vp8[loongarch]: Optimize vp8_sixtap_predict4x4

1. vp8_sixtap_predict4x4

Bug: webm:1755

Change-Id: If7d844496ef2cfe2252f2ef12bb7cded63ad03dd

2 files changed, 743 insertions, 4 deletions

diff --git a/vp8/common/loongarch/sixtap_filter_lsx.c b/vp8/common/loongarch/sixtap_filter_lsx.c
index 75fe533d9..a23ed16d2 100644
--- a/vp8/common/loongarch/sixtap_filter_lsx.c
+++ b/vp8/common/loongarch/sixtap_filter_lsx.c
@@ -50,9 +50,9 @@ static const uint8_t vp8_mc_filt_mask_arr[16 * 3] = {
     __m128i vec0_m, vec1_m, vec2_m;                                         \
     __m128i hz_out_m;                                                       \
                                                                             \
-    DUP2_ARG3(__lsx_vshuf_b, src0, src1, mask0, src0, src1, mask1, vec0_m,  \
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, mask0, src1, src0, mask1, vec0_m,  \
               vec1_m);                                                      \
-    vec2_m = __lsx_vshuf_b(src0, src1, mask2);                              \
+    vec2_m = __lsx_vshuf_b(src1, src0, mask2);                              \
     hz_out_m = DPADD_H3(vec0_m, vec1_m, vec2_m, filt_h0, filt_h1, filt_h2); \
                                                                             \
     hz_out_m = __lsx_vsrari_h(hz_out_m, VP8_FILTER_SHIFT);                  \
@@ -61,6 +61,24 @@ static const uint8_t vp8_mc_filt_mask_arr[16 * 3] = {
     hz_out_m;                                                               \
   })
 
+#define HORIZ_6TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1,   \
+                                   mask2, filt0, filt1, filt2, out0, out1) \
+  {                                                                        \
+    __m128i vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m;                \
+                                                                           \
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, mask0, src3, src2, mask0, vec0_m, \
+              vec1_m);                                                     \
+    DUP2_ARG2(__lsx_vdp2_h_b, vec0_m, filt0, vec1_m, filt0, out0, out1);   \
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, mask1, src3, src2, mask1, vec2_m, \
+              vec3_m);                                                     \
+    DUP2_ARG3(__lsx_vdp2add_h_b, out0, vec2_m, filt1, out1, vec3_m, filt1, \
+              out0, out1);                                                 \
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, mask2, src3, src2, mask2, vec4_m, \
+              vec5_m);                                                     \
+    DUP2_ARG3(__lsx_vdp2add_h_b, out0, vec4_m, filt2, out1, vec5_m, filt2, \
+              out0, out1);                                                 \
+  }
+
 #define HORIZ_6TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1,    \
                                    mask2, filt0, filt1, filt2, out0, out1,  \
                                    out2, out3)                              \
@@ -104,7 +122,7 @@ static const uint8_t vp8_mc_filt_mask_arr[16 * 3] = {
     __m128i vec0_m, vec1_m;                                                \
     __m128i hz_out_m;                                                      \
                                                                            \
-    DUP2_ARG3(__lsx_vshuf_b, src0, src1, mask0, src0, src1, mask1, vec0_m, \
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, mask0, src1, src0, mask1, vec0_m, \
               vec1_m);                                                     \
     hz_out_m = FILT_4TAP_DPADD_H(vec0_m, vec1_m, filt_h0, filt_h1);        \
     hz_out_m = __lsx_vsrari_h(hz_out_m, VP8_FILTER_SHIFT);                 \
@@ -113,6 +131,20 @@ static const uint8_t vp8_mc_filt_mask_arr[16 * 3] = {
     hz_out_m;                                                              \
   })
 
+#define HORIZ_4TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1,   \
+                                   filt0, filt1, out0, out1)               \
+  {                                                                        \
+    __m128i vec0_m, vec1_m, vec2_m, vec3_m;                                \
+                                                                           \
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, mask0, src3, src2, mask0, vec0_m, \
+              vec1_m);                                                     \
+    DUP2_ARG2(__lsx_vdp2_h_b, vec0_m, filt0, vec1_m, filt0, out0, out1);   \
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, mask1, src3, src2, mask1, vec2_m, \
+              vec3_m);                                                     \
+    DUP2_ARG3(__lsx_vdp2add_h_b, out0, vec2_m, filt1, out1, vec3_m, filt1, \
+              out0, out1);                                                 \
+  }
+
 #define HORIZ_4TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1,   \
                                    filt0, filt1, out0, out1, out2, out3)   \
   ({                                                                       \
@@ -133,6 +165,107 @@ static const uint8_t vp8_mc_filt_mask_arr[16 * 3] = {
               out3);                                                       \
   })
 
