, 128, src, 160, src, 192, src, 224,
+                  in4, in5, in6, in7);
+        DUP4_ARG2(__lsx_vld, src, 256, src, 288, src, 320, src, 352,
+                  in8, in9, in10, in11);
+        DUP4_ARG2(__lsx_vld, src, 384, src, 416, src, 448, src, 480,
+                  in12, in13, in14, in15);
+
+        DUP4_ARG2(__lsx_vilvl_h, in4, in0, in12, in8, in6, in2, in14, in10,
+                  src0_r, src1_r, src2_r, src3_r);
+        DUP4_ARG2(__lsx_vilvl_h, in5, in1, in13, in9, in3, in7, in11, in15,
+                  src4_r, src5_r, src6_r, src7_r);
+        DUP4_ARG2(__lsx_vilvh_h, in4, in0, in12, in8, in6, in2, in14, in10,
+                  src0_l, src1_l, src2_l, src3_l);
+        DUP4_ARG2(__lsx_vilvh_h, in5, in1, in13, in9, in3, in7, in11, in15,
+                  src4_l, src5_l, src6_l, src7_l);
+
+        HEVC_IDCT16x16_COL(src0_r, src1_r, src2_r, src3_r, src4_r, src5_r,
+                           src6_r, src7_r, src0_l, src1_l, src2_l, src3_l,
+                           src4_l, src5_l, src6_l, src7_l, 7);
+
+        src += 8;
+        buf_ptr = (&buf[0] + 8);
+        filter = &gt16x16_cnst[0];
+    }
+
+    src = &buf[0];
+    buf_ptr = coeffs;
+    filter = &gt16x16_cnst[0];
+
+    for (i = 2; i--;) {
+        DUP4_ARG2(__lsx_vld, src, 0, src, 16, src, 32, src, 48,
+                  in0, in8, in1, in9);
+        DUP4_ARG2(__lsx_vld, src, 64, src, 80, src, 96, src, 112,
+                  in2, in10, in3, in11);
+        DUP4_ARG2(__lsx_vld, src, 128, src, 144, src, 160, src, 176,
+                  in4, in12, in5, in13);
+        DUP4_ARG2(__lsx_vld, src, 192, src, 208, src, 224, src, 240,
+                  in6, in14, in7, in15);
+        LSX_TRANSPOSE8x8_H(in0, in1, in2, in3, in4, in5, in6, in7,
+                           in0, in1, in2, in3, in4, in5, in6, in7);
+        LSX_TRANSPOSE8x8_H(in8, in9, in10, in11, in12, in13, in14, in15,
+                           in8, in9, in10, in11, in12, in13, in14, in15);
+        DUP4_ARG2(__lsx_vilvl_h, in4, in0, in12, in8, in6, in2, in14, in10,
+                  src0_r, src1_r, src2_r, src3_r);
+        DUP4_ARG2(__lsx_vilvl_h, in5, in1, in13, in9, in3, in7, in11, in15,
+                  src4_r, src5_r, src6_r, src7_r);
+        DUP4_ARG2(__lsx_vilvh_h, in4, in0, in12, in8, in6, in2, in14, in10,
+                  src0_l, src1_l, src2_l, src3_l);
+        DUP4_ARG2(__lsx_vilvh_h, in5, in1, in13, in9, in3, in7, in11, in15,
+                  src4_l, src5_l, src6_l, src7_l);
+        HEVC_IDCT16x16_COL(src0_r, src1_r, src2_r, src3_r, src4_r, src5_r,
+                           src6_r, src7_r, src0_l, src1_l, src2_l, src3_l,
+                           src4_l, src5_l, src6_l, src7_l, 12);
+
+        src += 128;
+        buf_ptr = coeffs + 8;
+        filter = &gt16x16_cnst[0];
+    }
+
+    DUP4_ARG2(__lsx_vld, coeffs, 0, coeffs, 32, coeffs, 64, coeffs, 96,
+              in0, in1, in2, in3);
+    DUP4_ARG2(__lsx_vld, coeffs, 128, coeffs, 160, coeffs, 192, coeffs, 224,
+              in4, in5, in6, in7);
+    LSX_TRANSPOSE8x8_H(in0, in1, in2, in3, in4, in5, in6, in7,
+                       vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7);
+    __lsx_vst(vec0, coeffs, 0);
+    __lsx_vst(vec1, coeffs, 32);
+    __lsx_vst(vec2, coeffs, 64);
+    __lsx_vst(vec3, coeffs, 96);
+    __lsx_vst(vec4, coeffs, 128);
+    __lsx_vst(vec5, coeffs, 160);
+    __lsx_vst(vec6, coeffs, 192);
+    __lsx_vst(vec7, coeffs, 224);
+
+    src = coeffs + 8;
+    DUP4_ARG2(__lsx_vld, src, 0, src, 32, src, 64, src, 96, in0, in1, in2, in3);
+    DUP4_ARG2(__lsx_vld, src, 128, src, 160, src, 192, src, 224,
+              in4, in5, in6, in7);
+    LSX_TRANSPOSE8x8_H(in0, in1, in2, in3, in4, in5, in6, in7,
+                       vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7);
+    src = coeffs + 128;
+    DUP4_ARG2(__lsx_vld, src, 0, src, 32, src, 64, src, 96,
+              in8, in9, in10, in11);
+    DUP4_ARG2(__lsx_vld, src, 128, src, 160, src, 192, src, 224,
+              in12, in13, in14, in15);
+
+    __lsx_vst(vec0, src, 0);
+    __lsx_vst(vec1, src, 32);
+    __lsx_vst(vec2, src, 64);
+    __lsx_vst(vec3, src, 96);
+    __lsx_vst(vec4, src, 128);
+    __lsx_vst(vec5, src, 160);
+    __lsx_vst(vec6, src, 192);
+    __lsx_vst(vec7, src, 224);
+    LSX_TRANSPOSE8x8_H(in8, in9, in10, in11, in12, in13, in14, in15,
+                       vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7);
+    src = coeffs + 8;
+    __lsx_vst(vec0, src, 0);
+    __lsx_vst(vec1, src, 32);
+    __lsx_vst(vec2, src, 64);
+    __lsx_vst(vec3, src, 96);
+    __lsx_vst(vec4, src, 128);
+    __lsx_vst(vec5, src, 160);
+    __lsx_vst(vec6, src, 192);
+    __lsx_vst(vec7, src, 224);
+
+    src = coeffs + 136;
+    DUP4_ARG2(__lsx_vld, src, 0, src, 32, src, 64, src, 96,
+              in0, in1, in2, in3);
+    DUP4_ARG2(__lsx_vld, src, 128, src, 160, src, 192, src, 224,
+              in4, in5, in6, in7);
+    LSX_TRANSPOSE8x8_H(in0, in1, in2, in3, in4, in5, in6, in7,
+                       vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7);
+    __lsx_vst(vec0, src, 0);
+    __lsx_vst(vec1, src, 32);
+    __lsx_vst(vec2, src, 64);
+    __lsx_vst(vec3, src, 96);
+    __lsx_vst(vec4, src, 128);
+    __lsx_vst(vec5, src, 160);
+    __lsx_vst(vec6, src, 192);
+    __lsx_vst(vec7, src, 224);
+}
+
+static void hevc_idct_8x32_column_lsx(int16_t *coeffs, int32_t buf_pitch,
+                                      uint8_t round)
+{
+    uint8_t i;
+    int32_t buf_pitch_2  = buf_pitch << 1;
+    int32_t buf_pitch_4  = buf_pitch << 2;
+    int32_t buf_pitch_8  = buf_pitch << 3;
+    int32_t buf_pitch_16 = buf_pitch << 4;
+
+    const int16_t *filter_ptr0 = &gt32x32_cnst0[0];
+    const int16_t *filter_ptr1 = &gt32x32_cnst1[0];
+    const int16_t *filter_ptr2 = &gt32x32_cnst2[0];
+    const int16_t *filter_ptr3 = &gt8x8_cnst[0];
+    int16_t *src0 = (coeffs + buf_pitch);
+    int16_t *src1 = (coeffs + buf_pitch_2);
+    int16_t *src2 = (coeffs + buf_pitch_4);
+    int16_t *src3 = (coeffs);
+    int32_t tmp_buf[8 * 32 + 15];
+    int32_t *tmp_buf_ptr = tmp_buf + 15;
+    __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+    __m128i src0_r, src1_r, src2_r, src3_r, src4_r, src5_r, src6_r, src7_r;
+    __m128i src0_l, src1_l, src2_l, src3_l, src4_l, src5_l, src6_l, src7_l;
+    __m128i filter0, filter1, filter2, filter3;
+    __m128i sum0_r, sum0_l, sum1_r, sum1_l, tmp0_r, tmp0_l, tmp1_r, tmp1_l;
+
+    /* Align pointer to 64 byte boundary */
+    tmp_buf_ptr = (int32_t *)(((uintptr_t) tmp_buf_ptr) & ~(uintptr_t) 63);
+
+    /* process coeff 4, 12, 20, 28 */
+    in0 = __lsx_vld(src2, 0);
+    in1 = __lsx_vld(src2 + buf_pitch_8, 0);
+    in2 = __lsx_vld(src2 + buf_pitch_16, 0);
+    in3 = __lsx_vld(src2 + buf_pitch_16 + buf_pitch_8, 0);
+    in4 = __lsx_vld(src3, 0);
+    in5 = __lsx_vld(src3 + buf_pitch_8, 0);
+    in6 = __lsx_vld(src3 + buf_pitch_16, 0);
+    in7 = __lsx_vld(src3 + buf_pitch_16 + buf_pitch_8, 0);
+    DUP4_ARG2(__lsx_vilvl_h, in1, in0, in3, in2, in6, in4, in7, in5,
+              src0_r, src1_r, src2_r, src3_r);
+    DUP4_ARG2(__lsx_vilvh_h, in1, in0, in3, in2, in6, in4, in7, in5,
+              src0_l, src1_l, src2_l, src3_l);
+
+    filter0 = __lsx_vldrepl_w(filter_ptr2, 0);
+    filter1 = __lsx_vldrepl_w(filter_ptr2, 4);
+    sum0_r = __lsx_vdp2_w_h(src0_r, filter0);
+    sum0_l = __lsx_vdp2_w_h(src0_l, filter0);
+    sum0_r = __lsx_vdp2add_w_h(sum0_r, src1_r, filter1);
+    sum0_l = __lsx_vdp2add_w_h(sum0_l, src1_l, filter1);
+    __lsx_vst(sum0_r, tmp_buf_ptr, 0);
+    __lsx_vst(sum0_l, tmp_buf_ptr, 16);
+
+    filter0 = __lsx_vldrepl_w(filter_ptr2, 8);
+    filter1 = __lsx_vldrepl_w(filter_ptr2, 12);
+    sum0_r = __lsx_vdp2_w_h(src0_r, filter0);
+    sum0_l = __lsx_vdp2_w_h(src0_l, filter0);
+    sum0_r = __lsx_vdp2add_w_h(sum0_r, src1_r, filter1);
+    sum0_l = __lsx_vdp2add_w_h(sum0_l, src1_l, filter1);
+    __lsx_vst(sum0_r, tmp_buf_ptr, 32);
+    __lsx_vst(sum0_l, tmp_buf_ptr, 48);
+
+    filter0 = __lsx_vldrepl_w(filter_ptr2, 16);
+    filter1 = __lsx_vldrepl_w(filter_ptr2, 20);
+    sum0_r = __lsx_vdp2_w_h(src0_r, filter0);
+    sum0_l = __lsx_vdp2_w_h(src0_l, filter0);
+    sum0_r = __lsx_vdp2add_w_h(sum0_r, src1_r, filter1);
+    sum0_l = __lsx_vdp2add_w_h(sum0_l, src1_l, filter1);
+    __lsx_vst(sum0_r, tmp_buf_ptr, 64);
+    __lsx_vst(sum0_l, tmp_buf_ptr, 80);
+
+    filter0 = __lsx_vldrepl_w(filter_ptr2, 24);
+    filter1 = __lsx_vldrepl_w(filter_ptr2, 28);
+    sum0_r = __lsx_vdp2_w_h(src0_r, filter0);
+    sum0_l = __lsx_vdp2_w_h(src0_l, filter0);
+    sum0_r = __lsx_vdp2add_w_h(sum0_r, src1_r, filter1);
+    sum0_l = __lsx_vdp2add_w_h(sum0_l, src1_l, filter1);
+    __lsx_vst(sum0_r, tmp_buf_ptr, 96);
+    __lsx_vst(sum0_l, tmp_buf_ptr, 112);
+
+    /* process coeff 0, 8, 16, 24 */
+    filter0 = __lsx_vldrepl_w(filter_ptr3, 0);
+    filter1 = __lsx_vldrepl_w(filter_ptr3, 4);
+
+    DUP4_ARG2(__lsx_vdp2_w_h, src2_r, filter0, src2_l, filter0,
+              src3_r, filter1, src3_l, filter1, sum0_r, sum0_l, tmp1_r, tmp1_l);
+    sum1_r = __lsx_vsub_w(sum0_r, tmp1_r);
+    sum1_l = __lsx_vsub_w(sum0_l, tmp1_l);
+    sum0_r = __lsx_vadd_w(sum0_r, tmp1_r);
+    sum0_l = __lsx_vadd_w(sum0_l, tmp1_l);
+
+    HEVC_EVEN16_CALC(tmp_buf_ptr, sum0_r, sum0_l, 0, 7);
+    HEVC_EVEN16_CALC(tmp_buf_ptr, sum1_r, sum1_l, 3, 4);
+
+    filter0 = __lsx_vldrepl_w(filter_ptr3, 16);
+    filter1 = __lsx_vldrepl_w(filter_ptr3, 20);
+
+    DUP4_ARG2(__lsx_vdp2_w_h, src2_r, filter0, src2_l, filter0,
+              src3_r, filter1, src3_l, filter1, sum0_r, sum0_l, tmp1_r, tmp1_l);
+    sum1_r = __lsx_vsub_w(sum0_r, tmp1_r);
+    sum1_l = __lsx_vsub_w(sum0_l, tmp1_l);
+    sum0_r = __lsx_vadd_w(sum0_r, tmp1_r);
+    sum0_l = __lsx_vadd_w(sum0_l, tmp1_l);
+
+    HEVC_EVEN16_CALC(tmp_buf_ptr, sum0_r, sum0_l, 1, 6);
+    HEVC_EVEN16_CALC(tmp_buf_ptr, sum1_r, sum1_l, 2, 5);
+
+    /* process coeff 2 6 10 14 18 22 26 30 */
+    in0 = __lsx_vld(src1, 0);
+    in1 = __lsx_vld(src1 + buf_pitch_4, 0);
+    in2 = __lsx_vld(src1 + buf_pitch_8, 0);
+    in3 = __lsx_vld(src1 + buf_pitch_8 + buf_pitch_4, 0);
+    in4 = __lsx_vld(src1 + buf_pitch_16, 0);
+    in5 = __lsx_vld(src1 + buf_pitch_16 + buf_pitch_4, 0);
+    in6 = __lsx_vld(src1 + buf_pitch_16 + buf_pitch_8, 0);
+    in7 = __lsx_vld(src1 + buf_pitch_16 + buf_pitch_8 + buf_pitch_4, 0);
+
+    DUP4_ARG2(__lsx_vilvl_h, in1, in0, in3, in2, in5, in4, in7, in6,
+              src0_r, src1_r, src2_r, src3_r);
+    DUP4_ARG2(__lsx_vilvh_h, in1, in0, in3, in2, in5, in4, in7, in6,
+              src0_l, src1_l, src2_l, src3_l);
+
+    /* loop for all columns of constants */
+    for (i = 0; i < 8; i++) {
+        /* processing single column of constants */
+        filter0 = __lsx_vldrepl_w(filter_ptr1, 0);
+        filter1 = __lsx_vldrepl_w(filter_ptr1, 4);
+        filter2 = __lsx_vldrepl_w(filter_ptr1, 8);
+        filter3 = __lsx_vldrepl_w(filter_ptr1, 12);
+        sum0_r = __lsx_vdp2_w_h(src0_r, filter0);
+        sum0_l = __lsx_vdp2_w_h(src0_l, filter0);
+        sum0_r = __lsx_vdp2add_w_h(sum0_r, src1_r, filter1);
+        sum0_l = __lsx_vdp2add_w_h(sum0_l, src1_l, filter1);
+        sum0_r = __lsx_vdp2add_w_h(sum0_r, src2_r, filter2);
+        sum0_l = __lsx_vdp2add_w_h(sum0_l, src2_l, filter2);
+        sum0_r = __lsx_vdp2add_w_h(sum0_r, src3_r, filter3);
+        sum0_l = __lsx_vdp2add_w_h(sum0_l, src3_l, filter3);
+
+        tmp0_r = __lsx_vld(tmp_buf_ptr + (i << 3), 0);
+        tmp0_l = __lsx_vld(tmp_buf_ptr + (i << 3), 16);
+        tmp1_r = tmp0_r;
+        tmp1_l = tmp0_l;
+        tmp0_r = __lsx_vadd_w(tmp0_r, sum0_r);
+        tmp0_l = __lsx_vadd_w(tmp0_l, sum0_l);
+        tmp1_r = __lsx_vsub_w(tmp1_r, sum0_r);
+        tmp1_l = __lsx_vsub_w(tmp1_l, sum0_l);
+        __lsx_vst(tmp0_r, tmp_buf_ptr + (i << 3), 0);
+        __lsx_vst(tmp0_l, tmp_buf_ptr + (i << 3), 16);
+        __lsx_vst(tmp1_r, tmp_buf_ptr + ((15 - i) * 8), 0);
+        __lsx_vst(tmp1_l, tmp_buf_ptr + ((15 - i) * 8), 16);
+
+        filter_ptr1 += 8;
+    }
+
+    /* process coeff 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 */
+    in0 = __lsx_vld(src0, 0);
+    in1 = __lsx_vld(src0 + buf_pitch_2, 0);
+    in2 = __lsx_vld(src0 + buf_pitch_4, 0);
+    in3 = __lsx_vld(src0 + buf_pitch_4 + buf_pitch_2, 0);
+    in4 = __lsx_vld(src0 + buf_pitch_8, 0);
+    in5 = __lsx_vld(src0 + buf_pitch_8 + buf_pitch_2, 0);
+    in6 = __lsx_vld(src0 + buf_pitch_8 + buf_pitch_4, 0);
+    in7 = __lsx_vld(src0 + buf_pitch_8 + buf_pitch_4 + buf_pitch_2, 0);
+
+    src0 += 16 * buf_pitch;
+    DUP4_ARG2(__lsx_vilvl_h, in1, in0, in3, in2, in5, in4, in7, in6,
+              src0_r, src1_r, src2_r, src3_r);
+    DUP4_ARG2(__lsx_vilvh_h, in1, in0, in3, in2, in5, in4, in7, in6,
+              src0_l, src1_l, src2_l, src3_l);
+    in0 = __lsx_vld(src0, 0);
+    in1 = __lsx_vld(src0 + buf_pitch_2, 0);
+    in2 = __lsx_vld(src0 + buf_pitch_4, 0);
+    in3 = __lsx_vld(src0 + buf_pitch_4 + buf_pitch_2, 0);
+    in4 = __lsx_vld(src0 + buf_pitch_8, 0);
+    in5 = __lsx_vld(src0 + buf_pitch_8 + buf_pitch_2, 0);
+    in6 = __lsx_vld(src0 + buf_pitch_8 + buf_pitch_4, 0);
+    in7 = __lsx_vld(src0 + buf_pitch_8 + buf_pitch_4 + buf_pitch_2, 0);
+
+    DUP4_ARG2(__lsx_vilvl_h, in1, in0, in3, in2, in5, in4, in7, in6,
+              src4_r, src5_r, src6_r, src7_r);
+    DUP4_ARG2(__lsx_vilvh_h, in1, in0, in3, in2, in5, in4, in7, in6,
+              src4_l, src5_l, src6_l, src7_l);
+
+    /* loop for all columns of filter constants */
+    for (i = 0; i < 16; i++) {
+        /* processing single column of constants */
+        filter0 = __lsx_vldrepl_w(filter_ptr0, 0);
+        filter1 = __lsx_vldrepl_w(filter_ptr0, 4);
+        filter2 = __lsx_vldrepl_w(filter_ptr0, 8);
+        filter3 = __lsx_vldrepl_w(filter_ptr0, 12);
+        sum0_r = __lsx_vdp2_w_h(src0_r, filter0);
+        sum0_l = __lsx_vdp2_w_h(src0_l, filter0);
+        sum0_r = __lsx_vdp2add_w_h(sum0_r, src1_r, filter1);
+        sum0_l = __lsx_vdp2add_w_h(sum0_l, src1_l, filter1);
+        sum0_r = __lsx_vdp2add_w_h(sum0_r, src2_r, filter2);
+        sum0_l = __lsx_vdp2add_w_h(sum0_l, src2_l, filter2);
+        sum0_r = __lsx_vdp2add_w_h(sum0_r, src3_r, filter3);
+        sum0_l = __lsx_vdp2add_w_h(sum0_l, src3_l, filter3);
+        tmp1_r = sum0_r;
+        tmp1_l = sum0_l;
+
+        filter0 = __lsx_vldrepl_w(filter_ptr0, 16);
+        filter1 = __lsx_vldrepl_w(filter_ptr0, 20);
+        filter2 = __lsx_vldrepl_w(filter_ptr0, 24);
+        filter3 = __lsx_vldrepl_w(filter_ptr0, 28);
+        sum0_r = __lsx_vdp2_w_h(src4_r, filter0);
+        sum0_l = __lsx_vdp2_w_h(src4_l, filter0);
+        sum0_r = __lsx_vdp2add_w_h(sum0_r, src5_r, filter1);
+        sum0_l = __lsx_vdp2add_w_h(sum0_l, src5_l, filter1);
+        sum0_r = __lsx_vdp2add_w_h(sum0_r, src6_r, filter2);
+        sum0_l = __lsx_vdp2add_w_h(sum0_l, src6_l, filter2);
+        sum0_r = __lsx_vdp2add_w_h(sum0_r, src7_r, filter3);
+        sum0_l = __lsx_vdp2add_w_h(sum0_l, src7_l, filter3);
+        sum0_r = __lsx_vadd_w(sum0_r, tmp1_r);
+        sum0_l = __lsx_vadd_w(sum0_l, tmp1_l);
+
+        tmp0_r = __lsx_vld(tmp_buf_ptr + i * 8, 0);
+        tmp0_l = __lsx_vld(tmp_buf_ptr + i * 8, 16);
+        tmp1_r = tmp0_r;
+        tmp1_l = tmp0_l;
+        tmp0_r = __lsx_vadd_w(tmp0_r, sum0_r);
+        tmp0_l = __lsx_vadd_w(tmp0_l, sum0_l);
+        sum1_r = __lsx_vreplgr2vr_w(round);
+        tmp0_r = __lsx_vssrarn_h_w(tmp0_r, sum1_r);
+        tmp0_l = __lsx_vssrarn_h_w(tmp0_l, sum1_r);
+        in0    = __lsx_vpackev_d(tmp0_l, tmp0_r);
+        __lsx_vst(in0, (coeffs + i * buf_pitch), 0);
+        tmp1_r = __lsx_vsub_w(tmp1_r, sum0_r);
+        tmp1_l = __lsx_vsub_w(tmp1_l, sum0_l);
+        tmp1_r = __lsx_vssrarn_h_w(tmp1_r, sum1_r);
+        tmp1_l = __lsx_vssrarn_h_w(tmp1_l, sum1_r);
+        in0    = __lsx_vpackev_d(tmp1_l, tmp1_r);
+        __lsx_vst(in0, (coeffs + (31 - i) * buf_pitch), 0);
+
+        filter_ptr0 += 16;
+    }
+}
+
+static void hevc_idct_transpose_32x8_to_8x32(int16_t *coeffs, int16_t *tmp_buf)
+{
+    uint8_t i;
+    __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+
+    for (i = 0; i < 4; i++) {
+        DUP4_ARG2(__lsx_vld, coeffs, 0, coeffs, 64, coeffs, 128,
+                  coeffs, 192, in0, in1, in2, in3);
+        DUP4_ARG2(__lsx_vld, coeffs, 256, coeffs, 320, coeffs, 384,
+                  coeffs, 448, in4, in5, in6, in7);
+        coeffs += 8;
+        LSX_TRANSPOSE8x8_H(in0, in1, in2, in3, in4, in5, in6, in7,
+                           in0, in1, in2, in3, in4, in5, in6, in7);
+        __lsx_vst(in0, tmp_buf, 0);
+        __lsx_vst(in1, tmp_buf, 16);
+        __lsx_vst(in2, tmp_buf, 32);
+        __lsx_vst(in3, tmp_buf, 48);
+        __lsx_vst(in4, tmp_buf, 64);
+        __lsx_vst(in5, tmp_buf, 80);
+        __lsx_vst(in6, tmp_buf, 96);
+        __lsx_vst(in7, tmp_buf, 112);
+        tmp_buf += 64;
+    }
+}
+
+static void hevc_idct_transpose_8x32_to_32x8(int16_t *tmp_buf, int16_t *coeffs)
+{
+    uint8_t i;
+    __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+
+    for (i = 0; i < 4; i++) {
+        DUP4_ARG2(__lsx_vld, tmp_buf, 0, tmp_buf, 16, tmp_buf, 32,
+                  tmp_buf, 48, in0, in1, in2, in3);
+        DUP4_ARG2(__lsx_vld, tmp_buf, 64, tmp_buf, 80, tmp_buf, 96,
+                  tmp_buf, 112, in4, in5, in6, in7);
+        tmp_buf += 64;
+        LSX_TRANSPOSE8x8_H(in0, in1, in2, in3, in4, in5, in6, in7,