+static inline void common_hz_6t_4x4_lsx(uint8_t *RESTRICT src,
+                                        int32_t src_stride,
+                                        uint8_t *RESTRICT dst,
+                                        int32_t dst_stride,
+                                        const int8_t *filter) {
+  __m128i src0, src1, src2, src3, filt0, filt1, filt2;
+  __m128i mask0, mask1, mask2, out0, out1;
+  int32_t src_stride_x2 = src_stride << 1;
+  int32_t src_stride_x3 = src_stride_x2 + src_stride;
+
+  mask0 = __lsx_vld(vp8_mc_filt_mask_arr, 16);
+  src -= 2;
+
+  DUP2_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filt0, filt1);
+  filt2 = __lsx_vldrepl_h(filter, 4);
+
+  DUP2_ARG2(__lsx_vaddi_bu, mask0, 2, mask0, 4, mask1, mask2);
+  src0 = __lsx_vld(src, 0);
+  DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride_x2, src1, src2);
+  src3 = __lsx_vldx(src, src_stride_x3);
+
+  DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0,
+            src1, src2, src3);
+  HORIZ_6TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, filt0,
+                             filt1, filt2, out0, out1);
+  out0 = __lsx_vssrarni_b_h(out1, out0, VP8_FILTER_SHIFT);
+  out0 = __lsx_vxori_b(out0, 128);
+
+  __lsx_vstelm_w(out0, dst, 0, 0);
+  dst += dst_stride;
+  __lsx_vstelm_w(out0, dst, 0, 1);
+  dst += dst_stride;
+  __lsx_vstelm_w(out0, dst, 0, 2);
+  dst += dst_stride;
+  __lsx_vstelm_w(out0, dst, 0, 3);
+}
+
+static void common_hz_6t_4x8_lsx(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter) {
+  __m128i src0, src1, src2, src3, filt0, filt1, filt2;
+  __m128i mask0, mask1, mask2, out0, out1, out2, out3;
+  int32_t src_stride_x2 = src_stride << 1;
+  int32_t src_stride_x3 = src_stride_x2 + src_stride;
+  int32_t src_stride_x4 = src_stride_x2 << 1;
+
+  mask0 = __lsx_vld(vp8_mc_filt_mask_arr, 16);
+  src -= 2;
+
+  DUP2_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filt0, filt1);
+  filt2 = __lsx_vldrepl_h(filter, 4);
+  DUP2_ARG2(__lsx_vaddi_bu, mask0, 2, mask0, 4, mask1, mask2);
+
+  src0 = __lsx_vld(src, 0);
+  DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride_x2, src1, src2);
+  src3 = __lsx_vldx(src, src_stride_x3);
+  DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0,
+            src1, src2, src3);
+  src += src_stride_x4;
+  HORIZ_6TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, filt0,
+                             filt1, filt2, out0, out1);
+
+  src0 = __lsx_vld(src, 0);
+  DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride_x2, src1, src2);
+  src3 = __lsx_vldx(src, src_stride_x3);
+  DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0,
+            src1, src2, src3);
+  HORIZ_6TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, filt0,
+                             filt1, filt2, out2, out3);
+
+  DUP2_ARG3(__lsx_vssrarni_b_h, out1, out0, VP8_FILTER_SHIFT, out3, out2,
+            VP8_FILTER_SHIFT, out0, out1);
+  DUP2_ARG2(__lsx_vxori_b, out0, 128, out1, 128, out0, out1);
+  __lsx_vstelm_w(out0, dst, 0, 0);
+  dst += dst_stride;
+  __lsx_vstelm_w(out0, dst, 0, 1);
+  dst += dst_stride;
+  __lsx_vstelm_w(out0, dst, 0, 2);
+  dst += dst_stride;
+  __lsx_vstelm_w(out0, dst, 0, 3);
+  dst += dst_stride;
+
+  __lsx_vstelm_w(out1, dst, 0, 0);
+  dst += dst_stride;
+  __lsx_vstelm_w(out1, dst, 0, 1);
+  dst += dst_stride;
+  __lsx_vstelm_w(out1, dst, 0, 2);
+  dst += dst_stride;
+  __lsx_vstelm_w(out1, dst, 0, 3);
+}
+
+static void common_hz_6t_4w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
+                                uint8_t *RESTRICT dst, int32_t dst_stride,
+                                const int8_t *filter, int32_t height) {
+  if (height == 4) {
+    common_hz_6t_4x4_lsx(src, src_stride, dst, dst_stride, filter);
+  } else if (height == 8) {
+    common_hz_6t_4x8_lsx(src, src_stride, dst, dst_stride, filter);
+  }
+}
+
 static void common_hz_6t_8w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
                                 uint8_t *RESTRICT dst, int32_t dst_stride,
                                 const int8_t *filter, int32_t height) {
@@ -254,6 +387,64 @@ static void common_hz_6t_16w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
   }
 }
 
+static void common_vt_6t_4w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
+                                uint8_t *RESTRICT dst, int32_t dst_stride,
+                                const int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8;
+  __m128i src10_r, src32_r, src54_r, src76_r, src21_r, src43_r, src65_r;
+  __m128i src87_r, src2110, src4332, src6554, src8776, filt0, filt1, filt2;
+  __m128i out0, out1;
+  int32_t src_stride_x2 = src_stride << 1;
+  int32_t src_stride_x3 = src_stride_x2 + src_stride;
+  int32_t src_stride_x4 = src_stride << 2;
+
+  DUP2_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filt0, filt1);
+  filt2 = __lsx_vldrepl_h(filter, 4);
+
+  DUP2_ARG2(__lsx_vldx, src, -src_stride_x2, src, -src_stride, src0, src1);
+  src2 = __lsx_vld(src, 0);
+  DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride_x2, src3, src4);
+  src += src_stride_x3;
+
+  DUP4_ARG2(__lsx_vilvl_b, src1, src0, src2, src1, src3, src2, src4, src3,
+            src10_r, src21_r, src32_r, src43_r);
+  DUP2_ARG2(__lsx_vilvl_d, src21_r, src10_r, src43_r, src32_r, src2110,
+            src4332);
+  DUP2_ARG2(__lsx_vxori_b, src2110, 128, src4332, 128, src2110, src4332);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    src5 = __lsx_vld(src, 0);
+    DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride_x2, src6, src7);
+    src8 = __lsx_vldx(src, src_stride_x3);
+    src += src_stride_x4;
+
+    DUP4_ARG2(__lsx_vilvl_b, src5, src4, src6, src5, src7, src6, src8, src7,
+              src54_r, src65_r, src76_r, src87_r);
+    DUP2_ARG2(__lsx_vilvl_d, src65_r, src54_r, src87_r, src76_r, src6554,
+              src8776);
+    DUP2_ARG2(__lsx_vxori_b, src6554, 128, src8776, 128, src6554, src8776);
+    out0 = DPADD_H3(src2110, src4332, src6554, filt0, filt1, filt2);
+    out1 = DPADD_H3(src4332, src6554, src8776, filt0, filt1, filt2);
+
+    out0 = __lsx_vssrarni_b_h(out1, out0, VP8_FILTER_SHIFT);
+    out0 = __lsx_vxori_b(out0, 128);
+
+    __lsx_vstelm_w(out0, dst, 0, 0);
+    dst += dst_stride;
+    __lsx_vstelm_w(out0, dst, 0, 1);
+    dst += dst_stride;
+    __lsx_vstelm_w(out0, dst, 0, 2);
+    dst += dst_stride;
+    __lsx_vstelm_w(out0, dst, 0, 3);
+    dst += dst_stride;
+
+    src2110 = src6554;
+    src4332 = src8776;
+    src4 = src8;
+  }
+}
+
 static void common_vt_6t_8w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
                                 uint8_t *RESTRICT dst, int32_t dst_stride,
                                 const int8_t *filter, int32_t height) {
@@ -395,6 +586,92 @@ static void common_vt_6t_16w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
   }
 }
 
+static void common_hv_6ht_6vt_4w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
+                                     uint8_t *RESTRICT dst, int32_t dst_stride,
+                                     const int8_t *filter_horiz,
+                                     const int8_t *filter_vert,
+                                     int32_t height) {
+  uint32_t loop_cnt;
+  __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8, tmp0, tmp1;
+  __m128i filt_hz0, filt_hz1, filt_hz2, mask0, mask1, mask2;
+  __m128i hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+  __m128i hz_out7, filt_vt0, filt_vt1, filt_vt2, out0, out1, out2, out3;
+  __m128i shuff = { 0x0F0E0D0C0B0A0908, 0x1716151413121110 };
+  int32_t src_stride_x2 = src_stride << 1;
+  int32_t src_stride_x3 = src_stride_x2 + src_stride;
+
+  mask0 = __lsx_vld(vp8_mc_filt_mask_arr, 16);
+  src -= 2;
+
+  DUP2_ARG2(__lsx_vldrepl_h, filter_horiz, 0, filter_horiz, 2, filt_hz0,
+            filt_hz1);
+  filt_hz2 = __lsx_vldrepl_h(filter_horiz, 4);
+  DUP2_ARG2(__lsx_vldrepl_h, filter_vert, 0, filter_vert, 2, filt_vt0,
+            filt_vt1);
+  filt_vt2 = __lsx_vldrepl_h(filter_vert, 4);
+
+  DUP2_ARG2(__lsx_vaddi_bu, mask0, 2, mask0, 4, mask1, mask2);
+
+  DUP2_ARG2(__lsx_vldx, src, -src_stride_x2, src, -src_stride, src0, src1);
+  src2 = __lsx_vld(src, 0);
+  DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride_x2, src3, src4);
+  src += src_stride_x3;
+
+  DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0,
+            src1, src2, src3);
+  src4 = __lsx_vxori_b(src4, 128);
+
+  hz_out0 = HORIZ_6TAP_FILT(src0, src1, mask0, mask1, mask2, filt_hz0, filt_hz1,
+                            filt_hz2);
+  hz_out2 = HORIZ_6TAP_FILT(src2, src3, mask0, mask1, mask2, filt_hz0, filt_hz1,
+                            filt_hz2);
+  hz_out1 = __lsx_vshuf_b(hz_out2, hz_out0, shuff);
+  hz_out3 = HORIZ_6TAP_FILT(src3, src4, mask0, mask1, mask2, filt_hz0, filt_hz1,
+                            filt_hz2);
+  DUP2_ARG2(__lsx_vpackev_b, hz_out1, hz_out0, hz_out3, hz_out2, out0, out1);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    src5 = __lsx_vld(src, 0);
+    src6 = __lsx_vldx(src, src_stride);
+    src += src_stride_x2;
+
+    DUP2_ARG2(__lsx_vxori_b, src5, 128, src6, 128, src5, src6);
+    hz_out5 = HORIZ_6TAP_FILT(src5, src6, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+    hz_out4 = __lsx_vshuf_b(hz_out5, hz_out3, shuff);
+
+    src7 = __lsx_vld(src, 0);
+    src8 = __lsx_vldx(src, src_stride);
+    src += src_stride_x2;
+
+    DUP2_ARG2(__lsx_vxori_b, src7, 128, src8, 128, src7, src8);
+    hz_out7 = HORIZ_6TAP_FILT(src7, src8, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+    hz_out6 = __lsx_vshuf_b(hz_out7, hz_out5, shuff);
+
+    out2 = __lsx_vpackev_b(hz_out5, hz_out4);
+    tmp0 = DPADD_H3(out0, out1, out2, filt_vt0, filt_vt1, filt_vt2);
+
+    out3 = __lsx_vpackev_b(hz_out7, hz_out6);
+    tmp1 = DPADD_H3(out1, out2, out3, filt_vt0, filt_vt1, filt_vt2);
+
+    tmp0 = __lsx_vssrarni_b_h(tmp1, tmp0, 7);
+    tmp0 = __lsx_vxori_b(tmp0, 128);
+    __lsx_vstelm_w(tmp0, dst, 0, 0);
+    dst += dst_stride;
+    __lsx_vstelm_w(tmp0, dst, 0, 1);
+    dst += dst_stride;
+    __lsx_vstelm_w(tmp0, dst, 0, 2);
+    dst += dst_stride;
+    __lsx_vstelm_w(tmp0, dst, 0, 3);
+    dst += dst_stride;
+
+    hz_out3 = hz_out7;
+    out0 = out2;
+    out1 = out3;
+  }
+}
+
 static void common_hv_6ht_6vt_8w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
                                      uint8_t *RESTRICT dst, int32_t dst_stride,
                                      const int8_t *filter_horiz,
@@ -506,6 +783,102 @@ static void common_hv_6ht_6vt_16w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
                            filter_horiz, filter_vert, height);
 }
 