+                           in0, in1, in2, in3, in4, in5, in6, in7);
+        __lsx_vst(in0, coeffs, 0);
+        __lsx_vst(in1, coeffs, 64);
+        __lsx_vst(in2, coeffs, 128);
+        __lsx_vst(in3, coeffs, 192);
+        __lsx_vst(in4, coeffs, 256);
+        __lsx_vst(in5, coeffs, 320);
+        __lsx_vst(in6, coeffs, 384);
+        __lsx_vst(in7, coeffs, 448);
+        coeffs += 8;
+    }
+}
+
+void ff_hevc_idct_32x32_lsx(int16_t *coeffs, int col_limit)
+{
+    uint8_t row_cnt, col_cnt;
+    int16_t *src = coeffs;
+    int16_t tmp_buf[8 * 32 + 31];
+    int16_t *tmp_buf_ptr = tmp_buf + 31;
+    uint8_t round;
+    int32_t buf_pitch;
+
+    /* Align pointer to 64 byte boundary */
+    tmp_buf_ptr = (int16_t *)(((uintptr_t) tmp_buf_ptr) & ~(uintptr_t) 63);
+
+    /* column transform */
+    round = 7;
+    buf_pitch = 32;
+    for (col_cnt = 0; col_cnt < 4; col_cnt++) {
+        /* process 8x32 blocks */
+        hevc_idct_8x32_column_lsx((coeffs + col_cnt * 8), buf_pitch, round);
+    }
+
+    /* row transform */
+    round = 12;
+    buf_pitch = 8;
+    for (row_cnt = 0; row_cnt < 4; row_cnt++) {
+        /* process 32x8 blocks */
+        src = (coeffs + 32 * 8 * row_cnt);
+
+        hevc_idct_transpose_32x8_to_8x32(src, tmp_buf_ptr);
+        hevc_idct_8x32_column_lsx(tmp_buf_ptr, buf_pitch, round);
+        hevc_idct_transpose_8x32_to_32x8(tmp_buf_ptr, src);
+    }
+}
diff --git a/libavcodec/loongarch/hevc_lpf_sao_lsx.c b/libavcodec/loongarch/hevc_lpf_sao_lsx.c
new file mode 100644
index 0000000000..fc10e8eda8
--- /dev/null
+++ b/libavcodec/loongarch/hevc_lpf_sao_lsx.c
@@ -0,0 +1,2485 @@
+/*
+ * Copyright (c) 2022 Loongson Technology Corporation Limited
+ * Contributed by Lu Wang <wanglu at loongson.cn>
+ *                Hao Chen <chenhao at loongson.cn>
+ *
+ * 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/loongarch/loongson_intrinsics.h"
+#include "hevcdsp_lsx.h"
+
+void ff_hevc_loop_filter_luma_h_8_lsx(uint8_t *src, ptrdiff_t stride,
+                                      int32_t beta, int32_t *tc,
+                                      uint8_t *p_is_pcm, uint8_t *q_is_pcm)
+{
+    ptrdiff_t stride_2x = (stride << 1);
+    ptrdiff_t stride_4x = (stride << 2);
+    ptrdiff_t stride_3x = stride_2x + stride;
+    uint8_t *p3 = src - stride_4x;
+    uint8_t *p2 = src - stride_3x;
+    uint8_t *p1 = src - stride_2x;
+    uint8_t *p0 = src - stride;
+    uint8_t *q0 = src;
+    uint8_t *q1 = src + stride;
+    uint8_t *q2 = src + stride_2x;
+    uint8_t *q3 = src + stride_3x;
+    uint8_t flag0, flag1;
+    int32_t dp00, dq00, dp30, dq30, d00, d30, d0030, d0434;
+    int32_t dp04, dq04, dp34, dq34, d04, d34;
+    int32_t tc0, p_is_pcm0, q_is_pcm0, beta30, beta20, tc250;
+    int32_t tc4, p_is_pcm4, q_is_pcm4, tc254, tmp;
+
+    __m128i dst0, dst1, dst2, dst3, dst4, dst5;
+    __m128i cmp0, cmp1, cmp2, cmp3, p_is_pcm_vec, q_is_pcm_vec;
+    __m128i temp0, temp1;
+    __m128i temp2, tc_pos, tc_neg;
+    __m128i diff0, diff1, delta0, delta1, delta2, abs_delta0;
+    __m128i zero = {0};
+    __m128i p3_src, p2_src, p1_src, p0_src, q0_src, q1_src, q2_src, q3_src;
+
+    dp00 = abs(p2[0] - (p1[0] << 1) + p0[0]);
+    dq00 = abs(q2[0] - (q1[0] << 1) + q0[0]);
+    dp30 = abs(p2[3] - (p1[3] << 1) + p0[3]);
+    dq30 = abs(q2[3] - (q1[3] << 1) + q0[3]);
+    d00 = dp00 + dq00;
+    d30 = dp30 + dq30;
+    dp04 = abs(p2[4] - (p1[4] << 1) + p0[4]);
+    dq04 = abs(q2[4] - (q1[4] << 1) + q0[4]);
+    dp34 = abs(p2[7] - (p1[7] << 1) + p0[7]);
+    dq34 = abs(q2[7] - (q1[7] << 1) + q0[7]);
+    d04 = dp04 + dq04;
+    d34 = dp34 + dq34;
+
+    p_is_pcm0 = p_is_pcm[0];
+    p_is_pcm4 = p_is_pcm[1];
+    q_is_pcm0 = q_is_pcm[0];
+    q_is_pcm4 = q_is_pcm[1];
+
+    DUP2_ARG1(__lsx_vreplgr2vr_d, p_is_pcm0, p_is_pcm4, cmp0, cmp1);
+    p_is_pcm_vec = __lsx_vpackev_d(cmp1, cmp0);
+    p_is_pcm_vec = __lsx_vseqi_d(p_is_pcm_vec, 0);
+    d0030 = (d00 + d30) >= beta;
+    d0434 = (d04 + d34) >= beta;
+    DUP2_ARG1(__lsx_vreplgr2vr_w, d0030, d0434, cmp0, cmp1);
+    cmp3 = __lsx_vpackev_w(cmp1, cmp0);
+    cmp3 = __lsx_vseqi_w(cmp3, 0);
+
+    if ((!p_is_pcm0 || !p_is_pcm4 || !q_is_pcm0 || !q_is_pcm4) &&
+        (!d0030 || !d0434)) {
+        DUP4_ARG2(__lsx_vld, p3, 0, p2, 0, p1, 0, p0, 0,
+                  p3_src, p2_src, p1_src, p0_src);
+        DUP2_ARG1(__lsx_vreplgr2vr_d, q_is_pcm0, q_is_pcm4, cmp0, cmp1);
+        q_is_pcm_vec = __lsx_vpackev_d(cmp1, cmp0);
+        q_is_pcm_vec = __lsx_vseqi_d(q_is_pcm_vec, 0);
+
+        tc0 = tc[0];
+        beta30 = beta >> 3;
+        beta20 = beta >> 2;
+        tc250 = (((tc0 << 2) + tc0 + 1) >> 1);
+        tc4 = tc[1];
+        tc254 = (((tc4 << 2) + tc4 + 1) >> 1);
+
+        DUP2_ARG1(__lsx_vreplgr2vr_h, tc0, tc4, cmp0, cmp1);
+        DUP4_ARG2(__lsx_vilvl_b, zero, p3_src, zero, p2_src, zero, p1_src, zero,
+                  p0_src, p3_src, p2_src, p1_src, p0_src);
+        DUP4_ARG2(__lsx_vld, q0, 0, q1, 0, q2, 0, q3, 0,
+                  q0_src, q1_src, q2_src, q3_src);
+        flag0 = abs(p3[0] - p0[0]) + abs(q3[0] - q0[0]) < beta30 &&
+                abs(p0[0] - q0[0]) < tc250;
+        flag0 = flag0 && (abs(p3[3] - p0[3]) + abs(q3[3] - q0[3]) < beta30 &&
+                abs(p0[3] - q0[3]) < tc250 && (d00 << 1) < beta20 &&
+                (d30 << 1) < beta20);
+        tc_pos = __lsx_vpackev_d(cmp1, cmp0);
+        DUP4_ARG2(__lsx_vilvl_b, zero, q0_src, zero, q1_src, zero, q2_src,
+                  zero, q3_src, q0_src, q1_src, q2_src, q3_src);
+
+        flag1 = abs(p3[4] - p0[4]) + abs(q3[4] - q0[4]) < beta30 &&
+                abs(p0[4] - q0[4]) < tc254;
+        flag1 = flag1 && (abs(p3[7] - p0[7]) + abs(q3[7] - q0[7]) < beta30 &&
+                abs(p0[7] - q0[7]) < tc254 && (d04 << 1) < beta20 &&
+                (d34 << 1) < beta20);
+        DUP2_ARG1(__lsx_vreplgr2vr_w, flag0, flag1, cmp0, cmp1);
+        cmp2 = __lsx_vpackev_w(cmp1, cmp0);
+        cmp2 = __lsx_vseqi_w(cmp2, 0);
+
+        if (flag0 && flag1) { /* strong only */
+            /* strong filter */
+            tc_pos = __lsx_vslli_h(tc_pos, 1);
+            tc_neg = __lsx_vneg_h(tc_pos);
+
+            /* p part */
+            DUP2_ARG2(__lsx_vadd_h, p1_src, p0_src, temp0, q0_src,
+                      temp0, temp0);
+            temp1 = __lsx_vadd_h(p3_src, p2_src);
+            temp1 = __lsx_vslli_h(temp1, 1);
+            DUP2_ARG2(__lsx_vadd_h, temp1, p2_src, temp1, temp0, temp1, temp1);
+            temp1 = __lsx_vsrari_h(temp1, 3);
+            temp2 = __lsx_vsub_h(temp1, p2_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst0 = __lsx_vadd_h(temp2, p2_src);
+
+            temp1 = __lsx_vadd_h(temp0, p2_src);
+            temp1 = __lsx_vsrari_h(temp1, 2);
+            temp2 = __lsx_vsub_h(temp1, p1_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst1 = __lsx_vadd_h(temp2, p1_src);
+
+            temp1 = __lsx_vslli_h(temp0, 1);
+            DUP2_ARG2(__lsx_vadd_h, temp1, p2_src, temp1, q1_src,
+                      temp1, temp1);
+            temp1 = __lsx_vsrari_h(temp1, 3);
+            temp2 = __lsx_vsub_h(temp1, p0_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst2 = __lsx_vadd_h(temp2, p0_src);
+
+            p_is_pcm_vec = __lsx_vnor_v(p_is_pcm_vec, p_is_pcm_vec);
+            DUP2_ARG3(__lsx_vbitsel_v, dst0, p2_src, p_is_pcm_vec, dst1,
+                      p1_src, p_is_pcm_vec, dst0, dst1);
+            dst2 = __lsx_vbitsel_v(dst2, p0_src, p_is_pcm_vec);
+
+            /* q part */
+            DUP2_ARG2(__lsx_vadd_h, q1_src, p0_src, temp0, q0_src,
+                      temp0, temp0);
+            temp1 = __lsx_vadd_h(q3_src, q2_src);
+            temp1 = __lsx_vslli_h(temp1, 1);
+            DUP2_ARG2(__lsx_vadd_h, temp1, q2_src, temp1, temp0, temp1, temp1);
+            temp1 = __lsx_vsrari_h(temp1, 3);
+            temp2 = __lsx_vsub_h(temp1, q2_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst5 = __lsx_vadd_h(temp2, q2_src);
+
+            temp1 = __lsx_vadd_h(temp0, q2_src);
+            temp1 = __lsx_vsrari_h(temp1, 2);
+            temp2 = __lsx_vsub_h(temp1, q1_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst4 = __lsx_vadd_h(temp2, q1_src);
+
+            temp0 = __lsx_vslli_h(temp0, 1);
+            DUP2_ARG2(__lsx_vadd_h, temp0, p1_src, temp1, q2_src,
+                      temp1, temp1);
+            temp1 = __lsx_vsrari_h(temp1, 3);
+            temp2 = __lsx_vsub_h(temp1, q0_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst3 = __lsx_vadd_h(temp2, q0_src);
+
+            q_is_pcm_vec = __lsx_vnor_v(q_is_pcm_vec, q_is_pcm_vec);
+            DUP2_ARG3(__lsx_vbitsel_v, dst3, q0_src, q_is_pcm_vec, dst4,
+                      q1_src, q_is_pcm_vec, dst3, dst4);
+            dst5 = __lsx_vbitsel_v(dst5, q2_src, q_is_pcm_vec);
+
+            /* pack results to 8 bit */
+            DUP2_ARG2(__lsx_vpickev_b, dst1, dst0, dst3, dst2, dst0, dst1);
+            dst2 = __lsx_vpickev_b(dst5, dst4);
+
+            /* pack src to 8 bit */
+            DUP2_ARG2(__lsx_vpickev_b, p1_src, p2_src, q0_src, p0_src,
+                      dst3, dst4);
+            dst5 = __lsx_vpickev_b(q2_src, q1_src);
+
+            cmp3 = __lsx_vnor_v(cmp3, cmp3);
+            DUP2_ARG3(__lsx_vbitsel_v, dst0, dst3, cmp3, dst1, dst4, cmp3,
+                      dst0, dst1);
+            dst2 = __lsx_vbitsel_v(dst2, dst5, cmp3);
+
+            __lsx_vstelm_d(dst0, p2, 0, 0);
+            __lsx_vstelm_d(dst0, p2 + stride, 0, 1);
+            __lsx_vstelm_d(dst1, p2 + stride_2x, 0, 0);
+            __lsx_vstelm_d(dst1, p2 + stride_3x, 0, 1);
+            __lsx_vstelm_d(dst2, p2 + stride_4x, 0, 0);
+            __lsx_vstelm_d(dst2, p2 + stride_4x + stride, 0, 1);
+            /* strong filter ends */
+        } else if (flag0 == flag1) { /* weak only */
+            /* weak filter */
+            tc_neg = __lsx_vneg_h(tc_pos);
+            DUP2_ARG2(__lsx_vsub_h, q0_src, p0_src, q1_src, p1_src,
+                      diff0, diff1);
+            DUP2_ARG2(__lsx_vadd_h, __lsx_vslli_h(diff0, 3), diff0,
+                      __lsx_vslli_h(diff1, 1), diff1, diff0, diff1);
+            delta0 = __lsx_vsub_h(diff0, diff1);
+            delta0 = __lsx_vsrari_h(delta0, 4);
+            temp1 = __lsx_vadd_h(__lsx_vslli_h(tc_pos, 3),
+                                 __lsx_vslli_h(tc_pos, 1));
+            abs_delta0 = __lsx_vadda_h(delta0, zero);
+            abs_delta0 = __lsx_vsle_hu(temp1, abs_delta0);
+            abs_delta0 = __lsx_vnor_v(abs_delta0, abs_delta0);
+
+            delta0 = __lsx_vclip_h(delta0, tc_neg, tc_pos);
+            temp2 = __lsx_vadd_h(delta0, p0_src);
+            temp2 = __lsx_vclip255_h(temp2);
+            temp0 = __lsx_vbitsel_v(temp2, p0_src,
+                                    __lsx_vnor_v(p_is_pcm_vec, p_is_pcm_vec));
+            temp2 = __lsx_vsub_h(q0_src, delta0);
+            temp2 = __lsx_vclip255_h(temp2);
+            temp2 = __lsx_vbitsel_v(temp2, q0_src, __lsx_vnor_v(q_is_pcm_vec,
+                                    q_is_pcm_vec));
+            DUP2_ARG2(__lsx_vnor_v, p_is_pcm_vec, p_is_pcm_vec, q_is_pcm_vec,
+                      q_is_pcm_vec, p_is_pcm_vec, q_is_pcm_vec);
+
+            tmp = (beta + (beta >> 1)) >> 3;
+            DUP2_ARG1(__lsx_vreplgr2vr_d, dp00 + dp30 < tmp, dp04 + dp34 < tmp,
+                      cmp0, cmp1);
+            cmp0 = __lsx_vpackev_d(cmp1, cmp0);
+            cmp0 = __lsx_vseqi_d(cmp0, 0);
+            p_is_pcm_vec = __lsx_vor_v(p_is_pcm_vec, cmp0);
+
+            DUP2_ARG1(__lsx_vreplgr2vr_d, dq00 + dq30 < tmp, dq04 + dq34 < tmp,
+                      cmp0, cmp1);
+            cmp0 = __lsx_vpackev_d(cmp1, cmp0);
+            cmp0 = __lsx_vseqi_d(cmp0, 0);
+            q_is_pcm_vec = __lsx_vor_v(q_is_pcm_vec, cmp0);
+            tc_pos = __lsx_vsrai_h(tc_pos, 1);
+            tc_neg = __lsx_vneg_h(tc_pos);
+
+            DUP2_ARG2(__lsx_vavgr_hu, p2_src, p0_src, q0_src, q2_src,
+                      delta1, delta2);
+            DUP2_ARG2(__lsx_vsub_h, delta1, p1_src, delta2, q1_src,
+                      delta1, delta2);
+            delta1 = __lsx_vadd_h(delta1, delta0);
+            delta2 = __lsx_vsub_h(delta2, delta0);
+            DUP2_ARG2(__lsx_vsrai_h, delta1, 1, delta2, 1, delta1, delta2);
+            DUP2_ARG3(__lsx_vclip_h, delta1, tc_neg, tc_pos, delta2,
+                      tc_neg, tc_pos, delta1, delta2);
+            DUP2_ARG2(__lsx_vadd_h, p1_src, delta1, q1_src, delta2,
+                      delta1, delta2);
+            DUP2_ARG1(__lsx_vclip255_h, delta1, delta2, delta1, delta2);
+            DUP2_ARG3(__lsx_vbitsel_v, delta1, p1_src, p_is_pcm_vec, delta2,
+                      q1_src, q_is_pcm_vec, delta1, delta2);
+
+            abs_delta0 = __lsx_vnor_v(abs_delta0, abs_delta0);
+            DUP4_ARG3(__lsx_vbitsel_v, delta1, p1_src, abs_delta0, temp0,
+                      p0_src,  abs_delta0, temp2, q0_src, abs_delta0, delta2,
+                      q1_src, abs_delta0, dst1, dst2, dst3, dst4);
+            /* pack results to 8 bit */
+            DUP2_ARG2(__lsx_vpickev_b, dst2, dst1, dst4, dst3, dst0, dst1);
+            /* pack src to 8 bit */
+            DUP2_ARG2(__lsx_vpickev_b, p0_src, p1_src, q1_src, q0_src,
+                      dst2, dst3);
+            cmp3 = __lsx_vnor_v(cmp3, cmp3);
+            DUP2_ARG3(__lsx_vbitsel_v, dst0, dst2, cmp3, dst1, dst3, cmp3,
+                      dst0, dst1);
+
+            p2 += stride;
+            __lsx_vstelm_d(dst0, p2, 0, 0);
+            __lsx_vstelm_d(dst0, p2 + stride, 0, 1);
+            __lsx_vstelm_d(dst1, p2 + stride_2x, 0, 0);
+            __lsx_vstelm_d(dst1, p2 + stride_3x, 0, 1);
+            /* weak filter ends */
+        } else { /* strong + weak */
+            /* strong filter */
+            tc_pos = __lsx_vslli_h(tc_pos, 1);
+            tc_neg = __lsx_vneg_h(tc_pos);
+
+            /* p part */
+            DUP2_ARG2(__lsx_vadd_h, p1_src, p0_src, temp0, q0_src,
+                      temp0, temp0);
+            temp1 = __lsx_vadd_h(p3_src, p2_src);
+            temp1 = __lsx_vslli_h(temp1, 1);
+            DUP2_ARG2(__lsx_vadd_h, temp1, p2_src, temp1, temp0, temp1, temp1);
+            temp1 = __lsx_vsrari_h(temp1, 3);
+            temp2 = __lsx_vsub_h(temp1, p2_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst0 = __lsx_vadd_h(temp2, p2_src);
+
+            temp1 = __lsx_vadd_h(temp0, p2_src);
+            temp1 = __lsx_vsrari_h(temp1, 2);
+            temp2 = __lsx_vsub_h(temp1, p1_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst1 = __lsx_vadd_h(temp2, p1_src);
+
+            temp1 = __lsx_vslli_h(temp0, 1);
+            DUP2_ARG2(__lsx_vadd_h, temp1, p2_src, temp1, q1_src, temp1, temp1);
+            temp1 = __lsx_vsrari_h(temp1, 3);
+            temp2 = __lsx_vsub_h(temp1, p0_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst2 = __lsx_vadd_h(temp2, p0_src);
+
+            p_is_pcm_vec = __lsx_vnor_v(p_is_pcm_vec, p_is_pcm_vec);
+            DUP2_ARG3(__lsx_vbitsel_v, dst0, p2_src, p_is_pcm_vec, dst1,
+                      p1_src, p_is_pcm_vec, dst0, dst1);
+            dst2 = __lsx_vbitsel_v(dst2, p0_src, p_is_pcm_vec);
+
+            /* q part */
+            DUP2_ARG2(__lsx_vadd_h, q1_src, p0_src, temp0, q0_src,
+                      temp0, temp0);
+            temp1 = __lsx_vadd_h(q3_src, q2_src);
+            temp1 = __lsx_vslli_h(temp1, 1);
+            DUP2_ARG2(__lsx_vadd_h, temp1,  q2_src, temp1, temp0, temp1, temp1);
+            temp1 = __lsx_vsrari_h(temp1, 3);
+            temp2 = __lsx_vsub_h(temp1, q2_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst5 = __lsx_vadd_h(temp2, q2_src);
+
+            temp1 = __lsx_vadd_h(temp0, q2_src);
+            temp1 = __lsx_vsrari_h(temp1, 2);
+            temp2 = __lsx_vsub_h(temp1, q1_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst4 = __lsx_vadd_h(temp2, q1_src);
+
+            temp1 = __lsx_vslli_h(temp0, 1);
+            DUP2_ARG2(__lsx_vadd_h, temp1, p1_src, temp1, q2_src, temp1, temp1);
+            temp1 = __lsx_vsrari_h(temp1, 3);
+            temp2 = __lsx_vsub_h(temp1, q0_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst3 = __lsx_vadd_h(temp2, q0_src);
+
+            q_is_pcm_vec = __lsx_vnor_v(q_is_pcm_vec, q_is_pcm_vec);
+            DUP2_ARG3(__lsx_vbitsel_v, dst3, q0_src, q_is_pcm_vec, dst4,
+                      q1_src, q_is_pcm_vec, dst3, dst4);
+            dst5 = __lsx_vbitsel_v(dst5, q2_src, q_is_pcm_vec);
+
+            /* pack strong results to 8 bit */
+            DUP2_ARG2(__lsx_vpickev_b, dst1, dst0, dst3, dst2, dst0, dst1);
+            dst2 = __lsx_vpickev_b(dst5, dst4);
+            /* strong filter ends */
+
+            /* weak filter */
+            tc_pos = __lsx_vsrai_h(tc_pos, 1);
+            tc_neg = __lsx_vneg_h(tc_pos);
+
+            DUP2_ARG2(__lsx_vsub_h, q0_src, p0_src, q1_src, p1_src,
+                      diff0, diff1);
+            DUP2_ARG2(__lsx_vadd_h, __lsx_vslli_h(diff0, 3), diff0,
+                      __lsx_vslli_h(diff1, 1), diff1, diff0, diff1);
+            delta0 = __lsx_vsub_h(diff0, diff1);
+            delta0 = __lsx_vsrari_h(delta0, 4);
+            temp1 = __lsx_vadd_h(__lsx_vslli_h(tc_pos, 3),
+                                 __lsx_vslli_h(tc_pos, 1));
+            abs_delta0 = __lsx_vadda_h(delta0, zero);
+            abs_delta0 = __lsx_vsle_hu(temp1, abs_delta0);
+            abs_delta0 = __lsx_vnor_v(abs_delta0, abs_delta0);
+