+static void common_hz_4t_4x4_lsx(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter) {
+  __m128i src0, src1, src2, src3, filt0, filt1, mask0, mask1;
+  __m128i out0, out1;
+  int32_t src_stride_x2 = src_stride << 1;
+  int32_t src_stride_x3 = src_stride_x2 + src_stride;
+
+  mask0 = __lsx_vld(vp8_mc_filt_mask_arr, 16);
+  src -= 1;
+
+  DUP2_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filt0, filt1);
+  mask1 = __lsx_vaddi_bu(mask0, 2);
+
+  src0 = __lsx_vld(src, 0);
+  DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride_x2, src1, src2);
+  src3 = __lsx_vldx(src, src_stride_x3);
+  DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0,
+            src1, src2, src3);
+  HORIZ_4TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, filt0, filt1,
+                             out0, out1);
+
+  out0 = __lsx_vssrarni_b_h(out1, out0, VP8_FILTER_SHIFT);
+  out0 = __lsx_vxori_b(out0, 128);
+
+  __lsx_vstelm_w(out0, dst, 0, 0);
+  dst += dst_stride;
+  __lsx_vstelm_w(out0, dst, 0, 1);
+  dst += dst_stride;
+  __lsx_vstelm_w(out0, dst, 0, 2);
+  dst += dst_stride;
+  __lsx_vstelm_w(out0, dst, 0, 3);
+}
+
+static void common_hz_4t_4x8_lsx(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter) {
+  __m128i src0, src1, src2, src3, filt0, filt1, mask0, mask1;
+  __m128i out0, out1, out2, out3;
+  int32_t src_stride_x2 = src_stride << 1;
+  int32_t src_stride_x3 = src_stride_x2 + src_stride;
+  int32_t src_stride_x4 = src_stride << 2;
+
+  mask0 = __lsx_vld(vp8_mc_filt_mask_arr, 16);
+  src -= 1;
+
+  DUP2_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filt0, filt1);
+  mask1 = __lsx_vaddi_bu(mask0, 2);
+
+  src0 = __lsx_vld(src, 0);
+  DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride_x2, src1, src2);
+  src3 = __lsx_vldx(src, src_stride_x3);
+  src += src_stride_x4;
+  DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0,
+            src1, src2, src3);
+  HORIZ_4TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, filt0, filt1,
+                             out0, out1);
+
+  src0 = __lsx_vld(src, 0);
+  DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride_x2, src1, src2);
+  src3 = __lsx_vldx(src, src_stride_x3);
+  DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0,
+            src1, src2, src3);
+  HORIZ_4TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, filt0, filt1,
+                             out2, out3);
+  DUP2_ARG3(__lsx_vssrarni_b_h, out1, out0, VP8_FILTER_SHIFT, out3, out2,
+            VP8_FILTER_SHIFT, out0, out1);
+  DUP2_ARG2(__lsx_vxori_b, out0, 128, out1, 128, out0, out1);
+  __lsx_vstelm_w(out0, dst, 0, 0);
+  dst += dst_stride;
+  __lsx_vstelm_w(out0, dst, 0, 1);
+  dst += dst_stride;
+  __lsx_vstelm_w(out0, dst, 0, 2);
+  dst += dst_stride;
+  __lsx_vstelm_w(out0, dst, 0, 3);
+  dst += dst_stride;
+
+  __lsx_vstelm_w(out1, dst, 0, 0);
+  dst += dst_stride;
+  __lsx_vstelm_w(out1, dst, 0, 1);
+  dst += dst_stride;
+  __lsx_vstelm_w(out1, dst, 0, 2);
+  dst += dst_stride;
+  __lsx_vstelm_w(out1, dst, 0, 3);
+}
+
+static void common_hz_4t_4w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
+                                uint8_t *RESTRICT dst, int32_t dst_stride,
+                                const int8_t *filter, int32_t height) {
+  if (height == 4) {
+    common_hz_4t_4x4_lsx(src, src_stride, dst, dst_stride, filter);
+  } else if (height == 8) {
+    common_hz_4t_4x8_lsx(src, src_stride, dst, dst_stride, filter);
+  }
+}
+
 static void common_hz_4t_8w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
                                 uint8_t *RESTRICT dst, int32_t dst_stride,
                                 const int8_t *filter, int32_t height) {
@@ -597,6 +970,55 @@ static void common_hz_4t_16w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
   }
 }
 