+            delta0 = __lsx_vclip_h(delta0, tc_neg, tc_pos);
+            temp2 = __lsx_vadd_h(delta0, p0_src);
+            temp2 = __lsx_vclip255_h(temp2);
+            temp0 = __lsx_vbitsel_v(temp2, p0_src, p_is_pcm_vec);
+
+            temp2 = __lsx_vsub_h(q0_src, delta0);
+            temp2 = __lsx_vclip255_h(temp2);
+            temp2 = __lsx_vbitsel_v(temp2, q0_src, q_is_pcm_vec);
+
+            tmp = (beta + (beta >> 1)) >> 3;
+            DUP2_ARG1(__lsx_vreplgr2vr_d, dp00 + dp30 < tmp, dp04 + dp34 < tmp,
+                      cmp0, cmp1);
+            cmp0 = __lsx_vpackev_d(cmp1, cmp0);
+            p_is_pcm_vec = __lsx_vor_v(p_is_pcm_vec, __lsx_vseqi_d(cmp0, 0));
+            DUP2_ARG1(__lsx_vreplgr2vr_d, dq00 + dq30 < tmp, dq04 + dq34 < tmp,
+                      cmp0, cmp1);
+            cmp0 = __lsx_vpackev_d(cmp1, cmp0);
+            q_is_pcm_vec = __lsx_vor_v(q_is_pcm_vec, __lsx_vseqi_d(cmp0, 0));
+
+            tc_pos = __lsx_vsrai_h(tc_pos, 1);
+            tc_neg = __lsx_vneg_h(tc_pos);
+
+            DUP2_ARG2(__lsx_vavgr_hu, p2_src, p0_src, q0_src, q2_src,
+                      delta1, delta2);
+            DUP2_ARG2(__lsx_vsub_h, delta1, p1_src, delta2, q1_src,
+                      delta1, delta2);
+            delta1 = __lsx_vadd_h(delta1, delta0);
+            delta2 = __lsx_vsub_h(delta2, delta0);
+            DUP2_ARG2(__lsx_vsrai_h, delta1, 1, delta2, 1, delta1, delta2);
+            DUP2_ARG3(__lsx_vclip_h, delta1, tc_neg, tc_pos, delta2, tc_neg,
+                      tc_pos, delta1, delta2);
+            DUP2_ARG2(__lsx_vadd_h, p1_src, delta1, q1_src, delta2,
+                      delta1, delta2);
+            DUP2_ARG1(__lsx_vclip255_h, delta1, delta2, delta1, delta2);
+            DUP2_ARG3(__lsx_vbitsel_v, delta1, p1_src, p_is_pcm_vec, delta2,
+                      q1_src, q_is_pcm_vec, delta1, delta2);
+            abs_delta0 = __lsx_vnor_v(abs_delta0, abs_delta0);
+            DUP4_ARG3(__lsx_vbitsel_v, delta1, p1_src, abs_delta0, delta2,
+                      q1_src, abs_delta0, temp0, p0_src, abs_delta0, temp2,
+                      q0_src, abs_delta0, delta1, delta2, temp0, temp2);
+            /* weak filter ends */
+
+            /* pack weak results to 8 bit */
+            DUP2_ARG2(__lsx_vpickev_b, delta1, p2_src, temp2, temp0,
+                      dst3, dst4);
+            dst5 = __lsx_vpickev_b(q2_src, delta2);
+
+            /* select between weak or strong */
+            DUP2_ARG3(__lsx_vbitsel_v, dst0, dst3, cmp2, dst1, dst4, cmp2,
+                      dst0, dst1);
+            dst2 = __lsx_vbitsel_v(dst2, dst5, cmp2);
+
+            /* pack src to 8 bit */
+            DUP2_ARG2(__lsx_vpickev_b, p1_src, p2_src, q0_src, p0_src,
+                      dst3, dst4);
+            dst5 = __lsx_vpickev_b(q2_src, q1_src);
+
+            cmp3 = __lsx_vnor_v(cmp3, cmp3);
+            DUP2_ARG3(__lsx_vbitsel_v, dst0, dst3, cmp3, dst1, dst4, cmp3,
+                      dst0, dst1);
+            dst2 = __lsx_vbitsel_v(dst2, dst5, cmp3);
+
+            __lsx_vstelm_d(dst0, p2, 0, 0);
+            __lsx_vstelm_d(dst0, p2 + stride, 0, 1);
+            __lsx_vstelm_d(dst1, p2 + stride_2x, 0, 0);
+            __lsx_vstelm_d(dst1, p2 + stride_3x, 0, 1);
+            __lsx_vstelm_d(dst2, p2 + stride_4x, 0, 0);
+            __lsx_vstelm_d(dst2, p2 + stride_4x + stride, 0, 1);
+        }
+    }
+}
+
+void ff_hevc_loop_filter_luma_v_8_lsx(uint8_t *src, ptrdiff_t stride,
+                                      int32_t beta, int32_t *tc,
+                                      uint8_t *p_is_pcm, uint8_t *q_is_pcm)
+{
+    ptrdiff_t stride_2x = (stride << 1);
+    ptrdiff_t stride_4x = (stride << 2);
+    ptrdiff_t stride_3x = stride_2x + stride;
+    uint8_t *p3 = src;
+    uint8_t *p2 = src + stride_3x;
+    uint8_t *p1 = src + stride_4x;
+    uint8_t *p0 = src + stride_4x + stride_3x;
+    uint8_t flag0, flag1;
+    int32_t dp00, dq00, dp30, dq30, d00, d30;
+    int32_t d0030, d0434;
+    int32_t dp04, dq04, dp34, dq34, d04, d34;
+    int32_t tc0, p_is_pcm0, q_is_pcm0, beta30, beta20, tc250;
+    int32_t tc4, p_is_pcm4, q_is_pcm4, tc254, tmp;
+
+    __m128i dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+    __m128i cmp0, cmp1, cmp2, p_is_pcm_vec, q_is_pcm_vec;
+    __m128i cmp3;
+    __m128i temp0, temp1;
+    __m128i temp2;
+    __m128i tc_pos, tc_neg;
+    __m128i diff0, diff1, delta0, delta1, delta2, abs_delta0;
+    __m128i zero = {0};
+    __m128i p3_src, p2_src, p1_src, p0_src, q0_src, q1_src, q2_src, q3_src;
+
+    dp00 = abs(p3[-3] - (p3[-2] << 1) + p3[-1]);
+    dq00 = abs(p3[2] - (p3[1] << 1) + p3[0]);
+    dp30 = abs(p2[-3] - (p2[-2] << 1) + p2[-1]);
+    dq30 = abs(p2[2] - (p2[1] << 1) + p2[0]);
+    d00 = dp00 + dq00;
+    d30 = dp30 + dq30;
+    p_is_pcm0 = p_is_pcm[0];
+    q_is_pcm0 = q_is_pcm[0];
+
+    dp04 = abs(p1[-3] - (p1[-2] << 1) + p1[-1]);
+    dq04 = abs(p1[2] - (p1[1] << 1) + p1[0]);
+    dp34 = abs(p0[-3] - (p0[-2] << 1) + p0[-1]);
+    dq34 = abs(p0[2] - (p0[1] << 1) + p0[0]);
+    d04 = dp04 + dq04;
+    d34 = dp34 + dq34;
+    p_is_pcm4 = p_is_pcm[1];
+    q_is_pcm4 = q_is_pcm[1];
+
+    DUP2_ARG1(__lsx_vreplgr2vr_d, p_is_pcm0, p_is_pcm4, cmp0, cmp1);
+    p_is_pcm_vec = __lsx_vpackev_d(cmp1, cmp0);
+    p_is_pcm_vec = __lsx_vseqi_d(p_is_pcm_vec, 0);
+
+    d0030 = (d00 + d30) >= beta;
+    d0434 = (d04 + d34) >= beta;
+
+    DUP2_ARG1(__lsx_vreplgr2vr_d, d0030, d0434, cmp0, cmp1);
+    cmp3 = __lsx_vpackev_d(cmp1, cmp0);
+    cmp3 = __lsx_vseqi_d(cmp3, 0);
+
+    if ((!p_is_pcm0 || !p_is_pcm4 || !q_is_pcm0 || !q_is_pcm4) &&
+        (!d0030 || !d0434)) {
+        src -= 4;
+        DUP4_ARG2(__lsx_vld, src, 0, src + stride, 0, src + stride_2x, 0,
+                  src + stride_3x, 0, p3_src, p2_src, p1_src, p0_src);
+        src += stride_4x;
+        DUP4_ARG2(__lsx_vld, src, 0, src + stride, 0, src + stride_2x, 0,
+                  src + stride_3x, 0, q0_src, q1_src, q2_src, q3_src);
+        src -= stride_4x;
+
+        DUP2_ARG1(__lsx_vreplgr2vr_d, q_is_pcm0, q_is_pcm4, cmp0, cmp1);
+        q_is_pcm_vec = __lsx_vpackev_d(cmp1, cmp0);
+        q_is_pcm_vec = __lsx_vseqi_d(q_is_pcm_vec, 0);
+
+        tc0 = tc[0];
+        beta30 = beta >> 3;
+        beta20 = beta >> 2;
+        tc250 = (((tc0 << 2) + tc0 + 1) >> 1);
+        tc4 = tc[1];
+        tc254 = (((tc4 << 2) + tc4 + 1) >> 1);
+        DUP2_ARG1( __lsx_vreplgr2vr_h, tc0 << 1, tc4 << 1, cmp0, cmp1);
+        tc_pos = __lsx_vpackev_d(cmp1, cmp0);
+        LSX_TRANSPOSE8x8_B(p3_src, p2_src, p1_src, p0_src, q0_src, q1_src,
+                           q2_src, q3_src, p3_src, p2_src, p1_src, p0_src,
+                           q0_src, q1_src, q2_src, q3_src);
+
+        flag0 = abs(p3[-4] - p3[-1]) + abs(p3[3] - p3[0]) < beta30 &&
+                abs(p3[-1] - p3[0]) < tc250;
+        flag0 = flag0 && (abs(p2[-4] - p2[-1]) + abs(p2[3] - p2[0]) < beta30 &&
+                abs(p2[-1] - p2[0]) < tc250 && (d00 << 1) < beta20 &&
+                (d30 << 1) < beta20);
+        cmp0 = __lsx_vreplgr2vr_d(flag0);
+        DUP4_ARG2(__lsx_vilvl_b, zero, p3_src, zero, p2_src, zero, p1_src, zero,
+                  p0_src, p3_src, p2_src, p1_src, p0_src);
+
+        flag1 = abs(p1[-4] - p1[-1]) + abs(p1[3] - p1[0]) < beta30 &&
+                abs(p1[-1] - p1[0]) < tc254;
+        flag1 = flag1 && (abs(p0[-4] - p0[-1]) + abs(p0[3] - p0[0]) < beta30 &&
+                abs(p0[-1] - p0[0]) < tc254 && (d04 << 1) < beta20 &&
+                (d34 << 1) < beta20);
+        DUP4_ARG2(__lsx_vilvl_b, zero, q0_src, zero, q1_src, zero, q2_src, zero,
+                  q3_src, q0_src, q1_src, q2_src, q3_src);
+
+        cmp1 = __lsx_vreplgr2vr_d(flag1);
+        cmp2 = __lsx_vpackev_d(cmp1, cmp0);
+        cmp2 = __lsx_vseqi_d(cmp2, 0);
+
+        if (flag0 && flag1) { /* strong only */
+            /* strong filter */
+            tc_neg = __lsx_vneg_h(tc_pos);
+            /* p part */
+            DUP2_ARG2(__lsx_vadd_h, p1_src, p0_src, temp0, q0_src,
+                      temp0, temp0);
+            temp1 = __lsx_vadd_h(p3_src, p2_src);
+            temp1 = __lsx_vslli_h(temp1, 1);
+            DUP2_ARG2(__lsx_vadd_h, temp1, p2_src, temp1, temp0, temp1, temp1);
+            temp1 = __lsx_vsrari_h(temp1, 3);
+            temp2 = __lsx_vsub_h(temp1, p2_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst0 = __lsx_vadd_h(temp2, p2_src);
+
+            temp1 = __lsx_vadd_h(temp0, p2_src);
+            temp1 = __lsx_vsrari_h(temp1, 2);
+            temp2 = __lsx_vsub_h(temp1, p1_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst1 = __lsx_vadd_h(temp2, p1_src);
+
+            temp1 = __lsx_vslli_h(temp0, 1);
+            DUP2_ARG2(__lsx_vadd_h, temp1, p2_src, temp1, q1_src, temp1, temp1);
+            temp1 = __lsx_vsrari_h(temp1, 3);
+            temp2 = __lsx_vsub_h(temp1, p0_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst2 = __lsx_vadd_h(temp2, p0_src);
+
+            p_is_pcm_vec = __lsx_vnor_v(p_is_pcm_vec, p_is_pcm_vec);
+            DUP2_ARG3(__lsx_vbitsel_v, dst0, p2_src, p_is_pcm_vec, dst1, p1_src,
+                      p_is_pcm_vec, dst0, dst1);
+            dst2 = __lsx_vbitsel_v(dst2, p0_src, p_is_pcm_vec);
+
+            /* q part */
+            DUP2_ARG2(__lsx_vadd_h, q1_src, p0_src, temp0, q0_src,
+                      temp0, temp0);
+            temp1 = __lsx_vadd_h(q3_src, q2_src);
+            temp1 = __lsx_vslli_h(temp1, 1);
+            DUP2_ARG2(__lsx_vadd_h, temp1, q2_src, temp1, temp0, temp1, temp1);
+            temp1 = __lsx_vsrari_h(temp1, 3);
+            temp2 = __lsx_vsub_h(temp1, q2_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst5 = __lsx_vadd_h(temp2, q2_src);
+
+            temp1 = __lsx_vadd_h(temp0, q2_src);
+            temp1 = __lsx_vsrari_h(temp1, 2);
+            temp2 = __lsx_vsub_h(temp1, q1_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst4 = __lsx_vadd_h(temp2, q1_src);
+
+            temp1 = __lsx_vslli_h(temp0, 1);
+            DUP2_ARG2(__lsx_vadd_h, temp1, p1_src, temp1, q2_src, temp1, temp1);
+            temp1 = __lsx_vsrari_h(temp1, 3);
+            temp2 = __lsx_vsub_h(temp1, q0_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst3 = __lsx_vadd_h(temp2, q0_src);
+
+            q_is_pcm_vec = __lsx_vnor_v(q_is_pcm_vec, q_is_pcm_vec);
+            DUP2_ARG3(__lsx_vbitsel_v, dst3, q0_src, q_is_pcm_vec, dst4, q1_src,
+                      q_is_pcm_vec, dst3, dst4);
+            dst5 = __lsx_vbitsel_v(dst5, q2_src, q_is_pcm_vec);
+            /* strong filter ends */
+        } else if (flag0 == flag1) { /* weak only */
+            /* weak filter */
+            tc_pos = __lsx_vsrai_h(tc_pos, 1);
+            tc_neg = __lsx_vneg_h(tc_pos);
+
+            DUP2_ARG2(__lsx_vsub_h, q0_src, p0_src, q1_src, p1_src,
+                      diff0, diff1);
+            DUP2_ARG2(__lsx_vadd_h, __lsx_vslli_h(diff0, 3), diff0,
+                      __lsx_vslli_h(diff1, 1), diff1, diff0, diff1);
+            delta0 = __lsx_vsub_h(diff0, diff1);
+            delta0 = __lsx_vsrari_h(delta0, 4);
+            temp1 = __lsx_vadd_h(__lsx_vslli_h(tc_pos, 3),
+                                 __lsx_vslli_h(tc_pos, 1));
+            abs_delta0 = __lsx_vadda_h(delta0, zero);
+            abs_delta0 = __lsx_vsle_hu(temp1, abs_delta0);
+            abs_delta0 = __lsx_vnor_v(abs_delta0, abs_delta0);
+
+            delta0 = __lsx_vclip_h(delta0, tc_neg, tc_pos);
+            temp2 = __lsx_vadd_h(delta0, p0_src);
+            temp2 = __lsx_vclip255_h(temp2);
+            p_is_pcm_vec = __lsx_vnor_v(p_is_pcm_vec, p_is_pcm_vec);
+            temp0 = __lsx_vbitsel_v(temp2, p0_src, p_is_pcm_vec);
+
+            temp2 = __lsx_vsub_h(q0_src, delta0);
+            temp2 = __lsx_vclip255_h(temp2);
+            q_is_pcm_vec = __lsx_vnor_v(q_is_pcm_vec, q_is_pcm_vec);
+            temp2 = __lsx_vbitsel_v(temp2, q0_src, q_is_pcm_vec);
+
+            tmp = ((beta + (beta >> 1)) >> 3);
+            DUP2_ARG1(__lsx_vreplgr2vr_d, !p_is_pcm0 && ((dp00 + dp30) < tmp),
+                      !p_is_pcm4 && ((dp04 + dp34) < tmp), cmp0, cmp1);
+            p_is_pcm_vec = __lsx_vpackev_d(cmp1, cmp0);
+            p_is_pcm_vec = __lsx_vseqi_d(p_is_pcm_vec, 0);
+
+            DUP2_ARG1(__lsx_vreplgr2vr_h, (!q_is_pcm0) && (dq00 + dq30 < tmp),
+                      (!q_is_pcm4) && (dq04 + dq34 < tmp), cmp0, cmp1);
+            q_is_pcm_vec = __lsx_vpackev_d(cmp1, cmp0);
+            q_is_pcm_vec = __lsx_vseqi_d(q_is_pcm_vec, 0);
+            tc_pos = __lsx_vsrai_h(tc_pos, 1);
+            tc_neg = __lsx_vneg_h(tc_pos);
+
+            DUP2_ARG2(__lsx_vavgr_hu, p2_src, p0_src, q0_src, q2_src,
+                      delta1, delta2);
+            DUP2_ARG2(__lsx_vsub_h, delta1, p1_src, delta2, q1_src,
+                      delta1, delta2);
+            delta1 = __lsx_vadd_h(delta1, delta0);
+            delta2 = __lsx_vsub_h(delta2, delta0);
+            DUP2_ARG2(__lsx_vsrai_h, delta1, 1, delta2, 1, delta1, delta2);
+            DUP2_ARG3(__lsx_vclip_h, delta1, tc_neg, tc_pos, delta2, tc_neg,
+                      tc_pos, delta1, delta2);
+            DUP2_ARG2(__lsx_vadd_h, p1_src, delta1, q1_src, delta2,
+                      delta1, delta2);
+            DUP2_ARG1(__lsx_vclip255_h, delta1, delta2, delta1, delta2);
+            DUP2_ARG3(__lsx_vbitsel_v, delta1, p1_src, p_is_pcm_vec, delta2,
+                      q1_src, q_is_pcm_vec, delta1, delta2);
+
+            abs_delta0 = __lsx_vnor_v(abs_delta0, abs_delta0);
+            DUP4_ARG3(__lsx_vbitsel_v, delta1, p1_src, abs_delta0, temp0,
+                      p0_src, abs_delta0, temp2, q0_src, abs_delta0, delta2,
+                      q1_src, abs_delta0, dst0, dst1, dst2, dst3);
+            /* weak filter ends */
+
+            cmp3 = __lsx_vnor_v(cmp3, cmp3);
+            DUP4_ARG3(__lsx_vbitsel_v, dst0, p1_src, cmp3, dst1, p0_src,
+                      cmp3, dst2, q0_src, cmp3, dst3, q1_src, cmp3,
+                      dst0, dst1, dst2, dst3);
+            DUP2_ARG2(__lsx_vpickev_b, dst2, dst0, dst3, dst1, dst0, dst1);
+
+            /* transpose */
+            dst4 = __lsx_vilvl_b(dst1, dst0);
+            dst5 = __lsx_vilvh_b(dst1, dst0);
+            dst0 = __lsx_vilvl_h(dst5, dst4);
+            dst1 = __lsx_vilvh_h(dst5, dst4);
+
+            src += 2;
+            __lsx_vstelm_w(dst0, src, 0, 0);
+            __lsx_vstelm_w(dst0, src + stride, 0, 1);
+            __lsx_vstelm_w(dst0, src + stride_2x, 0, 2);
+            __lsx_vstelm_w(dst0, src + stride_3x, 0, 3);
+            src += stride_4x;
+            __lsx_vstelm_w(dst1, src, 0, 0);
+            __lsx_vstelm_w(dst1, src + stride, 0, 1);
+            __lsx_vstelm_w(dst1, src + stride_2x, 0, 2);
+            __lsx_vstelm_w(dst1, src + stride_3x, 0, 3);
+            return;
+        } else { /* strong + weak */
+            /* strong filter */
+            tc_neg = __lsx_vneg_h(tc_pos);
+
+            /* p part */
+            DUP2_ARG2(__lsx_vadd_h, p1_src, p0_src, temp0, q0_src,
+                      temp0, temp0);
+
+            temp1 = __lsx_vadd_h(p3_src, p2_src);
+            temp1 = __lsx_vslli_h(temp1, 1);
+            DUP2_ARG2(__lsx_vadd_h, temp1, p2_src, temp1, temp0, temp1, temp1);
+            temp1 = __lsx_vsrari_h(temp1, 3);
+            temp2 = __lsx_vsub_h(temp1, p2_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst0 = __lsx_vadd_h(temp2, p2_src);
+
+            temp1 = __lsx_vadd_h(temp0, p2_src);
+            temp1 = __lsx_vsrari_h(temp1, 2);
+            temp2 = __lsx_vsub_h(temp1, p1_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst1 = __lsx_vadd_h(temp2, p1_src);
+
+            temp1 = __lsx_vslli_h(temp0, 1);
+            DUP2_ARG2(__lsx_vadd_h, temp1, p2_src, temp1, q1_src, temp1, temp1);
+            temp1 = __lsx_vsrari_h(temp1, 3);
+            temp2 = __lsx_vsub_h(temp1, p0_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst2 = __lsx_vadd_h(temp2, p0_src);
+
+            p_is_pcm_vec = __lsx_vnor_v(p_is_pcm_vec, p_is_pcm_vec);
+            DUP2_ARG3(__lsx_vbitsel_v, dst0, p2_src, p_is_pcm_vec, dst1, p1_src,
+                      p_is_pcm_vec, dst0, dst1);
+            dst2 = __lsx_vbitsel_v(dst2, p0_src, p_is_pcm_vec);
+
+            /* q part */
+            DUP2_ARG2(__lsx_vadd_h, q1_src, p0_src, temp0, q0_src, temp0, temp0);
+            temp1 = __lsx_vadd_h(q3_src, q2_src);
+            temp1 = __lsx_vslli_h(temp1, 1);
+            DUP2_ARG2(__lsx_vadd_h, temp1, q2_src, temp1, temp0, temp1, temp1);
+            temp1 = __lsx_vsrari_h(temp1, 3);
+            temp2 = __lsx_vsub_h(temp1, q2_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst5 = __lsx_vadd_h(temp2, q2_src);
+
+            temp1 = __lsx_vadd_h(temp0, q2_src);
+            temp1 = __lsx_vsrari_h(temp1, 2);
+            temp2 = __lsx_vsub_h(temp1, q1_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst4 = __lsx_vadd_h(temp2, q1_src);
+
+            temp1 = __lsx_vslli_h(temp0, 1);
+            DUP2_ARG2(__lsx_vadd_h, temp1, p1_src, temp1, q2_src, temp1, temp1);
+            temp1 = __lsx_vsrari_h(temp1, 3);
+            temp2 = __lsx_vsub_h(temp1, q0_src);
+            temp2 = __lsx_vclip_h(temp2, tc_neg, tc_pos);
+            dst3 = __lsx_vadd_h(temp2, q0_src);
+
+            q_is_pcm_vec = __lsx_vnor_v(q_is_pcm_vec, q_is_pcm_vec);
+            DUP2_ARG3(__lsx_vbitsel_v, dst3, q0_src, q_is_pcm_vec, dst4, q1_src,
+                      q_is_pcm_vec, dst3, dst4);
+            dst5 = __lsx_vbitsel_v(dst5, q2_src, q_is_pcm_vec);
+            /* strong filter ends */
+
+            /* weak filter */
+            tc_pos = __lsx_vsrai_h(tc_pos, 1);
+            tc_neg = __lsx_vneg_h(tc_pos);
+
+            DUP2_ARG2(__lsx_vsub_h, q0_src, p0_src, q1_src, p1_src,
+                      diff0, diff1);
+            DUP2_ARG2(__lsx_vadd_h, __lsx_vslli_h(diff0, 3), diff0,
+                      __lsx_vslli_h(diff1, 1), diff1, diff0, diff1);
+            delta0 = __lsx_vsub_h(diff0, diff1);
+            delta0 = __lsx_vsrari_h(delta0, 4);
+
+            temp1 = __lsx_vadd_h(__lsx_vslli_h(tc_pos, 3),
+                    __lsx_vslli_h(tc_pos, 1));
+            abs_delta0 = __lsx_vadda_h(delta0, zero);
+            abs_delta0 = __lsx_vsle_hu(temp1, abs_delta0);
+            abs_delta0 = __lsx_vnor_v(abs_delta0, abs_delta0);
+            delta0 = __lsx_vclip_h(delta0, tc_neg, tc_pos);
+            temp2 = __lsx_vadd_h(delta0, p0_src);
+            temp2 = __lsx_vclip255_h(temp2);
+            temp0 = __lsx_vbitsel_v(temp2, p0_src, p_is_pcm_vec);
+            temp2 = __lsx_vsub_h(q0_src, delta0);
+            temp2 = __lsx_vclip255_h(temp2);
+            temp2 = __lsx_vbitsel_v(temp2, q0_src, q_is_pcm_vec);
+
+            tmp = (beta + (beta >> 1)) >> 3;
+            DUP2_ARG1(__lsx_vreplgr2vr_d, !p_is_pcm0 && ((dp00 + dp30) < tmp),
+                      !p_is_pcm4 && ((dp04 + dp34) < tmp), cmp0, cmp1);
+            p_is_pcm_vec = __lsx_vpackev_d(cmp1, cmp0);
+            p_is_pcm_vec = __lsx_vseqi_d(p_is_pcm_vec, 0);
+
+            DUP2_ARG1(__lsx_vreplgr2vr_h, (!q_is_pcm0) && (dq00 + dq30 < tmp),
+                      (!q_is_pcm4) && (dq04 + dq34 < tmp), cmp0, cmp1);
+            q_is_pcm_vec = __lsx_vpackev_d(cmp1, cmp0);
+            q_is_pcm_vec = __lsx_vseqi_d(q_is_pcm_vec, 0);
+            tc_pos = __lsx_vsrai_h(tc_pos, 1);
+            tc_neg = __lsx_vneg_h(tc_pos);
+
+            DUP2_ARG2(__lsx_vavgr_hu, p2_src, p0_src, q0_src, q2_src,
+                      delta1, delta2);
+            DUP2_ARG2(__lsx_vsub_h, delta1, p1_src, delta2, q1_src,
+                      delta1, delta2);
+            delta1 = __lsx_vadd_h(delta1, delta0);
+            delta2 = __lsx_vsub_h(delta2, delta0);
+            DUP2_ARG2(__lsx_vsrai_h, delta1, 1, delta2, 1, delta1, delta2);
+            DUP2_ARG3(__lsx_vclip_h, delta1, tc_neg, tc_pos, delta2, tc_neg,
+                      tc_pos, delta1, delta2);
+            DUP2_ARG2(__lsx_vadd_h, p1_src, delta1, q1_src, delta2,
+                      delta1, delta2);
+            DUP2_ARG1(__lsx_vclip255_h, delta1, delta2, delta1, delta2);
+            DUP2_ARG3(__lsx_vbitsel_v, delta1, p1_src, p_is_pcm_vec, delta2,
+                      q1_src, q_is_pcm_vec, delta1, delta2);
+
+            abs_delta0 = __lsx_vnor_v(abs_delta0, abs_delta0);
+            DUP4_ARG3(__lsx_vbitsel_v, delta1, p1_src, abs_delta0, delta2,
+                      q1_src, abs_delta0, temp0, p0_src, abs_delta0, temp2,
+                      q0_src, abs_delta0, delta1, delta2, temp0, temp2);
+            /* weak filter ends*/
+
+            /* select between weak or strong */
+            DUP4_ARG3(__lsx_vbitsel_v, dst0, p2_src, cmp2, dst1, delta1,
+                      cmp2, dst2, temp0, cmp2, dst3, temp2, cmp2,
+                      dst0, dst1, dst2, dst3);
+            DUP2_ARG3(__lsx_vbitsel_v, dst4, delta2, cmp2, dst5, q2_src, cmp2,
+                      dst4, dst5);
+        }
+
+        cmp3 = __lsx_vnor_v(cmp3, cmp3);
+        DUP4_ARG3(__lsx_vbitsel_v, dst0, p2_src, cmp3, dst1, p1_src, cmp3, dst2,
+                  p0_src, cmp3, dst3, q0_src, cmp3, dst0, dst1, dst2, dst3);
+        DUP2_ARG3(__lsx_vbitsel_v, dst4, q1_src, cmp3, dst5, q2_src, cmp3,
+                  dst4, dst5);
+
+        /* pack results to 8 bit */
+        DUP4_ARG2(__lsx_vpickev_b, dst2, dst0, dst3, dst1, dst4, dst4, dst5,
+                  dst5, dst0, dst1, dst2, dst3);
+
+        /* transpose */
+        DUP2_ARG2(__lsx_vilvl_b, dst1, dst0, dst3, dst2, dst4, dst6);
+        DUP2_ARG2(__lsx_vilvh_b, dst1, dst0, dst3, dst2, dst5, dst7);
+        DUP2_ARG2(__lsx_vilvl_h, dst5, dst4, dst7, dst6, dst0, dst2);
+        DUP2_ARG2(__lsx_vilvh_h, dst5, dst4, dst7, dst6, dst1, dst3);
+
+        src += 1;
+        __lsx_vstelm_w(dst0, src, 0, 0);
+        __lsx_vstelm_h(dst2, src, 4, 0);
+        src += stride;
+        __lsx_vstelm_w(dst0, src, 0, 1);
+        __lsx_vstelm_h(dst2, src, 4, 2);
+        src += stride;
+
+        __lsx_vstelm_w(dst0, src, 0, 2);
+        __lsx_vstelm_h(dst2, src, 4, 4);
+        src += stride;
+        __lsx_vstelm_w(dst0, src, 0, 3);
+        __lsx_vstelm_h(dst2, src, 4, 6);
+        src += stride;
+
+        __lsx_vstelm_w(dst1, src, 0, 0);
+        __lsx_vstelm_h(dst3, src, 4, 0);
+        src += stride;
+        __lsx_vstelm_w(dst1, src, 0, 1);
+        __lsx_vstelm_h(dst3, src, 4, 2);
+        src += stride;
+
+        __lsx_vstelm_w(dst1, src, 0, 2);
+        __lsx_vstelm_h(dst3, src, 4, 4);
+        src += stride;
+        __lsx_vstelm_w(dst1, src, 0, 3);
+        __lsx_vstelm_h(dst3, src, 4, 6);
+    }
+}
+
+void ff_hevc_loop_filter_chroma_h_8_lsx(uint8_t *src, ptrdiff_t stride,
+                                        int32_t *tc, uint8_t *p_is_pcm,
+                                        uint8_t *q_is_pcm)
+{
+    uint8_t *p1_ptr = src - (stride << 1);
+    uint8_t *p0_ptr = src - stride;
+    uint8_t *q0_ptr = src;
+    uint8_t *q1_ptr = src + stride;
+    __m128i cmp0, cmp1, p_is_pcm_vec, q_is_pcm_vec;
+    __m128i p1, p0, q0, q1;
+    __m128i tc_pos, tc_neg;
+    __m128i zero = {0};
+    __m128i temp0, temp1, delta;
+
+    if (!(tc[0] <= 0) || !(tc[1] <= 0)) {
+        DUP2_ARG1(__lsx_vreplgr2vr_h, tc[0], tc[1], cmp0, cmp1);
+        tc_pos = __lsx_vpackev_d(cmp1, cmp0);
+        tc_neg = __lsx_vneg_h(tc_pos);
+        DUP2_ARG1(__lsx_vreplgr2vr_d, p_is_pcm[0], p_is_pcm[1], cmp0, cmp1);
+        p_is_pcm_vec = __lsx_vpackev_d(cmp1, cmp0);
+        p_is_pcm_vec = __lsx_vseqi_d(p_is_pcm_vec, 0);
+
+        DUP2_ARG1(__lsx_vreplgr2vr_d, q_is_pcm[0], q_is_pcm[1], cmp0, cmp1);
+        q_is_pcm_vec = __lsx_vpackev_d(cmp1, cmp0);
+        q_is_pcm_vec = __lsx_vseqi_d(q_is_pcm_vec, 0);
+
+        DUP4_ARG2(__lsx_vld, p1_ptr, 0, p0_ptr, 0, q0_ptr, 0, q1_ptr, 0,
+                  p1, p0, q0, q1);
+        DUP4_ARG2(__lsx_vilvl_b, zero, p1, zero, p0, zero, q0, zero, q1,
+                  p1, p0, q0, q1);
+        DUP2_ARG2(__lsx_vsub_h, q0, p0, p1, q1, temp0, temp1);
+        temp0 = __lsx_vslli_h(temp0, 2);
+        temp0 = __lsx_vadd_h(temp0, temp1);
+        delta = __lsx_vsrari_h(temp0, 3);
+        delta = __lsx_vclip_h(delta, tc_neg, tc_pos);
+        temp0 = __lsx_vadd_h(p0, delta);
+        temp0 = __lsx_vclip255_h(temp0);
+        p_is_pcm_vec = __lsx_vnor_v(p_is_pcm_vec, p_is_pcm_vec);
+        temp0 = __lsx_vbitsel_v(temp0, p0, p_is_pcm_vec);
+
+        temp1 = __lsx_vsub_h(q0, delta);
+        temp1 = __lsx_vclip255_h(temp1);
+        q_is_pcm_vec = __lsx_vnor_v(q_is_pcm_vec, q_is_pcm_vec);
+        temp1 = __lsx_vbitsel_v(temp1, q0, q_is_pcm_vec);
+
+        tc_pos = __lsx_vslei_d(tc_pos, 0);
+        DUP2_ARG3(__lsx_vbitsel_v, temp0, p0, tc_pos, temp1, q0, tc_pos,
+                  temp0, temp1);
+        temp0 = __lsx_vpickev_b(temp1, temp0);
+        __lsx_vstelm_d(temp0, p0_ptr, 0, 0);
+        __lsx_vstelm_d(temp0, p0_ptr + stride, 0, 1);
+    }
+}
+
+void ff_hevc_loop_filter_chroma_v_8_lsx(uint8_t *src, ptrdiff_t stride,
+                                        int32_t *tc, uint8_t *p_is_pcm,
+                                        uint8_t *q_is_pcm)
+{
+    ptrdiff_t stride_2x = (stride << 1);
+    ptrdiff_t stride_4x = (stride << 2);
+    ptrdiff_t stride_3x = stride_2x + stride;
+    __m128i cmp0, cmp1, p_is_pcm_vec, q_is_pcm_vec;
+    __m128i src0, src1, src2, src3, src4, src5, src6, src7;
+    __m128i p1, p0, q0, q1;
+    __m128i tc_pos, tc_neg;
+    __m128i zero = {0};
+    __m128i temp0, temp1, delta;
+
+    if (!(tc[0] <= 0) || !(tc[1] <= 0)) {
+        DUP2_ARG1(__lsx_vreplgr2vr_h, tc[0], tc[1], cmp0, cmp1);
+        tc_pos = __lsx_vpackev_d(cmp1, cmp0);
+        tc_neg = __lsx_vneg_h(tc_pos);
+
+        DUP2_ARG1(__lsx_vreplgr2vr_d, p_is_pcm[0], p_is_pcm[1], cmp0, cmp1);
+        p_is_pcm_vec = __lsx_vpackev_d(cmp1, cmp0);
+        p_is_pcm_vec = __lsx_vseqi_d(p_is_pcm_vec, 0);
+        DUP2_ARG1(__lsx_vreplgr2vr_d, q_is_pcm[0], q_is_pcm[1], cmp0, cmp1);
+        q_is_pcm_vec = __lsx_vpackev_d(cmp1, cmp0);
+        q_is_pcm_vec = __lsx_vseqi_d(q_is_pcm_vec, 0);
+
+        src -= 2;
+        DUP4_ARG2(__lsx_vld, src, 0, src + stride, 0, src + stride_2x, 0,
+                  src + stride_3x, 0, src0, src1, src2, src3);
+        src += stride_4x;
+        DUP4_ARG2(__lsx_vld, src, 0, src + stride, 0, src + stride_2x, 0,
+                  src + stride_3x, 0, src4, src5, src6, src7);
+        src -= stride_4x;
+        LSX_TRANSPOSE8x4_B(src0, src1, src2, src3, src4, src5, src6, src7,
+                           p1, p0, q0, q1);
+        DUP4_ARG2(__lsx_vilvl_b, zero, p1, zero, p0, zero, q0, zero, q1,
+                  p1, p0, q0, q1);
+
+        DUP2_ARG2(__lsx_vsub_h, q0, p0, p1, q1, temp0, temp1);
+        temp0 = __lsx_vslli_h(temp0, 2);
+        temp0 = __lsx_vadd_h(temp0, temp1);
+        delta = __lsx_vsrari_h(temp0, 3);
+        delta = __lsx_vclip_h(delta, tc_neg, tc_pos);
+
+        temp0 = __lsx_vadd_h(p0, delta);
+        temp1 = __lsx_vsub_h(q0, delta);
+        DUP2_ARG1(__lsx_vclip255_h, temp0, temp1, temp0, temp1);
+        DUP2_ARG2(__lsx_vnor_v, p_is_pcm_vec, p_is_pcm_vec, q_is_pcm_vec,
+                  q_is_pcm_vec, p_is_pcm_vec, q_is_pcm_vec);
+        DUP2_ARG3(__lsx_vbitsel_v, temp0, p0, p_is_pcm_vec, temp1, q0,
+                  q_is_pcm_vec, temp0, temp1);
+
+        tc_pos = __lsx_vslei_d(tc_pos, 0);
+        DUP2_ARG3(__lsx_vbitsel_v, temp0, p0, tc_pos, temp1, q0, tc_pos,
+                  temp0, temp1);
+        temp0 = __lsx_vpackev_b(temp1, temp0);
+
+        src += 1;
+        __lsx_vstelm_h(temp0, src, 0, 0);
+        __lsx_vstelm_h(temp0, src + stride, 0, 1);
+        __lsx_vstelm_h(temp0, src + stride_2x, 0, 2);
+        __lsx_vstelm_h(temp0, src + stride_3x, 0, 3);
+        src += stride_4x;
+        __lsx_vstelm_h(temp0, src, 0, 4);
+        __lsx_vstelm_h(temp0, src + stride, 0, 5);
+        __lsx_vstelm_h(temp0, src + stride_2x, 0, 6);
+        __lsx_vstelm_h(temp0, src + stride_3x, 0, 7);
+        src -= stride_4x;
+    }
+}
+
+static void hevc_sao_edge_filter_0degree_4width_lsx(uint8_t *dst,
+                                                    int32_t dst_stride,
+                                                    uint8_t *src,
+                                                    int32_t src_stride,
+                                                    int16_t *sao_offset_val,
+                                                    int32_t height)
+{
+    const int32_t src_stride_2x = (src_stride << 1);
+    const int32_t dst_stride_2x = (dst_stride << 1);
+    __m128i shuf1 = {0x807060504030201, 0x100F0E0D0C0B0A09};
+    __m128i shuf2 = {0x908070605040302, 0x11100F0E0D0C0B0A};
+    __m128i edge_idx = {0x403000201, 0x0};
+    __m128i cmp_minus10, cmp_minus11, diff_minus10, diff_minus11;
+    __m128i sao_offset = __lsx_vld(sao_offset_val, 0);
+    __m128i src_minus10, src_minus11, src_plus10, offset, src0, dst0;
+    __m128i const1 = __lsx_vldi(1);
+    __m128i zero = {0};
+
+    sao_offset = __lsx_vpickev_b(sao_offset, sao_offset);
+    src -= 1;
+
+    /* load in advance */
+    DUP2_ARG2(__lsx_vld, src, 0, src + src_stride, 0, src_minus10, src_minus11);
+
+    for (height -= 2; height; height -= 2) {
+        src += src_stride_2x;
+        src_minus10 = __lsx_vpickev_d(src_minus11, src_minus10);
+        src0 = __lsx_vshuf_b(zero, src_minus10, shuf1);
+        src_plus10 = __lsx_vshuf_b(zero, src_minus10, shuf2);
+
+        DUP2_ARG2(__lsx_vseq_b, src0, src_minus10, src0, src_plus10,
+                  cmp_minus10, cmp_minus11);
+        DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+                  cmp_minus11, diff_minus10, diff_minus11);
+        DUP2_ARG2(__lsx_vsle_bu, src0, src_minus10, src0, src_plus10,
+                  cmp_minus10, cmp_minus11);
+        DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+                  cmp_minus11, cmp_minus10, cmp_minus11);
+        DUP2_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10,
+        diff_minus11, const1, cmp_minus11, diff_minus10, diff_minus11);
+
+        offset = __lsx_vadd_b(diff_minus10, diff_minus11);
+        offset = __lsx_vaddi_bu(offset, 2);
+
+        /* load in advance */
+        DUP2_ARG2(__lsx_vld, src, 0, src + src_stride, 0,
+                  src_minus10, src_minus11);
+        DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset,
+                  sao_offset, sao_offset, offset, offset, offset);
+        src0 = __lsx_vxori_b(src0, 128);
+        dst0 = __lsx_vsadd_b(src0, offset);
+        dst0 = __lsx_vxori_b(dst0, 128);
+
+        __lsx_vstelm_w(dst0, dst, 0, 0);
+        __lsx_vstelm_w(dst0, dst + dst_stride, 0, 2);
+        dst += dst_stride_2x;
+    }
+
+    src_minus10 = __lsx_vpickev_d(src_minus11, src_minus10);
+    src0 = __lsx_vshuf_b(zero, src_minus10, shuf1);
+    src_plus10 = __lsx_vshuf_b(zero, src_minus10, shuf2);
+
+    DUP2_ARG2(__lsx_vseq_b, src0, src_minus10, src0, src_plus10, cmp_minus10,
+              cmp_minus11);
+    DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11, cmp_minus11,
+              diff_minus10, diff_minus11);
+    DUP2_ARG2(__lsx_vsle_bu, src0, src_minus10, src0, src_plus10, cmp_minus10,
+              cmp_minus11);
+    DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11, cmp_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10, diff_minus11,
+              const1, cmp_minus11, diff_minus10, diff_minus11);
+
+    offset = __lsx_vadd_b(diff_minus10, diff_minus11);
+    offset = __lsx_vaddi_bu(offset, 2);
+    DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset, sao_offset, sao_offset,
+              offset, offset, offset);
+    src0 = __lsx_vxori_b(src0, 128);
+    dst0 = __lsx_vsadd_b(src0, offset);
+    dst0 = __lsx_vxori_b(dst0, 128);
+
+    __lsx_vstelm_w(dst0, dst, 0, 0);
+    __lsx_vstelm_w(dst0, dst + dst_stride, 0, 2);
+}
+
+static void hevc_sao_edge_filter_0degree_8width_lsx(uint8_t *dst,
+                                                    int32_t dst_stride,
+                                                    uint8_t *src,
+                                                    int32_t src_stride,
+                                                    int16_t *sao_offset_val,
+                                                    int32_t height)
+{
+    const int32_t src_stride_2x = (src_stride << 1);
+    const int32_t dst_stride_2x = (dst_stride << 1);
+    __m128i shuf1 = {0x807060504030201, 0x100F0E0D0C0B0A09};
+    __m128i shuf2 = {0x908070605040302, 0x11100F0E0D0C0B0A};
+    __m128i edge_idx = {0x403000201, 0x0};
+    __m128i const1 = __lsx_vldi(1);
+    __m128i cmp_minus10, cmp_minus11, diff_minus10, diff_minus11;
+    __m128i src0, src1, dst0, src_minus10, src_minus11, src_plus10, src_plus11;
+    __m128i offset, sao_offset = __lsx_vld(sao_offset_val, 0);
+    __m128i zeros = {0};
+
+    sao_offset = __lsx_vpickev_b(sao_offset, sao_offset);
+    src -= 1;
+
+    /* load in advance */
+    DUP2_ARG2(__lsx_vld, src, 0, src + src_stride, 0, src_minus10, src_minus11);
+
+    for (height -= 2; height; height -= 2) {
+        src += src_stride_2x;
+        DUP2_ARG3(__lsx_vshuf_b, zeros, src_minus10, shuf1, zeros,
+                  src_minus11, shuf1, src0, src1);
+        DUP2_ARG3(__lsx_vshuf_b, zeros, src_minus10, shuf2, zeros,
+                  src_minus11, shuf2, src_plus10, src_plus11);
+        DUP2_ARG2(__lsx_vpickev_d, src_minus11, src_minus10, src_plus11,
+                  src_plus10, src_minus10, src_plus10);
+        src0 = __lsx_vpickev_d(src1, src0);
+
+        DUP2_ARG2(__lsx_vseq_b, src0, src_minus10, src0, src_plus10,
+                  cmp_minus10, cmp_minus11);
+        DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+                  cmp_minus11, diff_minus10, diff_minus11);
+        DUP2_ARG2(__lsx_vsle_bu, src0, src_minus10, src0, src_plus10,
+                  cmp_minus10, cmp_minus11);
+        DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+                  cmp_minus11, cmp_minus10, cmp_minus11);
+        DUP2_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10,
+        diff_minus11, const1, cmp_minus11, diff_minus10, diff_minus11);
+
+        offset = __lsx_vadd_b(diff_minus10, diff_minus11);
+        offset = __lsx_vaddi_bu(offset, 2);
+
+        /* load in advance */
+        DUP2_ARG2(__lsx_vld, src, 0, src + src_stride, 0,
+                  src_minus10, src_minus11);
+        DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset, sao_offset,
+                  sao_offset, offset, offset, offset);
+        src0 = __lsx_vxori_b(src0, 128);
+        dst0 = __lsx_vsadd_b(src0, offset);
+        dst0 = __lsx_vxori_b(dst0, 128);
+
+        __lsx_vstelm_d(dst0, dst, 0, 0);
+        __lsx_vstelm_d(dst0, dst + dst_stride, 0, 1);
+        dst += dst_stride_2x;
+    }
+
+    DUP2_ARG3(__lsx_vshuf_b, zeros, src_minus10, shuf1, zeros, src_minus11,
+              shuf1, src0, src1);