+static void common_vt_4t_4w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
+                                uint8_t *RESTRICT dst, int32_t dst_stride,
+                                const int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  __m128i src0, src1, src2, src3, src4, src5;
+  __m128i src10_r, src32_r, src54_r, src21_r, src43_r, src65_r;
+  __m128i src2110, src4332, filt0, filt1, out0, out1;
+  int32_t src_stride_x2 = src_stride << 1;
+  int32_t src_stride_x3 = src_stride_x2 + src_stride;
+
+  DUP2_ARG2(__lsx_vldrepl_h, filter, 0, filter, 2, filt0, filt1);
+  DUP2_ARG2(__lsx_vldx, src, -src_stride, src, src_stride, src0, src2);
+  src1 = __lsx_vld(src, 0);
+  src += src_stride_x2;
+
+  DUP2_ARG2(__lsx_vilvl_b, src1, src0, src2, src1, src10_r, src21_r);
+
+  src2110 = __lsx_vilvl_d(src21_r, src10_r);
+  src2110 = __lsx_vxori_b(src2110, 128);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    src3 = __lsx_vld(src, 0);
+    DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride_x2, src4, src5);
+    src += src_stride_x3;
+    DUP2_ARG2(__lsx_vilvl_b, src3, src2, src4, src3, src32_r, src43_r);
+    src4332 = __lsx_vilvl_d(src43_r, src32_r);
+    src4332 = __lsx_vxori_b(src4332, 128);
+    out0 = FILT_4TAP_DPADD_H(src2110, src4332, filt0, filt1);
+
+    src2 = __lsx_vld(src, 0);
+    src += src_stride;
+    DUP2_ARG2(__lsx_vilvl_b, src5, src4, src2, src5, src54_r, src65_r);
+    src2110 = __lsx_vilvl_d(src65_r, src54_r);
+    src2110 = __lsx_vxori_b(src2110, 128);
+    out1 = FILT_4TAP_DPADD_H(src4332, src2110, filt0, filt1);
+    out0 = __lsx_vssrarni_b_h(out1, out0, VP8_FILTER_SHIFT);
+    out0 = __lsx_vxori_b(out0, 128);
+
+    __lsx_vstelm_w(out0, dst, 0, 0);
+    dst += dst_stride;
+    __lsx_vstelm_w(out0, dst, 0, 1);
+    dst += dst_stride;
+    __lsx_vstelm_w(out0, dst, 0, 2);
+    dst += dst_stride;
+    __lsx_vstelm_w(out0, dst, 0, 3);
+    dst += dst_stride;
+  }
+}
+
 static void common_vt_4t_8w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
                                 uint8_t *RESTRICT dst, int32_t dst_stride,
                                 const int8_t *filter, int32_t height) {
@@ -719,6 +1141,74 @@ static void common_vt_4t_16w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
   }
 }
 
+static void common_hv_4ht_4vt_4w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
+                                     uint8_t *RESTRICT dst, int32_t dst_stride,
+                                     const int8_t *filter_horiz,
+                                     const int8_t *filter_vert,
+                                     int32_t height) {
+  uint32_t loop_cnt;
+  __m128i src0, src1, src2, src3, src4, src5, src6, filt_hz0, filt_hz1;
+  __m128i mask0, mask1, filt_vt0, filt_vt1, tmp0, tmp1, vec0, vec1, vec2;
+  __m128i hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5;
+  __m128i shuff = { 0x0F0E0D0C0B0A0908, 0x1716151413121110 };
+  int32_t src_stride_x2 = src_stride << 1;
+  int32_t src_stride_x3 = src_stride_x2 + src_stride;
+  int32_t src_stride_x4 = src_stride << 2;
+
+  mask0 = __lsx_vld(vp8_mc_filt_mask_arr, 16);
+  src -= 1;
+
+  DUP2_ARG2(__lsx_vldrepl_h, filter_horiz, 0, filter_horiz, 2, filt_hz0,
+            filt_hz1);
+  mask1 = __lsx_vaddi_bu(mask0, 2);
+
+  src1 = __lsx_vld(src, 0);
+  DUP2_ARG2(__lsx_vldx, src, -src_stride, src, src_stride, src0, src2);
+  src += src_stride_x2;
+
+  DUP2_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src0, src1);
+  src2 = __lsx_vxori_b(src2, 128);
+  hz_out0 = HORIZ_4TAP_FILT(src0, src1, mask0, mask1, filt_hz0, filt_hz1);
+  hz_out1 = HORIZ_4TAP_FILT(src1, src2, mask0, mask1, filt_hz0, filt_hz1);
+  vec0 = __lsx_vpackev_b(hz_out1, hz_out0);
+
+  DUP2_ARG2(__lsx_vldrepl_h, filter_vert, 0, filter_vert, 2, filt_vt0,
+            filt_vt1);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    src3 = __lsx_vld(src, 0);
+    DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride_x2, src4, src5);
+    src6 = __lsx_vldx(src, src_stride_x3);
+    src += src_stride_x4;
+
+    DUP2_ARG2(__lsx_vxori_b, src3, 128, src4, 128, src3, src4);
+    hz_out3 = HORIZ_4TAP_FILT(src3, src4, mask0, mask1, filt_hz0, filt_hz1);
+    hz_out2 = __lsx_vshuf_b(hz_out3, hz_out1, shuff);
+    vec1 = __lsx_vpackev_b(hz_out3, hz_out2);
+    tmp0 = FILT_4TAP_DPADD_H(vec0, vec1, filt_vt0, filt_vt1);
+
+    DUP2_ARG2(__lsx_vxori_b, src5, 128, src6, 128, src5, src6);
+    hz_out5 = HORIZ_4TAP_FILT(src5, src6, mask0, mask1, filt_hz0, filt_hz1);
+    hz_out4 = __lsx_vshuf_b(hz_out5, hz_out3, shuff);
+    vec2 = __lsx_vpackev_b(hz_out5, hz_out4);
+    tmp1 = FILT_4TAP_DPADD_H(vec1, vec2, filt_vt0, filt_vt1);
+
+    tmp0 = __lsx_vssrarni_b_h(tmp1, tmp0, 7);
+    tmp0 = __lsx_vxori_b(tmp0, 128);
+    __lsx_vstelm_w(tmp0, dst, 0, 0);
+    dst += dst_stride;
+    __lsx_vstelm_w(tmp0, dst, 0, 1);
+    dst += dst_stride;
+    __lsx_vstelm_w(tmp0, dst, 0, 2);
+    dst += dst_stride;
+    __lsx_vstelm_w(tmp0, dst, 0, 3);
+    dst += dst_stride;
+
+    hz_out1 = hz_out5;
+    vec0 = vec2;
+  }
+}
+
 static inline void common_hv_4ht_4vt_8w_lsx(
     uint8_t *RESTRICT src, int32_t src_stride, uint8_t *RESTRICT dst,
     int32_t dst_stride, const int8_t *filter_horiz, const int8_t *filter_vert,
@@ -804,6 +1294,82 @@ static void common_hv_4ht_4vt_16w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
                            filter_horiz, filter_vert, height);
 }
 
+static void common_hv_6ht_4vt_4w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
+                                     uint8_t *RESTRICT dst, int32_t dst_stride,
+                                     const int8_t *filter_horiz,
+                                     const int8_t *filter_vert,
+                                     int32_t height) {
+  uint32_t loop_cnt;
+  __m128i src0, src1, src2, src3, src4, src5, src6;
+  __m128i filt_hz0, filt_hz1, filt_hz2, mask0, mask1, mask2;
+  __m128i filt_vt0, filt_vt1, tmp0, tmp1, vec0, vec1, vec2;
+  __m128i hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5;
+  __m128i shuff = { 0x0F0E0D0C0B0A0908, 0x1716151413121110 };
+  int32_t src_stride_x2 = src_stride << 1;
+  int32_t src_stride_x3 = src_stride_x2 + src_stride;
+  int32_t src_stride_x4 = src_stride << 2;
+
+  mask0 = __lsx_vld(vp8_mc_filt_mask_arr, 16);
+  src -= 2;
+
+  DUP2_ARG2(__lsx_vldrepl_h, filter_horiz, 0, filter_horiz, 2, filt_hz0,
+            filt_hz1);
+  filt_hz2 = __lsx_vldrepl_h(filter_horiz, 4);
+  DUP2_ARG2(__lsx_vaddi_bu, mask0, 2, mask0, 4, mask1, mask2);
+
+  src1 = __lsx_vld(src, 0);
+  DUP2_ARG2(__lsx_vldx, src, -src_stride, src, src_stride, src0, src2);
+  src += src_stride_x2;
+
+  DUP2_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src0, src1);
+  src2 = __lsx_vxori_b(src2, 128);
+
+  hz_out0 = HORIZ_6TAP_FILT(src0, src1, mask0, mask1, mask2, filt_hz0, filt_hz1,
+                            filt_hz2);
+  hz_out1 = HORIZ_6TAP_FILT(src1, src2, mask0, mask1, mask2, filt_hz0, filt_hz1,
+                            filt_hz2);
+  vec0 = __lsx_vpackev_b(hz_out1, hz_out0);
+
+  DUP2_ARG2(__lsx_vldrepl_h, filter_vert, 0, filter_vert, 2, filt_vt0,
+            filt_vt1);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    src3 = __lsx_vld(src, 0);
+    DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride_x2, src4, src5);
+    src6 = __lsx_vldx(src, src_stride_x3);
+    src += src_stride_x4;
+    DUP4_ARG2(__lsx_vxori_b, src3, 128, src4, 128, src5, 128, src6, 128, src3,
+              src4, src5, src6);
+
+    hz_out3 = HORIZ_6TAP_FILT(src3, src4, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+    hz_out2 = __lsx_vshuf_b(hz_out3, hz_out1, shuff);
+    vec1 = __lsx_vpackev_b(hz_out3, hz_out2);
+    tmp0 = FILT_4TAP_DPADD_H(vec0, vec1, filt_vt0, filt_vt1);
+
+    hz_out5 = HORIZ_6TAP_FILT(src5, src6, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+    hz_out4 = __lsx_vshuf_b(hz_out5, hz_out3, shuff);
+    vec2 = __lsx_vpackev_b(hz_out5, hz_out4);
+    tmp1 = FILT_4TAP_DPADD_H(vec1, vec2, filt_vt0, filt_vt1);
+
+    DUP2_ARG3(__lsx_vssrarni_b_h, tmp0, tmp0, 7, tmp1, tmp1, 7, tmp0, tmp1);
+    DUP2_ARG2(__lsx_vxori_b, tmp0, 128, tmp1, 128, tmp0, tmp1);
+
+    __lsx_vstelm_w(tmp0, dst, 0, 0);
+    dst += dst_stride;
+    __lsx_vstelm_w(tmp0, dst, 0, 1);
+    dst += dst_stride;
+    __lsx_vstelm_w(tmp1, dst, 0, 0);
+    dst += dst_stride;
+    __lsx_vstelm_w(tmp1, dst, 0, 1);
+    dst += dst_stride;
+
+    hz_out1 = hz_out5;
+    vec0 = vec2;
+  }
+}
+
 static inline void common_hv_6ht_4vt_8w_lsx(
     uint8_t *RESTRICT src, int32_t src_stride, uint8_t *RESTRICT dst,
     int32_t dst_stride, const int8_t *filter_horiz, const int8_t *filter_vert,
@@ -895,6 +1461,82 @@ static void common_hv_6ht_4vt_16w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
                            filter_horiz, filter_vert, height);
 }
 