+    DUP2_ARG3(__lsx_vshuf_b, zeros, src_minus10, shuf2, zeros, src_minus11,
+              shuf2, src_plus10, src_plus11);
+    DUP2_ARG2(__lsx_vpickev_d, src_minus11, src_minus10, src_plus11,
+              src_plus10, src_minus10, src_plus10);
+    src0 =  __lsx_vpickev_d(src1, src0);
+
+    DUP2_ARG2(__lsx_vseq_b, src0, src_minus10, src0, src_plus10, cmp_minus10,
+              cmp_minus11);
+    DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11, cmp_minus11,
+              diff_minus10, diff_minus11);
+    DUP2_ARG2(__lsx_vsle_bu, src0, src_minus10, src0, src_plus10, cmp_minus10,
+              cmp_minus11);
+    DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11, cmp_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10, diff_minus11,
+              const1, cmp_minus11, diff_minus10, diff_minus11);
+
+    offset = __lsx_vadd_b(diff_minus10, diff_minus11);
+    offset = __lsx_vaddi_bu(offset, 2);
+    DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset, sao_offset,
+              sao_offset, offset, offset, offset);
+    src0 = __lsx_vxori_b(src0, 128);
+    dst0 = __lsx_vsadd_b(src0, offset);
+    dst0 = __lsx_vxori_b(dst0, 128);
+
+    __lsx_vstelm_d(dst0, dst, 0, 0);
+    __lsx_vstelm_d(dst0, dst + dst_stride, 0, 1);
+}
+
+static void hevc_sao_edge_filter_0degree_16multiple_lsx(uint8_t *dst,
+                                                        int32_t dst_stride,
+                                                        uint8_t *src,
+                                                        int32_t src_stride,
+                                                        int16_t *sao_offset_val,
+                                                        int32_t width,
+                                                        int32_t height)
+{
+    uint8_t *dst_ptr, *src_minus1;
+    int32_t v_cnt;
+    const int32_t src_stride_2x = (src_stride << 1);
+    const int32_t dst_stride_2x = (dst_stride << 1);
+    const int32_t src_stride_4x = (src_stride << 2);
+    const int32_t dst_stride_4x = (dst_stride << 2);
+    const int32_t src_stride_3x = src_stride_2x + src_stride;
+    const int32_t dst_stride_3x = dst_stride_2x + dst_stride;
+
+    __m128i shuf1 = {0x807060504030201, 0x100F0E0D0C0B0A09};
+    __m128i shuf2 = {0x908070605040302, 0x11100F0E0D0C0B0A};
+    __m128i edge_idx = {0x403000201, 0x0};
+    __m128i const1 = __lsx_vldi(1);
+    __m128i sao_offset;
+    __m128i cmp_minus10, cmp_plus10, diff_minus10, diff_plus10, cmp_minus11;
+    __m128i cmp_plus11, diff_minus11, diff_plus11, cmp_minus12, cmp_plus12;
+    __m128i diff_minus12, diff_plus12, cmp_minus13, cmp_plus13, diff_minus13;
+    __m128i diff_plus13;
+    __m128i src10, src11, src12, src13, dst0, dst1, dst2, dst3;
+    __m128i src_minus10, src_minus11, src_minus12, src_minus13;
+    __m128i offset_mask0, offset_mask1, offset_mask2, offset_mask3;
+    __m128i src_zero0, src_zero1, src_zero2, src_zero3;
+    __m128i src_plus10, src_plus11, src_plus12, src_plus13;
+
+    sao_offset = __lsx_vld(sao_offset_val, 0);
+    sao_offset = __lsx_vpickev_b(sao_offset, sao_offset);
+
+    for (; height; height -= 4) {
+        src_minus1 = src - 1;
+        src_minus10 = __lsx_vld(src_minus1, 0);
+        DUP2_ARG2(__lsx_vldx, src_minus1, src_stride, src_minus1,
+                  src_stride_2x, src_minus11, src_minus12);
+        src_minus13 = __lsx_vldx(src_minus1, src_stride_3x);
+
+        for (v_cnt = 0; v_cnt < width; v_cnt += 16) {
+            src_minus1 += 16;
+            dst_ptr = dst + v_cnt;
+            src10 = __lsx_vld(src_minus1, 0);
+            DUP2_ARG2(__lsx_vldx, src_minus1, src_stride, src_minus1,
+                      src_stride_2x, src11, src12);
+            src13 = __lsx_vldx(src_minus1, src_stride_3x);
+            DUP4_ARG3(__lsx_vshuf_b, src10, src_minus10, shuf1, src11,
+                      src_minus11, shuf1, src12, src_minus12, shuf1, src13,
+                      src_minus13, shuf1, src_zero0, src_zero1,
+                      src_zero2, src_zero3);
+            DUP4_ARG3(__lsx_vshuf_b, src10, src_minus10, shuf2, src11,
+                      src_minus11, shuf2, src12, src_minus12, shuf2, src13,
+                      src_minus13, shuf2, src_plus10, src_plus11,
+                      src_plus12, src_plus13);
+            DUP4_ARG2(__lsx_vseq_b, src_zero0, src_minus10, src_zero0,
+                      src_plus10, src_zero1, src_minus11, src_zero1, src_plus11,
+                      cmp_minus10, cmp_plus10, cmp_minus11, cmp_plus11);
+            DUP4_ARG2(__lsx_vseq_b, src_zero2, src_minus12, src_zero2,
+                      src_plus12, src_zero3, src_minus13, src_zero3, src_plus13,
+                      cmp_minus12, cmp_plus12, cmp_minus13, cmp_plus13);
+            DUP4_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_plus10,
+                      cmp_plus10, cmp_minus11, cmp_minus11, cmp_plus11,
+                      cmp_plus11, diff_minus10, diff_plus10, diff_minus11,
+                      diff_plus11);
+            DUP4_ARG2(__lsx_vnor_v, cmp_minus12, cmp_minus12, cmp_plus12,
+                      cmp_plus12, cmp_minus13, cmp_minus13, cmp_plus13,
+                      cmp_plus13, diff_minus12, diff_plus12, diff_minus13,
+                      diff_plus13);
+            DUP4_ARG2(__lsx_vsle_bu, src_zero0, src_minus10, src_zero0,
+                      src_plus10, src_zero1, src_minus11, src_zero1, src_plus11,
+                      cmp_minus10, cmp_plus10, cmp_minus11, cmp_plus11);
+            DUP4_ARG2(__lsx_vsle_bu, src_zero2, src_minus12, src_zero2,
+                      src_plus12, src_zero3, src_minus13, src_zero3, src_plus13,
+                      cmp_minus12, cmp_plus12, cmp_minus13, cmp_plus13);
+            DUP4_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_plus10,
+                      cmp_plus10, cmp_minus11, cmp_minus11, cmp_plus11,
+                      cmp_plus11, cmp_minus10, cmp_plus10, cmp_minus11,
+                      cmp_plus11);
+            DUP4_ARG2(__lsx_vnor_v, cmp_minus12, cmp_minus12, cmp_plus12,
+                      cmp_plus12, cmp_minus13, cmp_minus13, cmp_plus13,
+                      cmp_plus13, cmp_minus12, cmp_plus12, cmp_minus13,
+                      cmp_plus13);
+            DUP4_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10,
+                      diff_plus10, const1, cmp_plus10, diff_minus11, const1,
+                      cmp_minus11, diff_plus11, const1, cmp_plus11,
+                      diff_minus10, diff_plus10, diff_minus11, diff_plus11);
+            DUP4_ARG3(__lsx_vbitsel_v, diff_minus12, const1, cmp_minus12,
+                      diff_plus12, const1, cmp_plus12, diff_minus13, const1,
+                      cmp_minus13, diff_plus13, const1, cmp_plus13,
+                      diff_minus12, diff_plus12, diff_minus13, diff_plus13);
+
+            DUP4_ARG2(__lsx_vadd_b, diff_minus10, diff_plus10, diff_minus11,
+                      diff_plus11, diff_minus12, diff_plus12, diff_minus13,
+                      diff_plus13, offset_mask0, offset_mask1, offset_mask2,
+                      offset_mask3);
+            DUP4_ARG2(__lsx_vaddi_bu, offset_mask0, 2, offset_mask1, 2,
+                      offset_mask2, 2, offset_mask3, 2, offset_mask0,
+                      offset_mask1, offset_mask2, offset_mask3);
+            DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset_mask0,
+                      sao_offset, sao_offset, offset_mask0, offset_mask0,
+                      offset_mask0);
+            DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset_mask1,
+                      sao_offset, sao_offset, offset_mask1, offset_mask1,
+                      offset_mask1);
+            DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset_mask2,
+                      sao_offset, sao_offset, offset_mask2, offset_mask2,
+                      offset_mask2);
+            DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset_mask3,
+                      sao_offset, sao_offset, offset_mask3, offset_mask3,
+                      offset_mask3);
+
+            DUP4_ARG2(__lsx_vxori_b, src_zero0, 128, src_zero1, 128,
+                      src_zero2, 128, src_zero3, 128, src_zero0, src_zero1,
+                      src_zero2, src_zero3);
+            DUP4_ARG2(__lsx_vsadd_b, src_zero0, offset_mask0, src_zero1,
+                      offset_mask1, src_zero2, offset_mask2, src_zero3,
+                      offset_mask3, dst0, dst1, dst2, dst3);
+            DUP4_ARG2(__lsx_vxori_b, dst0, 128, dst1, 128, dst2, 128, dst3,
+                      128, dst0, dst1, dst2, dst3);
+
+            src_minus10 = src10;
+            src_minus11 = src11;
+            src_minus12 = src12;
+            src_minus13 = src13;
+
+            __lsx_vst(dst0, dst_ptr, 0);
+            __lsx_vst(dst1, dst_ptr + dst_stride, 0);
+            __lsx_vst(dst2, dst_ptr + dst_stride_2x, 0);
+            __lsx_vst(dst3, dst_ptr + dst_stride_3x, 0);
+        }
+        src += src_stride_4x;
+        dst += dst_stride_4x;
+    }
+}
+
+static void hevc_sao_edge_filter_90degree_4width_lsx(uint8_t *dst,
+                                                     int32_t dst_stride,
+                                                     uint8_t *src,
+                                                     int32_t src_stride,
+                                                     int16_t *sao_offset_val,
+                                                     int32_t height)
+{
+    const int32_t src_stride_2x = (src_stride << 1);
+    const int32_t dst_stride_2x = (dst_stride << 1);
+    __m128i edge_idx = {0x403000201, 0x0};
+    __m128i const1 = __lsx_vldi(1);
+    __m128i dst0;
+    __m128i sao_offset = __lsx_vld(sao_offset_val, 0);
+    __m128i cmp_minus10, diff_minus10, cmp_minus11, diff_minus11;
+    __m128i src_minus10, src_minus11, src10, src11;
+    __m128i src_zero0, src_zero1;
+    __m128i offset;
+    __m128i offset_mask0, offset_mask1;
+
+    sao_offset = __lsx_vpickev_b(sao_offset, sao_offset);
+
+    /* load in advance */
+    DUP4_ARG2(__lsx_vld, src - src_stride, 0, src, 0, src + src_stride, 0,
+              src + src_stride_2x, 0, src_minus10, src_minus11, src10, src11);
+
+    for (height -= 2; height; height -= 2) {
+        src += src_stride_2x;
+        DUP4_ARG2(__lsx_vilvl_b, src10, src_minus10, src_minus11, src_minus11,
+                  src11, src_minus11, src10, src10, src_minus10, src_zero0,
+                  src_minus11, src_zero1);
+        DUP2_ARG2(__lsx_vseq_b, src_zero0, src_minus10, src_zero1, src_minus11,
+                  cmp_minus10, cmp_minus11);
+        DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+                  cmp_minus11, diff_minus10, diff_minus11);
+        DUP2_ARG2(__lsx_vsle_bu, src_zero0, src_minus10, src_zero1,
+                  src_minus11, cmp_minus10, cmp_minus11);
+        DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+                  cmp_minus11, cmp_minus10, cmp_minus11);
+        DUP2_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10,
+                 diff_minus11, const1, cmp_minus11, diff_minus10, diff_minus11);
+
+        DUP2_ARG2(__lsx_vhaddw_hu_bu, diff_minus10, diff_minus10, diff_minus11,
+                  diff_minus11, offset_mask0, offset_mask1);
+        DUP2_ARG2(__lsx_vaddi_hu, offset_mask0, 2, offset_mask1, 2,
+                  offset_mask0, offset_mask1);
+        DUP2_ARG2(__lsx_vpickev_b, offset_mask1, offset_mask0, src_zero1,
+                  src_zero0, offset, dst0);
+        DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset, sao_offset,
+                  sao_offset, offset, offset, offset);
+
+        dst0 = __lsx_vxori_b(dst0, 128);
+        dst0 = __lsx_vsadd_b(dst0, offset);
+        dst0 = __lsx_vxori_b(dst0, 128);
+        src_minus10 = src10;
+        src_minus11 = src11;
+
+        /* load in advance */
+        DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride_2x,
+                  src10, src11);
+
+        __lsx_vstelm_w(dst0, dst, 0, 0);
+        __lsx_vstelm_w(dst0, dst + dst_stride, 0, 2);
+        dst += dst_stride_2x;
+    }
+
+    DUP4_ARG2(__lsx_vilvl_b, src10, src_minus10, src_minus11, src_minus11,
+              src11,  src_minus11, src10, src10, src_minus10, src_zero0,
+              src_minus11, src_zero1);
+    DUP2_ARG2(__lsx_vseq_b, src_zero0, src_minus10, src_zero1, src_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11, cmp_minus11,
+              diff_minus10, diff_minus11);
+    DUP2_ARG2(__lsx_vsle_bu, src_zero0, src_minus10, src_zero1, src_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11, cmp_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10, diff_minus11,
+              const1, cmp_minus11, diff_minus10, diff_minus11);
+
+    DUP2_ARG2(__lsx_vhaddw_hu_bu, diff_minus10, diff_minus10, diff_minus11,
+              diff_minus11, offset_mask0, offset_mask1);
+    DUP2_ARG2(__lsx_vaddi_bu, offset_mask0, 2, offset_mask1, 2,
+              offset_mask0, offset_mask1);
+    DUP2_ARG2(__lsx_vpickev_b, offset_mask1, offset_mask0, src_zero1,
+              src_zero0, offset, dst0);
+    DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset, sao_offset,
+              sao_offset, offset, offset, offset);
+    dst0 = __lsx_vxori_b(dst0, 128);
+    dst0 = __lsx_vsadd_b(dst0, offset);
+    dst0 = __lsx_vxori_b(dst0, 128);
+
+    __lsx_vstelm_w(dst0, dst, 0, 0);
+    __lsx_vstelm_w(dst0, dst + dst_stride, 0, 2);
+}
+
+static void hevc_sao_edge_filter_90degree_8width_lsx(uint8_t *dst,
+                                                     int32_t dst_stride,
+                                                     uint8_t *src,
+                                                     int32_t src_stride,
+                                                     int16_t *sao_offset_val,
+                                                     int32_t height)
+{
+    const int32_t src_stride_2x = (src_stride << 1);
+    const int32_t dst_stride_2x = (dst_stride << 1);
+    __m128i edge_idx = {0x403000201, 0x0};
+    __m128i const1 = __lsx_vldi(1);
+    __m128i offset, sao_offset = __lsx_vld(sao_offset_val, 0);
+    __m128i src_zero0, src_zero1, dst0;
+    __m128i cmp_minus10, diff_minus10, cmp_minus11, diff_minus11;
+    __m128i src_minus10, src_minus11, src10, src11;
+    __m128i offset_mask0, offset_mask1;
+
+    sao_offset = __lsx_vpickev_b(sao_offset, sao_offset);
+
+    /* load in advance */
+    DUP2_ARG2(__lsx_vld, src - src_stride, 0, src, 0, src_minus10, src_minus11);
+    DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride_2x, src10, src11);
+
+    for (height -= 2; height; height -= 2) {
+        src += src_stride_2x;
+        DUP4_ARG2(__lsx_vilvl_b, src10, src_minus10, src_minus11, src_minus11,
+                  src11, src_minus11, src10, src10, src_minus10, src_zero0,
+                  src_minus11, src_zero1);
+        DUP2_ARG2(__lsx_vseq_b, src_zero0, src_minus10, src_zero1, src_minus11,
+                  cmp_minus10, cmp_minus11);
+        DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+                  cmp_minus11, diff_minus10, diff_minus11);
+        DUP2_ARG2(__lsx_vsle_bu, src_zero0, src_minus10, src_zero1,
+                  src_minus11, cmp_minus10, cmp_minus11);
+        DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+                  cmp_minus11, cmp_minus10, cmp_minus11);
+        DUP2_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10,
+                diff_minus11, const1, cmp_minus11, diff_minus10, diff_minus11);
+
+        DUP2_ARG2(__lsx_vhaddw_hu_bu, diff_minus10, diff_minus10, diff_minus11,
+                  diff_minus11, offset_mask0, offset_mask1);
+        DUP2_ARG2(__lsx_vaddi_hu, offset_mask0, 2, offset_mask1, 2,
+                  offset_mask0, offset_mask1);
+        DUP2_ARG2(__lsx_vpickev_b, offset_mask1, offset_mask0, src_zero1,
+                  src_zero0, offset, dst0);
+        DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset, sao_offset,
+                  sao_offset, offset, offset, offset);
+
+        dst0 = __lsx_vxori_b(dst0, 128);
+        dst0 = __lsx_vsadd_b(dst0, offset);
+        dst0 = __lsx_vxori_b(dst0, 128);
+        src_minus10 = src10;
+        src_minus11 = src11;
+
+        /* load in advance */
+        DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride_2x,
+                  src10, src11);
+
+        __lsx_vstelm_d(dst0, dst, 0, 0);
+        __lsx_vstelm_d(dst0, dst + dst_stride, 0, 1);
+        dst += dst_stride_2x;
+    }
+
+    DUP4_ARG2(__lsx_vilvl_b, src10, src_minus10, src_minus11, src_minus11,
+              src11, src_minus11, src10, src10, src_minus10, src_zero0,
+              src_minus11, src_zero1);
+    DUP2_ARG2(__lsx_vseq_b, src_zero0, src_minus10, src_zero1, src_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11, cmp_minus11,
+              diff_minus10, diff_minus11);
+    DUP2_ARG2(__lsx_vsle_bu, src_zero0, src_minus10, src_zero1, src_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11, cmp_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10, diff_minus11,
+              const1, cmp_minus11, diff_minus10, diff_minus11);
+
+    DUP2_ARG2(__lsx_vhaddw_hu_bu, diff_minus10, diff_minus10, diff_minus11,
+              diff_minus11, offset_mask0, offset_mask1);
+    DUP2_ARG2(__lsx_vaddi_hu, offset_mask0, 2, offset_mask1, 2,
+              offset_mask0, offset_mask1);
+    DUP2_ARG2(__lsx_vpickev_b, offset_mask1, offset_mask0, src_zero1,