+static void common_hv_4ht_6vt_4w_lsx(uint8_t *RESTRICT src, int32_t src_stride,
+                                     uint8_t *RESTRICT dst, int32_t dst_stride,
+                                     const int8_t *filter_horiz,
+                                     const int8_t *filter_vert,
+                                     int32_t height) {
+  uint32_t loop_cnt;
+  __m128i src0, src1, src2, src3, src4, src5, src6, src7, src8;
+  __m128i filt_hz0, filt_hz1, filt_vt0, filt_vt1, filt_vt2, mask0, mask1;
+  __m128i hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+  __m128i hz_out7, tmp0, tmp1, out0, out1, out2, out3;
+  __m128i shuff = { 0x0F0E0D0C0B0A0908, 0x1716151413121110 };
+  int32_t src_stride_x2 = src_stride << 1;
+  int32_t src_stride_x3 = src_stride_x2 + src_stride;
+  int32_t src_stride_x4 = src_stride << 2;
+
+  mask0 = __lsx_vld(vp8_mc_filt_mask_arr, 16);
+
+  src -= 1;
+
+  DUP2_ARG2(__lsx_vldrepl_h, filter_horiz, 0, filter_horiz, 2, filt_hz0,
+            filt_hz1);
+  mask1 = __lsx_vaddi_bu(mask0, 2);
+
+  DUP4_ARG2(__lsx_vldx, src, -src_stride_x2, src, -src_stride, src, src_stride,
+            src, src_stride_x2, src0, src1, src3, src4);
+  src2 = __lsx_vld(src, 0);
+  src += src_stride_x3;
+
+  DUP4_ARG2(__lsx_vxori_b, src0, 128, src1, 128, src2, 128, src3, 128, src0,
+            src1, src2, src3);
+  src4 = __lsx_vxori_b(src4, 128);
+  hz_out0 = HORIZ_4TAP_FILT(src0, src1, mask0, mask1, filt_hz0, filt_hz1);
+  hz_out2 = HORIZ_4TAP_FILT(src2, src3, mask0, mask1, filt_hz0, filt_hz1);
+  hz_out3 = HORIZ_4TAP_FILT(src3, src4, mask0, mask1, filt_hz0, filt_hz1);
+  hz_out1 = __lsx_vshuf_b(hz_out2, hz_out0, shuff);
+  DUP2_ARG2(__lsx_vpackev_b, hz_out1, hz_out0, hz_out3, hz_out2, out0, out1);
+
+  DUP2_ARG2(__lsx_vldrepl_h, filter_vert, 0, filter_vert, 2, filt_vt0,
+            filt_vt1);
+  filt_vt2 = __lsx_vldrepl_h(filter_vert, 4);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    src5 = __lsx_vld(src, 0);
+    DUP2_ARG2(__lsx_vldx, src, src_stride, src, src_stride_x2, src6, src7);
+    src8 = __lsx_vldx(src, src_stride_x3);
+    DUP4_ARG2(__lsx_vxori_b, src5, 128, src6, 128, src7, 128, src8, 128, src5,
+              src6, src7, src8);
+    src += src_stride_x4;
+
+    hz_out5 = HORIZ_4TAP_FILT(src5, src6, mask0, mask1, filt_hz0, filt_hz1);
+    hz_out4 = __lsx_vshuf_b(hz_out5, hz_out3, shuff);
+    out2 = __lsx_vpackev_b(hz_out5, hz_out4);
+    tmp0 = DPADD_H3(out0, out1, out2, filt_vt0, filt_vt1, filt_vt2);
+
+    hz_out7 = HORIZ_4TAP_FILT(src7, src8, mask0, mask1, filt_hz0, filt_hz1);
+    hz_out6 = __lsx_vshuf_b(hz_out7, hz_out5, shuff);
+    out3 = __lsx_vpackev_b(hz_out7, hz_out6);
+    tmp1 = DPADD_H3(out1, out2, out3, filt_vt0, filt_vt1, filt_vt2);
+
+    tmp0 = __lsx_vssrarni_b_h(tmp1, tmp0, 7);
+    tmp0 = __lsx_vxori_b(tmp0, 128);
+    __lsx_vstelm_w(tmp0, dst, 0, 0);
+    dst += dst_stride;
+    __lsx_vstelm_w(tmp0, dst, 0, 1);
+    dst += dst_stride;
+    __lsx_vstelm_w(tmp0, dst, 0, 2);
+    dst += dst_stride;
+    __lsx_vstelm_w(tmp0, dst, 0, 3);
+    dst += dst_stride;
+
+    hz_out3 = hz_out7;
+    out0 = out2;
+    out1 = out3;
+  }
+}
+
 static inline void common_hv_4ht_6vt_8w_lsx(
     uint8_t *RESTRICT src, int32_t src_stride, uint8_t *RESTRICT dst,
     int32_t dst_stride, const int8_t *filter_horiz, const int8_t *filter_vert,
@@ -1000,6 +1642,103 @@ typedef void (*PVp8SixtapPredictFunc2)(uint8_t *RESTRICT src,
                                        int32_t dst_stride, const int8_t *filter,
                                        int32_t height);
 