+              src_zero0, offset, dst0);
+    DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset, sao_offset,
+              sao_offset, offset, offset, offset);
+    dst0 =  __lsx_vxori_b(dst0, 128);
+    dst0 = __lsx_vsadd_b(dst0, offset);
+    dst0 = __lsx_vxori_b(dst0, 128);
+
+    __lsx_vstelm_d(dst0, dst, 0, 0);
+    __lsx_vstelm_d(dst0, dst + dst_stride, 0, 1);
+}
+
+static void hevc_sao_edge_filter_90degree_16multiple_lsx(uint8_t *dst,
+                                                         int32_t dst_stride,
+                                                         uint8_t *src,
+                                                         int32_t src_stride,
+                                                         int16_t *
+                                                         sao_offset_val,
+                                                         int32_t width,
+                                                         int32_t height)
+{
+    uint8_t *src_orig = src;
+    uint8_t *dst_orig = dst;
+    int32_t h_cnt, v_cnt;
+    const int32_t src_stride_2x = (src_stride << 1);
+    const int32_t dst_stride_2x = (dst_stride << 1);
+    const int32_t src_stride_4x = (src_stride << 2);
+    const int32_t dst_stride_4x = (dst_stride << 2);
+    const int32_t src_stride_3x = src_stride_2x + src_stride;
+    const int32_t dst_stride_3x = dst_stride_2x + dst_stride;
+    __m128i edge_idx = {0x403000201, 0x0};
+    __m128i const1 = __lsx_vldi(1);
+    __m128i cmp_minus10, cmp_plus10, diff_minus10, diff_plus10, cmp_minus11;
+    __m128i cmp_plus11, diff_minus11, diff_plus11, cmp_minus12, cmp_plus12;
+    __m128i diff_minus12, diff_plus12, cmp_minus13, cmp_plus13, diff_minus13;
+    __m128i diff_plus13;
+    __m128i src10, src_minus10, dst0, src11, src_minus11, dst1;
+    __m128i src12, dst2, src13, dst3;
+    __m128i offset_mask0, offset_mask1, offset_mask2, offset_mask3, sao_offset;
+
+    sao_offset = __lsx_vld(sao_offset_val, 0);
+    sao_offset = __lsx_vpickev_b(sao_offset, sao_offset);
+
+    for (v_cnt = 0; v_cnt < width; v_cnt += 16) {
+        src = src_orig + v_cnt;
+        dst = dst_orig + v_cnt;
+
+        DUP2_ARG2(__lsx_vld, src - src_stride, 0, src, 0,
+                  src_minus10, src_minus11);
+
+        for (h_cnt = (height >> 2); h_cnt--;) {
+            DUP4_ARG2(__lsx_vldx, src, src_stride, src, src_stride_2x,
+                      src, src_stride_3x, src, src_stride_4x,
+                      src10, src11, src12, src13);
+            DUP4_ARG2(__lsx_vseq_b, src_minus11, src_minus10, src_minus11,
+                      src10, src10, src_minus11, src10, src11, cmp_minus10,
+                      cmp_plus10, cmp_minus11, cmp_plus11);
+            DUP4_ARG2(__lsx_vseq_b, src11, src10, src11, src12, src12, src11,
+                      src12, src13, cmp_minus12, cmp_plus12,
+                      cmp_minus13, cmp_plus13);
+            DUP4_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_plus10,
+                      cmp_plus10, cmp_minus11, cmp_minus11, cmp_plus11,
+                      cmp_plus11, diff_minus10, diff_plus10, diff_minus11,
+                      diff_plus11);
+            DUP4_ARG2(__lsx_vnor_v, cmp_minus12, cmp_minus12, cmp_plus12,
+                      cmp_plus12, cmp_minus13, cmp_minus13, cmp_plus13,
+                      cmp_plus13, diff_minus12, diff_plus12, diff_minus13,
+                      diff_plus13);
+            DUP4_ARG2(__lsx_vsle_bu, src_minus11, src_minus10, src_minus11,
+                      src10, src10, src_minus11, src10, src11, cmp_minus10,
+                      cmp_plus10, cmp_minus11, cmp_plus11);
+            DUP4_ARG2(__lsx_vsle_bu, src11, src10, src11, src12, src12, src11,
+                      src12, src13, cmp_minus12, cmp_plus12, cmp_minus13,
+                      cmp_plus13);
+            DUP4_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_plus10,
+                      cmp_plus10, cmp_minus11, cmp_minus11, cmp_plus11,
+                      cmp_plus11, cmp_minus10, cmp_plus10, cmp_minus11,
+                      cmp_plus11);
+            DUP4_ARG2(__lsx_vnor_v, cmp_minus12, cmp_minus12, cmp_plus12,
+                      cmp_plus12, cmp_minus13, cmp_minus13, cmp_plus13,
+                      cmp_plus13, cmp_minus12, cmp_plus12, cmp_minus13,
+                      cmp_plus13);
+            DUP4_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10,
+                      diff_plus10, const1, cmp_plus10, diff_minus11, const1,
+                      cmp_minus11, diff_plus11, const1, cmp_plus11,
+                      diff_minus10, diff_plus10, diff_minus11, diff_plus11);
+            DUP4_ARG3(__lsx_vbitsel_v, diff_minus12, const1, cmp_minus12,
+                      diff_plus12, const1, cmp_plus12, diff_minus13, const1,
+                      cmp_minus13, diff_plus13, const1, cmp_plus13,
+                      diff_minus12, diff_plus12, diff_minus13, diff_plus13);
+
+            DUP4_ARG2(__lsx_vadd_b, diff_minus10, diff_plus10, diff_minus11,
+                      diff_plus11, diff_minus12, diff_plus12, diff_minus13,
+                      diff_plus13, offset_mask0, offset_mask1, offset_mask2,
+                      offset_mask3);
+            DUP4_ARG2(__lsx_vaddi_bu, offset_mask0, 2, offset_mask1, 2,
+                      offset_mask2, 2, offset_mask3, 2, offset_mask0,
+                      offset_mask1, offset_mask2, offset_mask3);
+            DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset_mask0,
+                      sao_offset, sao_offset, offset_mask0,\
+                      offset_mask0, offset_mask0);
+            DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset_mask1,
+                      sao_offset, sao_offset, offset_mask1, offset_mask1,
+                      offset_mask1);
+            DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset_mask2,
+                      sao_offset, sao_offset, offset_mask2, offset_mask2,
+                      offset_mask2);
+            DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset_mask3,
+                      sao_offset, sao_offset, offset_mask3, offset_mask3,
+                      offset_mask3);
+
+            src_minus10 = src12;
+            DUP4_ARG2(__lsx_vxori_b, src_minus11, 128, src10, 128, src11, 128,
+                      src12, 128, src_minus11, src10, src11, src12);
+            DUP4_ARG2(__lsx_vsadd_b, src_minus11, offset_mask0, src10,
+                      offset_mask1, src11, offset_mask2, src12,
+                      offset_mask3, dst0, dst1, dst2, dst3);
+            DUP4_ARG2(__lsx_vxori_b, dst0, 128, dst1, 128, dst2, 128, dst3,
+                      128, dst0, dst1, dst2, dst3);
+            src_minus11 = src13;
+
+            __lsx_vst(dst0, dst, 0);
+            __lsx_vstx(dst1, dst, dst_stride);
+            __lsx_vstx(dst2, dst, dst_stride_2x);
+            __lsx_vstx(dst3, dst, dst_stride_3x);
+            src += src_stride_4x;
+            dst += dst_stride_4x;
+        }
+    }
+}
+
+static void hevc_sao_edge_filter_45degree_4width_lsx(uint8_t *dst,
+                                                     int32_t dst_stride,
+                                                     uint8_t *src,
+                                                     int32_t src_stride,
+                                                     int16_t *sao_offset_val,
+                                                     int32_t height)
+{
+    uint8_t *src_orig;
+    const int32_t src_stride_2x = (src_stride << 1);
+    const int32_t dst_stride_2x = (dst_stride << 1);
+    __m128i shuf1 = {0x807060504030201, 0x100F0E0D0C0B0A09};
+    __m128i shuf2 = {0x908070605040302, 0x11100F0E0D0C0B0A};
+    __m128i edge_idx = {0x403000201, 0x0};
+    __m128i const1 = __lsx_vldi(1);
+    __m128i offset, sao_offset = __lsx_vld(sao_offset_val, 0);
+    __m128i cmp_minus10, diff_minus10, src_minus10, cmp_minus11, diff_minus11;
+    __m128i src_minus11, src10, src11;
+    __m128i src_plus0, src_zero0, src_plus1, src_zero1, dst0;
+    __m128i offset_mask0, offset_mask1;
+    __m128i zeros = {0};
+
+    sao_offset = __lsx_vpickev_b(sao_offset, sao_offset);
+    src_orig = src - 1;
+
+    /* load in advance */
+    DUP2_ARG2(__lsx_vld, src_orig - src_stride, 0, src_orig, 0,
+              src_minus10, src_minus11);
+    DUP2_ARG2(__lsx_vldx, src_orig, src_stride, src_orig, src_stride_2x,
+              src10, src11);
+
+    for (height -= 2; height; height -= 2) {
+        src_orig += src_stride_2x;
+
+        DUP2_ARG3(__lsx_vshuf_b, zeros, src_minus11, shuf1, zeros, src10,
+                  shuf1, src_zero0, src_zero1);
+        DUP2_ARG3(__lsx_vshuf_b, zeros, src10, shuf2, zeros, src11, shuf2,
+                  src_plus0, src_plus1);
+
+        DUP2_ARG2(__lsx_vilvl_b, src_plus0, src_minus10, src_plus1,
+                  src_minus11, src_minus10, src_minus11);
+        DUP2_ARG2(__lsx_vilvl_b, src_zero0, src_zero0, src_zero1,
+                  src_zero1, src_zero0, src_zero1);
+        DUP2_ARG2(__lsx_vseq_b, src_zero0, src_minus10, src_zero1,
+                  src_minus11, cmp_minus10, cmp_minus11);
+        DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+                  cmp_minus11, diff_minus10, diff_minus11);
+        DUP2_ARG2(__lsx_vsle_bu, src_zero0, src_minus10, src_zero1,
+                  src_minus11, cmp_minus10, cmp_minus11);
+        DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+                  cmp_minus11, cmp_minus10, cmp_minus11);
+        DUP2_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10,
+             diff_minus11, const1, cmp_minus11, diff_minus10, diff_minus11);
+
+        DUP2_ARG2(__lsx_vhaddw_hu_bu, diff_minus10, diff_minus10, diff_minus11,
+                  diff_minus11, offset_mask0, offset_mask1);
+        DUP2_ARG2(__lsx_vaddi_hu, offset_mask0, 2, offset_mask1, 2,
+                  offset_mask0, offset_mask1);
+        DUP2_ARG2(__lsx_vpickev_b, offset_mask1, offset_mask0, src_zero1,
+                  src_zero0, offset, dst0);
+        DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset, sao_offset,
+                  sao_offset, offset, offset, offset);
+        dst0 = __lsx_vxori_b(dst0, 128);
+        dst0 = __lsx_vsadd_b(dst0, offset);
+        dst0 = __lsx_vxori_b(dst0, 128);
+
+        src_minus10 = src10;
+        src_minus11 = src11;
+
+        /* load in advance */
+        DUP2_ARG2(__lsx_vldx, src_orig, src_stride, src_orig, src_stride_2x,
+                  src10, src11);
+
+        __lsx_vstelm_w(dst0, dst, 0, 0);
+        __lsx_vstelm_w(dst0, dst + dst_stride, 0, 2);
+        dst += dst_stride_2x;
+    }
+
+    DUP2_ARG3(__lsx_vshuf_b, zeros, src_minus11, shuf1, zeros, src10, shuf1,
+              src_zero0, src_zero1);
+    DUP2_ARG3(__lsx_vshuf_b, zeros, src10, shuf2, zeros, src11, shuf2,
+              src_plus0, src_plus1);
+
+    DUP2_ARG2(__lsx_vilvl_b, src_plus0, src_minus10, src_plus1, src_minus11,
+              src_minus10, src_minus11);
+    DUP2_ARG2(__lsx_vilvl_b, src_zero0, src_zero0, src_zero1, src_zero1,
+              src_zero0, src_zero1);
+    DUP2_ARG2(__lsx_vseq_b, src_zero0, src_minus10, src_zero1, src_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11, cmp_minus11,
+              diff_minus10, diff_minus11);
+    DUP2_ARG2(__lsx_vsle_bu, src_zero0, src_minus10, src_zero1, src_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11, cmp_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10, diff_minus11,
+              const1, cmp_minus11, diff_minus10, diff_minus11);
+
+    DUP2_ARG2(__lsx_vhaddw_hu_bu, diff_minus10, diff_minus10, diff_minus11,
+              diff_minus11, offset_mask0, offset_mask1);
+    DUP2_ARG2(__lsx_vaddi_hu, offset_mask0, 2, offset_mask1, 2, offset_mask0,
+              offset_mask1);
+    DUP2_ARG2(__lsx_vpickev_b, offset_mask1, offset_mask0, src_zero1,
+              src_zero0, offset, dst0);
+    DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset, sao_offset,
+              sao_offset, offset, offset, offset);
+    dst0 = __lsx_vxori_b(dst0, 128);
+    dst0 = __lsx_vsadd_b(dst0, offset);
+    dst0 = __lsx_vxori_b(dst0, 128);
+
+    __lsx_vstelm_w(dst0, dst, 0, 0);
+    __lsx_vstelm_w(dst0, dst + dst_stride, 0, 2);
+}
+
+static void hevc_sao_edge_filter_45degree_8width_lsx(uint8_t *dst,
+                                                     int32_t dst_stride,
+                                                     uint8_t *src,
+                                                     int32_t src_stride,
+                                                     int16_t *sao_offset_val,
+                                                     int32_t height)
+{
+    uint8_t *src_orig;
+    const int32_t src_stride_2x = (src_stride << 1);
+    const int32_t dst_stride_2x = (dst_stride << 1);
+    __m128i shuf1 = {0x807060504030201, 0x100F0E0D0C0B0A09};
+    __m128i shuf2 = {0x908070605040302, 0x11100F0E0D0C0B0A};
+    __m128i edge_idx = {0x403000201, 0x0};
+    __m128i const1 = __lsx_vldi(1);
+    __m128i offset, sao_offset = __lsx_vld(sao_offset_val, 0);
+    __m128i cmp_minus10, diff_minus10, cmp_minus11, diff_minus11;
+    __m128i src_minus10, src10, src_minus11, src11;
+    __m128i src_zero0, src_plus10, src_zero1, src_plus11, dst0;
+    __m128i offset_mask0, offset_mask1;
+    __m128i zeros = {0};
+
+    sao_offset = __lsx_vpickev_b(sao_offset, sao_offset);
+    src_orig = src - 1;
+
+    /* load in advance */
+    DUP2_ARG2(__lsx_vld, src_orig - src_stride, 0, src_orig, 0, src_minus10,
+              src_minus11);
+    DUP2_ARG2(__lsx_vldx, src_orig, src_stride, src_orig, src_stride_2x,
+              src10, src11);
+
+    for (height -= 2; height; height -= 2) {
+        src_orig += src_stride_2x;
+
+        DUP2_ARG3(__lsx_vshuf_b, zeros, src_minus11, shuf1, zeros, src10,
+                  shuf1, src_zero0, src_zero1);
+        DUP2_ARG3(__lsx_vshuf_b, zeros, src10, shuf2, zeros, src11, shuf2,
+                  src_plus10, src_plus11);
+
+        DUP2_ARG2(__lsx_vilvl_b, src_plus10, src_minus10, src_plus11,
+                  src_minus11, src_minus10, src_minus11);
+        DUP2_ARG2(__lsx_vilvl_b, src_zero0, src_zero0, src_zero1, src_zero1,
+                  src_zero0, src_zero1);
+        DUP2_ARG2(__lsx_vseq_b, src_zero0, src_minus10, src_zero1, src_minus11,
+                  cmp_minus10, cmp_minus11);
+        DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+                  cmp_minus11, diff_minus10, diff_minus11);
+        DUP2_ARG2(__lsx_vsle_bu, src_zero0, src_minus10, src_zero1,
+                  src_minus11, cmp_minus10, cmp_minus11);
+        DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+                  cmp_minus11, cmp_minus10, cmp_minus11);
+        DUP2_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10,
+               diff_minus11, const1, cmp_minus11,  diff_minus10, diff_minus11);
+
+        DUP2_ARG2(__lsx_vhaddw_hu_bu, diff_minus10, diff_minus10, diff_minus11,
+                  diff_minus11, offset_mask0, offset_mask1);
+        DUP2_ARG2(__lsx_vaddi_hu, offset_mask0, 2, offset_mask1, 2,
+                  offset_mask0, offset_mask1);
+        DUP2_ARG2(__lsx_vpickev_b, offset_mask1, offset_mask0, src_zero1,
+                  src_zero0, offset, dst0);
+        DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset, sao_offset,
+                  sao_offset, offset, offset, offset);
+        dst0 = __lsx_vxori_b(dst0, 128);
+        dst0 = __lsx_vsadd_b(dst0, offset);
+        dst0 = __lsx_vxori_b(dst0, 128);
+
+        src_minus10 = src10;
+        src_minus11 = src11;
+
+        /* load in advance */
+        DUP2_ARG2(__lsx_vldx, src_orig, src_stride, src_orig, src_stride_2x,
+                  src10, src11)
+        __lsx_vstelm_d(dst0, dst, 0, 0);
+        __lsx_vstelm_d(dst0, dst + dst_stride, 0, 1);
+        dst += dst_stride_2x;
+    }
+
+    DUP2_ARG3(__lsx_vshuf_b, zeros, src_minus11, shuf1, zeros, src10, shuf1,
+              src_zero0, src_zero1);
+    DUP2_ARG3(__lsx_vshuf_b, zeros, src10, shuf2, zeros, src11, shuf2,
+              src_plus10, src_plus11);
+    DUP2_ARG2(__lsx_vilvl_b, src_plus10, src_minus10, src_plus11, src_minus11,
+              src_minus10, src_minus11);
+    DUP2_ARG2(__lsx_vilvl_b, src_zero0, src_zero0, src_zero1, src_zero1,
+              src_zero0, src_zero1);
+
+    DUP2_ARG2(__lsx_vseq_b, src_zero0, src_minus10, src_zero1, src_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+              cmp_minus11, diff_minus10, diff_minus11);
+    DUP2_ARG2(__lsx_vsle_bu, src_zero0, src_minus10, src_zero1, src_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11, cmp_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10, diff_minus11,
+              const1, cmp_minus11, diff_minus10, diff_minus11);
+
+    DUP2_ARG2(__lsx_vhaddw_hu_bu, diff_minus10, diff_minus10, diff_minus11,
+              diff_minus11, offset_mask0, offset_mask1);
+    DUP2_ARG2(__lsx_vaddi_hu, offset_mask0, 2, offset_mask1, 2, offset_mask0,
+              offset_mask1);
+    DUP2_ARG2(__lsx_vpickev_b, offset_mask1, offset_mask0, src_zero1,
+              src_zero0, offset, dst0);
+    DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset, sao_offset,
+              sao_offset, offset, offset, offset);
+    dst0 = __lsx_vxori_b(dst0, 128);
+    dst0 = __lsx_vsadd_b(dst0, offset);
+    dst0 = __lsx_vxori_b(dst0, 128);
+
+    src_minus10 = src10;
+    src_minus11 = src11;
+
+    /* load in advance */
+    DUP2_ARG2(__lsx_vldx, src_orig, src_stride, src_orig, src_stride_2x,
+              src10, src11);
+
+    __lsx_vstelm_d(dst0, dst, 0, 0);
+    __lsx_vstelm_d(dst0, dst + dst_stride, 0, 1);
+}
+
+static void hevc_sao_edge_filter_45degree_16multiple_lsx(uint8_t *dst,
+                                                         int32_t dst_stride,
+                                                         uint8_t *src,
+                                                         int32_t src_stride,
+                                                         int16_t *
+                                                         sao_offset_val,
+                                                         int32_t width,
+                                                         int32_t height)
+{
+    uint8_t *src_orig = src;
+    uint8_t *dst_orig = dst;
+    int32_t v_cnt;
+    const int32_t src_stride_2x = (src_stride << 1);
+    const int32_t dst_stride_2x = (dst_stride << 1);
+    const int32_t src_stride_4x = (src_stride << 2);
+    const int32_t dst_stride_4x = (dst_stride << 2);
+    const int32_t src_stride_3x = src_stride_2x + src_stride;
+    const int32_t dst_stride_3x = dst_stride_2x + dst_stride;
+
+    __m128i shuf1 = {0x807060504030201, 0x100F0E0D0C0B0A09};
+    __m128i shuf2 = {0x908070605040302, 0x11100F0E0D0C0B0A};
+    __m128i edge_idx = {0x403000201, 0x0};
+    __m128i const1 = __lsx_vldi(1);
+    __m128i cmp_minus10, cmp_plus10, diff_minus10, diff_plus10, cmp_minus11;
+    __m128i cmp_plus11, diff_minus11, diff_plus11, cmp_minus12, cmp_plus12;
+    __m128i diff_minus12, diff_plus12, cmp_minus13, cmp_plus13, diff_minus13;
+    __m128i diff_plus13, src_minus14, src_plus13;
+    __m128i offset_mask0, offset_mask1, offset_mask2, offset_mask3;
+    __m128i src10, src_minus10, dst0, src11, src_minus11, dst1;
+    __m128i src12, src_minus12, dst2, src13, src_minus13, dst3;
+    __m128i src_zero0, src_plus10, src_zero1, src_plus11, src_zero2;
+    __m128i src_zero3, sao_offset, src_plus12;
+
+    sao_offset = __lsx_vld(sao_offset_val, 0);
+    sao_offset = __lsx_vpickev_b(sao_offset, sao_offset);
+
+    for (; height; height -= 4) {
+        src_orig = src - 1;
+        dst_orig = dst;
+        src_minus11 = __lsx_vld(src_orig, 0);
+        DUP2_ARG2(__lsx_vldx, src_orig, src_stride, src_orig, src_stride_2x,
+                  src_minus12, src_minus13);
+        src_minus14 = __lsx_vldx(src_orig, src_stride_3x);
+
+        for (v_cnt = 0; v_cnt < width; v_cnt += 16) {
+            src_minus10 = __lsx_vld(src_orig - src_stride, 0);
+            src_orig += 16;
+            src10 = __lsx_vld(src_orig, 0);
+            DUP2_ARG2(__lsx_vldx, src_orig, src_stride, src_orig,
+                      src_stride_2x, src11, src12);
+            src13 = __lsx_vldx(src_orig, src_stride_3x);
+            src_plus13 = __lsx_vld(src + v_cnt + src_stride_4x, 1);
+
+            DUP4_ARG3(__lsx_vshuf_b, src10, src_minus11, shuf1, src11,
+                      src_minus12, shuf1, src12, src_minus13, shuf1,
+                      src13, src_minus14, shuf1, src_zero0, src_zero1,
+                      src_zero2, src_zero3);
+            DUP2_ARG3(__lsx_vshuf_b, src11, src_minus12, shuf2, src12,
+                      src_minus13, shuf2, src_plus10, src_plus11);
+            src_plus12 = __lsx_vshuf_b(src13, src_minus14, shuf2);
+
+            DUP4_ARG2(__lsx_vseq_b, src_zero0, src_minus10, src_zero0,
+                      src_plus10, src_zero1, src_minus11, src_zero1,
+                      src_plus11, cmp_minus10, cmp_plus10,
+                      cmp_minus11, cmp_plus11);
+            DUP4_ARG2(__lsx_vseq_b, src_zero2, src_minus12, src_zero2,
+                      src_plus12, src_zero3, src_minus13, src_zero3,
+                      src_plus13, cmp_minus12, cmp_plus12,
+                      cmp_minus13, cmp_plus13);
+            DUP4_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_plus10,
+                      cmp_plus10, cmp_minus11, cmp_minus11, cmp_plus11,
+                      cmp_plus11, diff_minus10, diff_plus10, diff_minus11,
+                      diff_plus11);
+            DUP4_ARG2(__lsx_vnor_v, cmp_minus12, cmp_minus12, cmp_plus12,
+                      cmp_plus12, cmp_minus13, cmp_minus13, cmp_plus13,
+                      cmp_plus13, diff_minus12, diff_plus12, diff_minus13,
+                      diff_plus13);
+            DUP4_ARG2(__lsx_vsle_bu, src_zero0, src_minus10, src_zero0,
+                      src_plus10, src_zero1, src_minus11, src_zero1,
+                      src_plus11, cmp_minus10, cmp_plus10, cmp_minus11,
+                      cmp_plus11);
+            DUP4_ARG2(__lsx_vsle_bu, src_zero2, src_minus12, src_zero2,
+                      src_plus12, src_zero3, src_minus13, src_zero3,
+                      src_plus13, cmp_minus12, cmp_plus12, cmp_minus13,
+                      cmp_plus13);
+            DUP4_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_plus10,
+                      cmp_plus10, cmp_minus11, cmp_minus11, cmp_plus11,
+                      cmp_plus11, cmp_minus10, cmp_plus10, cmp_minus11,
+                      cmp_plus11);
+            DUP4_ARG2(__lsx_vnor_v, cmp_minus12, cmp_minus12, cmp_plus12,
+                      cmp_plus12, cmp_minus13, cmp_minus13, cmp_plus13,
+                      cmp_plus13, cmp_minus12, cmp_plus12, cmp_minus13,
+                      cmp_plus13);
+            DUP4_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10,
+                      diff_plus10, const1, cmp_plus10, diff_minus11, const1,
+                      cmp_minus11, diff_plus11, const1, cmp_plus11,
+                      diff_minus10, diff_plus10, diff_minus11, diff_plus11);
+            DUP4_ARG3(__lsx_vbitsel_v, diff_minus12, const1, cmp_minus12,
+                      diff_plus12, const1, cmp_plus12, diff_minus13, const1,
+                      cmp_minus13, diff_plus13, const1, cmp_plus13,
+                      diff_minus12, diff_plus12, diff_minus13, diff_plus13);
+
+            DUP4_ARG2(__lsx_vadd_b, diff_minus10, diff_plus10, diff_minus11,
+                      diff_plus11, diff_minus12, diff_plus12, diff_minus13,
+                      diff_plus13, offset_mask0, offset_mask1, offset_mask2,
+                      offset_mask3);
+            DUP4_ARG2(__lsx_vaddi_bu, offset_mask0, 2, offset_mask1, 2,
+                      offset_mask2, 2, offset_mask3, 2, offset_mask0,
+                      offset_mask1, offset_mask2, offset_mask3);
+
+            DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset_mask0,
+                      sao_offset, sao_offset, offset_mask0, offset_mask0,
+                      offset_mask0);
+            DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset_mask1,
+                      sao_offset, sao_offset, offset_mask1, offset_mask1,
+                      offset_mask1);
+            DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset_mask2,
+                      sao_offset, sao_offset, offset_mask2, offset_mask2,
+                      offset_mask2);
+            DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset_mask3,
+                      sao_offset, sao_offset, offset_mask3, offset_mask3,
+                      offset_mask3);
+
+            DUP4_ARG2(__lsx_vxori_b, src_zero0, 128, src_zero1, 128, src_zero2,
+                      128, src_zero3, 128, src_zero0, src_zero1, src_zero2,
+                      src_zero3);
+            DUP4_ARG2(__lsx_vsadd_b, src_zero0, offset_mask0, src_zero1,
+                      offset_mask1, src_zero2, offset_mask2, src_zero3,
+                      offset_mask3, dst0, dst1, dst2, dst3);
+            DUP4_ARG2(__lsx_vxori_b, dst0, 128, dst1, 128, dst2, 128, dst3,
+                      128, dst0, dst1, dst2, dst3);
+
+            src_minus11 = src10;
+            src_minus12 = src11;
+            src_minus13 = src12;
+            src_minus14 = src13;
+
+            __lsx_vst(dst0, dst_orig, 0);
+            __lsx_vstx(dst1, dst_orig, dst_stride);
+            __lsx_vstx(dst2, dst_orig, dst_stride_2x);
+            __lsx_vstx(dst3, dst_orig, dst_stride_3x);
+            dst_orig += 16;
+        }
+        src += src_stride_4x;
+        dst += dst_stride_4x;
+    }
+}
+
+static void hevc_sao_edge_filter_135degree_4width_lsx(uint8_t *dst,
+                                                      int32_t dst_stride,
+                                                      uint8_t *src,
+                                                      int32_t src_stride,
+                                                      int16_t *sao_offset_val,
+                                                      int32_t height)
+{
+    uint8_t *src_orig;
+    const int32_t src_stride_2x = (src_stride << 1);
+    const int32_t dst_stride_2x = (dst_stride << 1);
+
+    __m128i shuf1 = {0x807060504030201, 0x100F0E0D0C0B0A09};
+    __m128i shuf2 = {0x908070605040302, 0x11100F0E0D0C0B0A};
+    __m128i edge_idx = {0x403000201, 0x0};
+    __m128i const1 = __lsx_vldi(1);
+    __m128i offset, sao_offset = __lsx_vld(sao_offset_val, 0);
+    __m128i src_zero0, src_zero1, dst0;
+    __m128i cmp_minus10, diff_minus10, cmp_minus11, diff_minus11;
+    __m128i src_minus10, src10, src_minus11, src11;
+    __m128i offset_mask0, offset_mask1;
+    __m128i zeros = {0};
+
+    sao_offset = __lsx_vpickev_b(sao_offset, sao_offset);
+    src_orig = src - 1;
+
+    /* load in advance */
+    DUP2_ARG2(__lsx_vld, src_orig - src_stride, 0, src_orig, 0,
+              src_minus10, src_minus11);
+    DUP2_ARG2(__lsx_vldx, src_orig, src_stride, src_orig, src_stride_2x,
+              src10, src11);
+
+    for (height -= 2; height; height -= 2) {
+        src_orig += src_stride_2x;
+
+        DUP2_ARG3(__lsx_vshuf_b, zeros, src_minus11, shuf1, zeros, src10,
+                  shuf1, src_zero0, src_zero1);
+        DUP2_ARG3(__lsx_vshuf_b, zeros, src_minus10, shuf2, zeros, src_minus11,
+                  shuf2, src_minus10, src_minus11);
+
+        DUP2_ARG2(__lsx_vilvl_b, src10, src_minus10, src11, src_minus11,
+                  src_minus10, src_minus11);
+        DUP2_ARG2(__lsx_vilvl_b, src_zero0, src_zero0, src_zero1, src_zero1,
+                  src_zero0, src_zero1);
+        DUP2_ARG2(__lsx_vseq_b, src_zero0, src_minus10, src_zero1, src_minus11,
+                  cmp_minus10, cmp_minus11);
+        DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+                  cmp_minus11, diff_minus10, diff_minus11);
+        DUP2_ARG2(__lsx_vsle_bu, src_zero0, src_minus10, src_zero1,
+                  src_minus11, cmp_minus10, cmp_minus11);
+        DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+                  cmp_minus11, cmp_minus10, cmp_minus11);
+        DUP2_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10,
+               diff_minus11, const1, cmp_minus11,  diff_minus10, diff_minus11);
+
+        DUP2_ARG2(__lsx_vhaddw_hu_bu, diff_minus10, diff_minus10, diff_minus11,
+                  diff_minus11, offset_mask0, offset_mask1);
+        DUP2_ARG2(__lsx_vaddi_hu, offset_mask0, 2, offset_mask1, 2,
+                  offset_mask0, offset_mask1);
+        DUP2_ARG2(__lsx_vpickev_b, offset_mask1, offset_mask0, src_zero1,
+                  src_zero0, offset, dst0);
+        DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset, sao_offset,
+                  sao_offset, offset, offset, offset);
+        dst0 = __lsx_vxori_b(dst0, 128);
+        dst0 = __lsx_vsadd_b(dst0, offset);
+        dst0 = __lsx_vxori_b(dst0, 128);
+
+        src_minus10 = src10;
+        src_minus11 = src11;
+
+        /* load in advance */
+        DUP2_ARG2(__lsx_vldx, src_orig, src_stride, src_orig, src_stride_2x,
+                  src10, src11);
+
+        __lsx_vstelm_w(dst0, dst, 0, 0);
+        __lsx_vstelm_w(dst0, dst + dst_stride, 0, 2);
+        dst += dst_stride_2x;
+    }
+
+    DUP2_ARG3(__lsx_vshuf_b, zeros, src_minus11, shuf1, zeros, src10, shuf1,
+              src_zero0, src_zero1);
+    DUP2_ARG3(__lsx_vshuf_b, zeros, src_minus10, shuf2, zeros, src_minus11,
+              shuf2, src_minus10, src_minus11);
+
+    DUP2_ARG2(__lsx_vilvl_b, src10, src_minus10, src11, src_minus11,
+              src_minus10, src_minus11);
+    DUP2_ARG2(__lsx_vilvl_b, src_zero0, src_zero0, src_zero1, src_zero1,
+              src_zero0, src_zero1);
+    DUP2_ARG2(__lsx_vseq_b, src_zero0, src_minus10, src_zero1, src_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+              cmp_minus11, diff_minus10, diff_minus11);
+    DUP2_ARG2(__lsx_vsle_bu, src_zero0, src_minus10, src_zero1, src_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11, cmp_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10, diff_minus11,
+              const1, cmp_minus11, diff_minus10, diff_minus11);
+
+    DUP2_ARG2(__lsx_vhaddw_hu_bu, diff_minus10, diff_minus10, diff_minus11,
+              diff_minus11, offset_mask0, offset_mask1);
+    DUP2_ARG2(__lsx_vaddi_hu, offset_mask0, 2, offset_mask1, 2, offset_mask0,
+              offset_mask1);
+    DUP2_ARG2(__lsx_vpickev_b, offset_mask1, offset_mask0, src_zero1,
+              src_zero0, offset, dst0);
+    DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset, sao_offset,
+              sao_offset, offset, offset, offset);
+    dst0 = __lsx_vxori_b(dst0, 128);
+    dst0 = __lsx_vsadd_b(dst0, offset);
+    dst0 = __lsx_vxori_b(dst0, 128);
+
+    __lsx_vstelm_w(dst0, dst, 0, 0);
+    __lsx_vstelm_w(dst0, dst + dst_stride, 0, 2);
+    dst += dst_stride_2x;
+}
+
+static void hevc_sao_edge_filter_135degree_8width_lsx(uint8_t *dst,
+                                                      int32_t dst_stride,
+                                                      uint8_t *src,
+                                                      int32_t src_stride,
+                                                      int16_t *sao_offset_val,
+                                                      int32_t height)
+{
+    uint8_t *src_orig;
+    const int32_t src_stride_2x = (src_stride << 1);
+    const int32_t dst_stride_2x = (dst_stride << 1);
+
+    __m128i shuf1 = {0x807060504030201, 0x100F0E0D0C0B0A09};
+    __m128i shuf2 = {0x908070605040302, 0x11100F0E0D0C0B0A};
+    __m128i edge_idx = {0x403000201, 0x0};
+    __m128i const1 = __lsx_vldi(1);
+    __m128i offset, sao_offset = __lsx_vld(sao_offset_val, 0);
+    __m128i cmp_minus10, diff_minus10, cmp_minus11, diff_minus11;
+    __m128i src_minus10, src10, src_minus11, src11;
+    __m128i src_zero0, src_zero1, dst0;
+    __m128i offset_mask0, offset_mask1;
+    __m128i zeros = {0};
+
+    sao_offset = __lsx_vpickev_b(sao_offset, sao_offset);
+    src_orig = src - 1;
+
+    /* load in advance */
+    DUP2_ARG2(__lsx_vld, src_orig - src_stride, 0, src_orig, 0,
+              src_minus10, src_minus11);
+    DUP2_ARG2(__lsx_vldx, src_orig, src_stride, src_orig, src_stride_2x,
+              src10, src11);
+
+    for (height -= 2; height; height -= 2) {
+        src_orig += src_stride_2x;
+
+        DUP2_ARG3(__lsx_vshuf_b, zeros, src_minus11, shuf1, zeros, src10,
+                  shuf1, src_zero0, src_zero1);
+        DUP2_ARG3(__lsx_vshuf_b, zeros, src_minus10, shuf2, zeros, src_minus11,
+                  shuf2, src_minus10, src_minus11);
+
+        DUP2_ARG2(__lsx_vilvl_b, src10, src_minus10, src11, src_minus11,
+                  src_minus10, src_minus11);
+        DUP2_ARG2(__lsx_vilvl_b, src_zero0, src_zero0, src_zero1, src_zero1,
+                  src_zero0, src_zero1);
+        DUP2_ARG2(__lsx_vseq_b, src_zero0, src_minus10, src_zero1, src_minus11,
+                  cmp_minus10, cmp_minus11);
+        DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+                  cmp_minus11, diff_minus10, diff_minus11);
+        DUP2_ARG2(__lsx_vsle_bu, src_zero0, src_minus10, src_zero1,
+                  src_minus11, cmp_minus10, cmp_minus11);
+        DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11,
+                  cmp_minus11, cmp_minus10, cmp_minus11);
+        DUP2_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10,
+              diff_minus11, const1, cmp_minus11,  diff_minus10, diff_minus11);
+
+        DUP2_ARG2(__lsx_vhaddw_hu_bu, diff_minus10, diff_minus10, diff_minus11,
+                  diff_minus11, offset_mask0, offset_mask1);
+        DUP2_ARG2(__lsx_vaddi_hu, offset_mask0, 2, offset_mask1, 2,
+                  offset_mask0, offset_mask1);
+        DUP2_ARG2(__lsx_vpickev_b, offset_mask1, offset_mask0, src_zero1,
+                  src_zero0, offset, dst0);
+        DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset, sao_offset,
+                  sao_offset, offset, offset, offset);
+        dst0 = __lsx_vxori_b(dst0, 128);
+        dst0 = __lsx_vsadd_b(dst0, offset);
+        dst0 = __lsx_vxori_b(dst0, 128);
+
+        src_minus10 = src10;
+        src_minus11 = src11;
+
+        /* load in advance */
+        DUP2_ARG2(__lsx_vldx, src_orig, src_stride, src_orig, src_stride_2x,
+                  src10, src11);
+
+        __lsx_vstelm_d(dst0, dst, 0, 0);
+        __lsx_vstelm_d(dst0, dst + dst_stride, 0, 1);
+        dst += dst_stride_2x;
+    }
+
+    DUP2_ARG3(__lsx_vshuf_b, zeros, src_minus11, shuf1, zeros, src10, shuf1,
+              src_zero0, src_zero1);
+    DUP2_ARG3(__lsx_vshuf_b, zeros, src_minus10, shuf2, zeros, src_minus11,
+              shuf2, src_minus10, src_minus11);
+