+void vp8_sixtap_predict4x4_lsx(uint8_t *RESTRICT src, int32_t src_stride,
+                               int32_t xoffset, int32_t yoffset,
+                               uint8_t *RESTRICT dst, int32_t dst_stride) {
+  const int8_t *h_filter = vp8_subpel_filters_lsx[xoffset - 1];
+  const int8_t *v_filter = vp8_subpel_filters_lsx[yoffset - 1];
+
+  static PVp8SixtapPredictFunc1 Predict4x4Funcs1[4] = {
+    common_hv_6ht_6vt_4w_lsx,
+    common_hv_6ht_4vt_4w_lsx,
+    common_hv_4ht_6vt_4w_lsx,
+    common_hv_4ht_4vt_4w_lsx,
+  };
+
+  static PVp8SixtapPredictFunc2 Predict4x4Funcs2[4] = { common_vt_6t_4w_lsx,
+                                                        common_vt_4t_4w_lsx,
+                                                        common_hz_6t_4w_lsx,
+                                                        common_hz_4t_4w_lsx };
+  if (yoffset < 8 && xoffset < 8) {
+    if (yoffset) {
+      if (xoffset) {
+        switch (xoffset & 1) {
+          case 0:
+            switch (yoffset & 1) {
+              case 0:
+                Predict4x4Funcs1[0](src, src_stride, dst, dst_stride, h_filter,
+                                    v_filter, 4);
+                break;
+              case 1:
+                Predict4x4Funcs1[1](src, src_stride, dst, dst_stride, h_filter,
+                                    v_filter + 1, 4);
+                break;
+            }
+            break;
+
+          case 1:
+            switch (yoffset & 1) {
+              case 0:
+                Predict4x4Funcs1[2](src, src_stride, dst, dst_stride,
+                                    h_filter + 1, v_filter, 4);
+                break;
+
+              case 1:
+                Predict4x4Funcs1[3](src, src_stride, dst, dst_stride,
+                                    h_filter + 1, v_filter + 1, 4);
+                break;
+            }
+            break;
+        }
+      } else {
+        switch (yoffset & 1) {
+          case 0:
+            Predict4x4Funcs2[0](src, src_stride, dst, dst_stride, v_filter, 4);
+            break;
+
+          case 1:
+            Predict4x4Funcs2[1](src, src_stride, dst, dst_stride, v_filter + 1,
+                                4);
+            break;
+        }
+      }
+    } else {
+      switch (xoffset) {
+        case 0: {
+          __m128i tp0;
+          tp0 = __lsx_vinsgr2vr_w(tp0, src, 0);
+          src += src_stride;
+          tp0 = __lsx_vinsgr2vr_w(tp0, src, 0);
+          src += src_stride;
+          tp0 = __lsx_vinsgr2vr_w(tp0, src, 0);
+          src += src_stride;
+          tp0 = __lsx_vinsgr2vr_w(tp0, src, 0);
+
+          __lsx_vstelm_w(tp0, dst, 0, 0);
+          dst += dst_stride;
+          __lsx_vstelm_w(tp0, dst, 0, 1);
+          dst += dst_stride;
+          __lsx_vstelm_w(tp0, dst, 0, 2);
+          dst += dst_stride;
+          __lsx_vstelm_w(tp0, dst, 0, 3);
+          break;
+        }
+        case 2:
+        case 4:
+        case 6:
+          Predict4x4Funcs2[2](src, src_stride, dst, dst_stride, h_filter, 4);
+          break;
+      }
+      switch (xoffset & 1) {
+        case 1:
+          Predict4x4Funcs2[3](src, src_stride, dst, dst_stride, h_filter + 1,
+                              4);
+          break;
+      }
+    }
+  }
+}
+
 void vp8_sixtap_predict8x8_lsx(uint8_t *RESTRICT src, int32_t src_stride,
                                int32_t xoffset, int32_t yoffset,
                                uint8_t *RESTRICT dst, int32_t dst_stride) {
diff --git a/vp8/common/rtcd_defs.pl b/vp8/common/rtcd_defs.pl
index 7bc866faa..739a61284 100644
--- a/vp8/common/rtcd_defs.pl
+++ b/vp8/common/rtcd_defs.pl
@@ -155,7 +155,7 @@ add_proto qw/void vp8_sixtap_predict8x4/, "unsigned char *src_ptr, int src_pixel
 specialize qw/vp8_sixtap_predict8x4 sse2 ssse3 neon dspr2 msa mmi/;
 
 add_proto qw/void vp8_sixtap_predict4x4/, "unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch";
-specialize qw/vp8_sixtap_predict4x4 mmx ssse3 neon dspr2 msa mmi/;
+specialize qw/vp8_sixtap_predict4x4 mmx ssse3 neon dspr2 msa mmi lsx/;
 
 add_proto qw/void vp8_bilinear_predict16x16/, "unsigned char *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, unsigned char *dst_ptr, int dst_pitch";
 specialize qw/vp8_bilinear_predict16x16 sse2 ssse3 neon msa/;