+    DUP2_ARG2(__lsx_vilvl_b, src10, src_minus10, src11, src_minus11,
+              src_minus10, src_minus11);
+    DUP2_ARG2(__lsx_vilvl_b, src_zero0, src_zero0, src_zero1, src_zero1,
+              src_zero0, src_zero1);
+    DUP2_ARG2(__lsx_vseq_b, src_zero0, src_minus10, src_zero1, src_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11, cmp_minus11,
+              diff_minus10, diff_minus11);
+    DUP2_ARG2(__lsx_vsle_bu, src_zero0, src_minus10, src_zero1, src_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_minus11, cmp_minus11,
+              cmp_minus10, cmp_minus11);
+    DUP2_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10, diff_minus11,
+              const1, cmp_minus11, diff_minus10, diff_minus11);
+
+    DUP2_ARG2(__lsx_vhaddw_hu_bu, diff_minus10, diff_minus10, diff_minus11,
+              diff_minus11, offset_mask0, offset_mask1);
+    DUP2_ARG2(__lsx_vaddi_hu, offset_mask0, 2, offset_mask1, 2, offset_mask0,
+              offset_mask1);
+    DUP2_ARG2(__lsx_vpickev_b, offset_mask1, offset_mask0, src_zero1,
+              src_zero0, offset, dst0);
+    DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset, sao_offset,
+              sao_offset, offset, offset, offset);
+    dst0 = __lsx_vxori_b(dst0, 128);
+    dst0 = __lsx_vsadd_b(dst0, offset);
+    dst0 = __lsx_vxori_b(dst0, 128);
+
+    __lsx_vstelm_d(dst0, dst, 0, 0);
+    __lsx_vstelm_d(dst0, dst + dst_stride, 0, 1);
+}
+
+static void hevc_sao_edge_filter_135degree_16multiple_lsx(uint8_t *dst,
+                                                          int32_t dst_stride,
+                                                          uint8_t *src,
+                                                          int32_t src_stride,
+                                                          int16_t *sao_offset_val,
+                                                          int32_t width,
+                                                          int32_t height)
+{
+    uint8_t *src_orig, *dst_orig;
+    int32_t v_cnt;
+    const int32_t src_stride_2x = (src_stride << 1);
+    const int32_t dst_stride_2x = (dst_stride << 1);
+    const int32_t src_stride_4x = (src_stride << 2);
+    const int32_t dst_stride_4x = (dst_stride << 2);
+    const int32_t src_stride_3x = src_stride_2x + src_stride;
+    const int32_t dst_stride_3x = dst_stride_2x + dst_stride;
+
+    __m128i shuf1 = {0x807060504030201, 0x100F0E0D0C0B0A09};
+    __m128i shuf2 = {0x908070605040302, 0x11100F0E0D0C0B0A};
+    __m128i edge_idx = {0x403000201, 0x0};
+    __m128i const1 = __lsx_vldi(1);
+    __m128i dst0, dst1, dst2, dst3;
+    __m128i cmp_minus10, cmp_minus11, cmp_minus12, cmp_minus13, cmp_plus10;
+    __m128i cmp_plus11, cmp_plus12, cmp_plus13, diff_minus10, diff_minus11;
+    __m128i diff_minus12, diff_minus13, diff_plus10, diff_plus11, diff_plus12;
+    __m128i diff_plus13, src10, src11, src12, src13, src_minus10, src_minus11;
+    __m128i src_plus10, src_plus11, src_plus12, src_plus13;
+    __m128i src_minus12, src_minus13, src_zero0, src_zero1, src_zero2, src_zero3;
+    __m128i offset_mask0, offset_mask1, offset_mask2, offset_mask3, sao_offset;
+
+    sao_offset = __lsx_vld(sao_offset_val, 0);
+    sao_offset = __lsx_vpickev_b(sao_offset, sao_offset);
+
+    for (; height; height -= 4) {
+        src_orig = src - 1;
+        dst_orig = dst;
+
+        src_minus11 = __lsx_vld(src_orig, 0);
+        DUP2_ARG2(__lsx_vldx, src_orig, src_stride, src_orig, src_stride_2x,
+                  src_plus10, src_plus11);
+        src_plus12 = __lsx_vldx(src_orig, src_stride_3x);
+
+        for (v_cnt = 0; v_cnt < width; v_cnt += 16) {
+            src_minus10 = __lsx_vld(src_orig - src_stride, 2);
+            src_plus13 = __lsx_vldx(src_orig, src_stride_4x);
+            src_orig += 16;
+            src10 = __lsx_vld(src_orig, 0);
+            DUP2_ARG2(__lsx_vldx, src_orig, src_stride, src_orig, src_stride_2x,
+                      src11, src12);
+            src13 =__lsx_vldx(src_orig, src_stride_3x);
+
+            DUP4_ARG3(__lsx_vshuf_b, src10, src_minus11, shuf1, src11,
+                      src_plus10,  shuf1, src12, src_plus11, shuf1, src13,
+                      src_plus12, shuf1, src_zero0, src_zero1, src_zero2,
+                      src_zero3);
+            src_minus11 = __lsx_vshuf_b(src10, src_minus11, shuf2);
+            DUP2_ARG3(__lsx_vshuf_b, src11, src_plus10, shuf2, src12,
+                      src_plus11, shuf2, src_minus12, src_minus13);
+
+            DUP4_ARG2(__lsx_vseq_b, src_zero0, src_minus10, src_zero0,
+                      src_plus10,  src_zero1, src_minus11, src_zero1,
+                      src_plus11, cmp_minus10, cmp_plus10, cmp_minus11,
+                      cmp_plus11);
+            DUP4_ARG2(__lsx_vseq_b, src_zero2, src_minus12, src_zero2,
+                      src_plus12, src_zero3, src_minus13, src_zero3,
+                      src_plus13, cmp_minus12, cmp_plus12, cmp_minus13,
+                      cmp_plus13);
+            DUP4_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_plus10,
+                      cmp_plus10, cmp_minus11, cmp_minus11, cmp_plus11,
+                      cmp_plus11, diff_minus10, diff_plus10, diff_minus11,
+                      diff_plus11);
+            DUP4_ARG2(__lsx_vnor_v, cmp_minus12, cmp_minus12, cmp_plus12,
+                      cmp_plus12, cmp_minus13, cmp_minus13, cmp_plus13,
+                      cmp_plus13, diff_minus12, diff_plus12, diff_minus13,
+                      diff_plus13);
+            DUP4_ARG2(__lsx_vsle_bu, src_zero0, src_minus10, src_zero0,
+                      src_plus10, src_zero1, src_minus11, src_zero1, src_plus11,
+                      cmp_minus10, cmp_plus10, cmp_minus11, cmp_plus11);
+            DUP4_ARG2(__lsx_vsle_bu, src_zero2, src_minus12, src_zero2,
+                      src_plus12, src_zero3, src_minus13, src_zero3, src_plus13,
+                      cmp_minus12, cmp_plus12, cmp_minus13, cmp_plus13);
+            DUP4_ARG2(__lsx_vnor_v, cmp_minus10, cmp_minus10, cmp_plus10,
+                      cmp_plus10, cmp_minus11, cmp_minus11, cmp_plus11,
+                      cmp_plus11, cmp_minus10, cmp_plus10, cmp_minus11,
+                      cmp_plus11);
+            DUP4_ARG2(__lsx_vnor_v, cmp_minus12, cmp_minus12, cmp_plus12,
+                      cmp_plus12, cmp_minus13, cmp_minus13, cmp_plus13,
+                      cmp_plus13, cmp_minus12, cmp_plus12, cmp_minus13,
+                      cmp_plus13);
+            DUP4_ARG3(__lsx_vbitsel_v, diff_minus10, const1, cmp_minus10,
+                      diff_plus10, const1, cmp_plus10, diff_minus11, const1,
+                      cmp_minus11, diff_plus11, const1, cmp_plus11,
+                      diff_minus10, diff_plus10, diff_minus11, diff_plus11);
+            DUP4_ARG3(__lsx_vbitsel_v, diff_minus12, const1, cmp_minus12,
+                      diff_plus12, const1, cmp_plus12, diff_minus13, const1,
+                      cmp_minus13, diff_plus13, const1, cmp_plus13,
+                      diff_minus12, diff_plus12, diff_minus13, diff_plus13);
+
+            DUP4_ARG2(__lsx_vadd_b, diff_minus10, diff_plus10, diff_minus11,
+                      diff_plus11, diff_minus12, diff_plus12, diff_minus13,
+                      diff_plus13, offset_mask0, offset_mask1, offset_mask2,
+                      offset_mask3);
+            DUP4_ARG2(__lsx_vaddi_bu, offset_mask0, 2, offset_mask1, 2,
+                      offset_mask2, 2, offset_mask3, 2, offset_mask0,
+                      offset_mask1, offset_mask2, offset_mask3);
+
+            DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset_mask0,
+                      sao_offset, sao_offset, offset_mask0, offset_mask0,
+                      offset_mask0);
+            DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset_mask1,
+                      sao_offset, sao_offset, offset_mask1, offset_mask1,
+                      offset_mask1);
+            DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset_mask2,
+                      sao_offset, sao_offset, offset_mask2, offset_mask2,
+                      offset_mask2);
+            DUP2_ARG3(__lsx_vshuf_b, edge_idx, edge_idx, offset_mask3,
+                      sao_offset, sao_offset, offset_mask3, offset_mask3,
+                      offset_mask3);
+
+            DUP4_ARG2(__lsx_vxori_b, src_zero0, 128, src_zero1, 128,
+                      src_zero2, 128, src_zero3, 128, src_zero0, src_zero1,
+                      src_zero2, src_zero3);
+            DUP4_ARG2(__lsx_vsadd_b, src_zero0, offset_mask0, src_zero1,
+                      offset_mask1, src_zero2, offset_mask2, src_zero3,
+                      offset_mask3, dst0, dst1, dst2, dst3);
+            DUP4_ARG2(__lsx_vxori_b, dst0, 128, dst1, 128, dst2, 128, dst3,
+                      128, dst0, dst1, dst2, dst3);
+
+            src_minus11 = src10;
+            src_plus10 = src11;
+            src_plus11 = src12;
+            src_plus12 = src13;
+
+            __lsx_vst(dst0, dst_orig, 0);
+            __lsx_vstx(dst1, dst_orig, dst_stride);
+            __lsx_vstx(dst2, dst_orig, dst_stride_2x);
+            __lsx_vstx(dst3, dst_orig, dst_stride_3x);
+            dst_orig += 16;
+        }
+
+        src += src_stride_4x;
+        dst += dst_stride_4x;
+    }
+}
+
+void ff_hevc_sao_edge_filter_8_lsx(uint8_t *dst, uint8_t *src,
+                                   ptrdiff_t stride_dst,
+                                   int16_t *sao_offset_val,
+                                   int eo, int width, int height)
+{
+    ptrdiff_t stride_src = (2 * MAX_PB_SIZE + AV_INPUT_BUFFER_PADDING_SIZE);
+
+    switch (eo) {
+    case 0:
+        if (width >> 4) {
+            hevc_sao_edge_filter_0degree_16multiple_lsx(dst, stride_dst,
+                                                        src, stride_src,
+                                                        sao_offset_val,
+                                                        width - (width & 0x0F),
+                                                        height);
+            dst += width & 0xFFFFFFF0;
+            src += width & 0xFFFFFFF0;
+            width &= 0x0F;
+        }
+
+        if (width >> 3) {
+            hevc_sao_edge_filter_0degree_8width_lsx(dst, stride_dst,
+                                                    src, stride_src,
+                                                    sao_offset_val, height);
+            dst += 8;
+            src += 8;
+            width &= 0x07;
+        }
+
+        if (width) {
+            hevc_sao_edge_filter_0degree_4width_lsx(dst, stride_dst,
+                                                    src, stride_src,
+                                                    sao_offset_val, height);
+        }
+        break;
+
+    case 1:
+        if (width >> 4) {
+            hevc_sao_edge_filter_90degree_16multiple_lsx(dst, stride_dst,
+                                                         src, stride_src,
+                                                         sao_offset_val,
+                                                         width - (width & 0x0F),
+                                                         height);
+            dst += width & 0xFFFFFFF0;
+            src += width & 0xFFFFFFF0;
+            width &= 0x0F;
+        }
+
+        if (width >> 3) {
+            hevc_sao_edge_filter_90degree_8width_lsx(dst, stride_dst,
+                                                     src, stride_src,
+                                                     sao_offset_val, height);
+            dst += 8;
+            src += 8;
+            width &= 0x07;
+        }
+
+        if (width) {
+            hevc_sao_edge_filter_90degree_4width_lsx(dst, stride_dst,
+                                                     src, stride_src,
+                                                     sao_offset_val, height);
+        }
+        break;
+
+    case 2:
+        if (width >> 4) {
+            hevc_sao_edge_filter_45degree_16multiple_lsx(dst, stride_dst,
+                                                         src, stride_src,
+                                                         sao_offset_val,
+                                                         width - (width & 0x0F),
+                                                         height);
+            dst += width & 0xFFFFFFF0;
+            src += width & 0xFFFFFFF0;
+            width &= 0x0F;
+        }
+
+        if (width >> 3) {
+            hevc_sao_edge_filter_45degree_8width_lsx(dst, stride_dst,
+                                                     src, stride_src,
+                                                     sao_offset_val, height);
+            dst += 8;
+            src += 8;
+            width &= 0x07;
+        }
+
+        if (width) {
+            hevc_sao_edge_filter_45degree_4width_lsx(dst, stride_dst,
+                                                     src, stride_src,
+                                                     sao_offset_val, height);
+        }
+        break;
+
+    case 3:
+        if (width >> 4) {
+            hevc_sao_edge_filter_135degree_16multiple_lsx(dst, stride_dst,
+                                                          src, stride_src,
+                                                          sao_offset_val,
+                                                          width - (width & 0x0F),
+                                                          height);
+            dst += width & 0xFFFFFFF0;
+            src += width & 0xFFFFFFF0;
+            width &= 0x0F;
+        }
+
+        if (width >> 3) {
+            hevc_sao_edge_filter_135degree_8width_lsx(dst, stride_dst,
+                                                      src, stride_src,
+                                                      sao_offset_val, height);
+            dst += 8;
+            src += 8;
+            width &= 0x07;
+        }
+
+        if (width) {
+            hevc_sao_edge_filter_135degree_4width_lsx(dst, stride_dst,
+                                                      src, stride_src,
+                                                      sao_offset_val, height);
+        }
+        break;
+    }
+}
diff --git a/libavcodec/loongarch/hevcdsp_init_loongarch.c b/libavcodec/loongarch/hevcdsp_init_loongarch.c
index fc0e8fb0df..f39674be64 100644
--- a/libavcodec/loongarch/hevcdsp_init_loongarch.c
+++ b/libavcodec/loongarch/hevcdsp_init_loongarch.c
@@ -85,6 +85,25 @@ void ff_hevc_dsp_init_loongarch(HEVCDSPContext *c, const int bit_depth)
             c->put_hevc_epel[5][1][1] = ff_hevc_put_hevc_epel_hv16_8_lsx;
             c->put_hevc_epel[6][1][1] = ff_hevc_put_hevc_epel_hv24_8_lsx;
             c->put_hevc_epel[7][1][1] = ff_hevc_put_hevc_epel_hv32_8_lsx;
+
+            c->sao_edge_filter[4] = ff_hevc_sao_edge_filter_8_lsx;
+
+            c->hevc_h_loop_filter_luma = ff_hevc_loop_filter_luma_h_8_lsx;
+            c->hevc_v_loop_filter_luma = ff_hevc_loop_filter_luma_v_8_lsx;
+
+            c->hevc_h_loop_filter_luma_c = ff_hevc_loop_filter_luma_h_8_lsx;
+            c->hevc_v_loop_filter_luma_c = ff_hevc_loop_filter_luma_v_8_lsx;
+
+            c->hevc_h_loop_filter_chroma = ff_hevc_loop_filter_chroma_h_8_lsx;
+            c->hevc_v_loop_filter_chroma = ff_hevc_loop_filter_chroma_v_8_lsx;
+
+            c->hevc_h_loop_filter_chroma_c = ff_hevc_loop_filter_chroma_h_8_lsx;
+            c->hevc_v_loop_filter_chroma_c = ff_hevc_loop_filter_chroma_v_8_lsx;
+
+            c->idct[0] = ff_hevc_idct_4x4_lsx;
+            c->idct[1] = ff_hevc_idct_8x8_lsx;
+            c->idct[2] = ff_hevc_idct_16x16_lsx;
+            c->idct[3] = ff_hevc_idct_32x32_lsx;
         }
     }
 }
diff --git a/libavcodec/loongarch/hevcdsp_lsx.h b/libavcodec/loongarch/hevcdsp_lsx.h
index 3259e03f13..0e73fd1f8e 100644
--- a/libavcodec/loongarch/hevcdsp_lsx.h
+++ b/libavcodec/loongarch/hevcdsp_lsx.h
@@ -85,4 +85,30 @@ MC(epel, hv, 32);
 
 #undef MC
 
+void ff_hevc_loop_filter_luma_h_8_lsx(uint8_t *src, ptrdiff_t stride,
+                                      int32_t beta, int32_t *tc,
+                                      uint8_t *p_is_pcm, uint8_t *q_is_pcm);
+
+void ff_hevc_loop_filter_luma_v_8_lsx(uint8_t *src, ptrdiff_t stride,
+                                      int32_t beta, int32_t *tc,
+                                      uint8_t *p_is_pcm, uint8_t *q_is_pcm);
+
+void ff_hevc_loop_filter_chroma_h_8_lsx(uint8_t *src, ptrdiff_t stride,
+                                        int32_t *tc, uint8_t *p_is_pcm,
+                                        uint8_t *q_is_pcm);
+
+void ff_hevc_loop_filter_chroma_v_8_lsx(uint8_t *src, ptrdiff_t stride,
+                                        int32_t *tc, uint8_t *p_is_pcm,
+                                        uint8_t *q_is_pcm);
+
+void ff_hevc_sao_edge_filter_8_lsx(uint8_t *dst, uint8_t *src,
+                                   ptrdiff_t stride_dst,
+                                   int16_t *sao_offset_val,
+                                   int eo, int width, int height);
+
+void ff_hevc_idct_4x4_lsx(int16_t *coeffs, int col_limit);
+void ff_hevc_idct_8x8_lsx(int16_t *coeffs, int col_limit);
+void ff_hevc_idct_16x16_lsx(int16_t *coeffs, int col_limit);
+void ff_hevc_idct_32x32_lsx(int16_t *coeffs, int col_limit);
+
 #endif  // #ifndef AVCODEC_LOONGARCH_HEVCDSP_LSX_H
-- 
2.20.1



More information about the ffmpeg-devel mailing list